Most GEO advice stops at "get cited." The orchestration layer behind ChatGPT, Perplexity, and Gemini deep-research modes adds three stages after retrieval: fan-out, worker synthesis, and critic review. Each one changes what visibility means. Here are the three links nobody connect

1. Fan-out turns one buyer prompt into a dozen sub-queries you never tracked

An orchestrator receives "best project management tool for a 40-person agency" and splits it into narrow worker tasks: pricing comparison, integrations check, review sentiment, migration pain. Each worker retrieves independently. Your brand has to win the sub-query, not the headline prompt.

This is why a brand can rank first on a tracked prompt and vanish in the agentic version of the same question. The tracked prompt is the parent; the answer is built from the children. If your tracking only logs the parent, you are measuring a layer the model barely uses.

2. The critic agent enforces citation-level attribution

Orchestrator, narrow workers, verification: the verifier is typically a second model call with one job, rejecting any claim it cannot tie to a retrieved source span. That is citation-level attribution as a gate, not a courtesy.

The consequence is mechanical. A claim like "Acme integrates with Salesforce" survives only if some retrieved passage states it in extractable form. Marketing copy that says "seamless connectivity" fails the check and gets dropped, even when the worker included it. Your page was retrieved. Your claim was not verifiable. The brand disappears without leaving a trace in the output.

Write claims the way a critic reads them: subject, verb, object, number, one sentence, on a page the crawler can fetch.

3. Entity resolution is the silent merge step, and knowledge graphs set its outcome

When three workers return three descriptions of your product, something has to decide they are the same entity. Systems lean on graph-style resolution: names, domains, and relationships across sources. If your company name is ambiguous, or your product is described differently on your site, G2, and Wikipedia, the merge fails and your evidence splits into weak fragments. Each fragment is below the critic's confidence threshold. Together they would have passed.

Hybrid retrieval and GraphRAG make this sharper. Relationship edges (founded by, integrates with, competes with) carry multi-hop reasoning. A brand with consistent edges gets pulled in on hops it was never directly queried for.

How to see which stage drops you

You cannot debug this from a rank tracker. You need per-prompt, per-engine data: which prompts mention you, where you sit in the answer, which sources get cited, and where a competitor replaces you. LLM Search Console tracks visibility score, share of voice, sentiment, and citations found across ChatGPT, Gemini, Perplexity, Claude, Google AI Overview, and Google AI Mode. When a prompt shows retrieval-like citations to your domain but no brand mention, that is your verifier-drop signature. Plans start at 49 euros per month.

Quick wins for GEO

• Rewrite your top 10 product claims as single-sentence, number-bearing facts a verifier can quote.

• Split your tracked prompts into parent and child variants (pricing, integrations, alternatives) and score each separately.

• Align your company description word for word across your site, LinkedIn, Crunchbase, and review profiles.

• Add Organization and Product schema with sameAs links to make entity resolution unambiguous.

• Audit robots.txt for AI crawlers; a blocked fetch means the critic never sees the evidence.

• Flag prompts where your domain is cited but your brand is unnamed, and fix those pages first.

That gap is why most teams argue about whether LLM visibility pays off. They are looking at a report that cannot see the channel. Fix the report and the argument ends.

Why AI referral traffic deserves its own line in your report

The volume is small next to organic search. The intent is not. A visitor who clicks through from an AI answer has already read a summary of what you do, who you compete with, and why the model picked you. They skip the awareness stage. They arrive with a question half answered.

That changes how you read the numbers. Ten sessions from Perplexity can outweigh a hundred from a broad blog post, because the ten came pre-qualified by a model that explained your value in its own words.

You will only see that pattern if the traffic is separated from everything else.

Where AI referral traffic shows up in your analytics

The sources you can see

Several assistants pass a referrer, so GA4 records them as ordinary referral traffic. Look for these in your source list.

• chatgpt.com

• perplexity.ai

• gemini.google.com

• copilot.microsoft.com

• claude.ai

ChatGPT also tags many outbound links with a utm_source=chatgpt.com parameter, which makes it the easiest to isolate. The others rely on the referrer alone.

The sources you cannot see

Google AI Overviews and AI Mode land inside your Google organic numbers. There is no clean split today. Links copied out of a chat window and pasted into Slack or email arrive as direct. Mobile apps often strip the referrer entirely.

Treat what you can measure as a floor, not a total. If chatgpt.com shows 400 sessions a month, the real number is higher. Plan around that.

Set up AI referral tracking in GA4 in 20 minutes

You do not need a new tool for the first pass. You need one custom channel group.

1. Go to Admin, then Data display, then Channel groups, and copy the default group.

2. Add a channel called AI Assistants and place it above Referral in the order.

3. Set the condition to Source matches regex and paste the pattern below.

chatgpt\.com|perplexity|gemini\.google|copilot\.microsoft|claude\.ai|deepseek

Custom channel groups apply to your historical data, so you get a trend line on day one instead of waiting a month.

Then build a simple Exploration. Rows are landing page. Columns are session source. Values are sessions, engagement rate and key events. Save it.

What to read in the data

Raw session counts are the least useful number on the page. Read these instead.

• Landing pages

• Engagement rate against organic

• Conversion rate against organic

• Source split by engine

Landing pages tell you the most. If AI traffic concentrates on three comparison pages and a pricing page, you now know which content the models trust enough to cite. Those pages are your highest-leverage assets. Keep them accurate, fast and current.

If the engines send people to a blog post from 2023 that quotes old pricing, that is a leak. Fix it this week.

Traffic is a lagging indicator

Here is the part referral reports cannot show. A click happens only after the model has mentioned you and a user has chosen to follow the link. Every prompt where you were left out, or named without a link, produces zero sessions and zero data.

So a flat referral line can mean two very different things. Buyers may not be asking. Or they are asking and the model recommends someone else. Analytics cannot tell those apart.

That is the job of LLM brand visibility tracking. It runs the prompts your buyers type across ChatGPT, Perplexity, Gemini and the other engines, then records whether you appear, in what position, and which sources the model cited. Put that next to your GA4 channel and you can see the full path from mention to click to conversion.

A weekly loop that takes 30 minutes

• Check AI Assistants sessions, week over week

• Note the top three landing pages

• Compare conversion rate to organic

• Pull the prompts where a competitor won

• Assign one fix

One fix a week beats a quarterly audit. Engines re-weight sources constantly, and a page that earned citations in September can lose them by November if nobody is watching.

Do this before Monday

Create the AI Assistants channel group today. It takes less time than this article took to read. On Monday morning you will have a number you did not have on Friday, and a list of landing pages the engines already favor.

If you want the other half of the picture, LLM Search Console shows which prompts mention your brand, where you rank against named competitors, and which sources drive the answers, so the traffic you measure has a cause you can act on.

Subscribe to the Substack newsletter for one tested AI search tactic every week, delivered before your Monday standup.

That blind spot has a cost. A buyer who asks "what are the best tools for X" and gets three competitor names never visits your site, never shows up in your funnel, and never appears in a lost-deal report. Fixing it starts with one question. Which prompts mention my brand, and which ones skip it? Teams that track LLM visibility at the prompt level can answer that in a week.

Why the prompt is the unit that matters

Keywords told you what people searched. Prompts tell you what people want to decide. "CRM" is a keyword. "Which CRM works best for a 20-person agency that sells retainers" is a prompt, and the model answers it with a shortlist.

Your brand can dominate one prompt and vanish from the next one, even when the two look almost the same. That is why a single "are we in ChatGPT" spot check tells you nothing. You need a map.

The four prompt types to track

Most brand-relevant prompts fall into four groups. Cover all four or your picture will skew.

Category prompts

Broad shortlist questions such as "best project management software for agencies". These drive discovery. Appear here and you enter the consideration set.

Comparison prompts

"X vs Y" and "alternatives to X". Buyers use these late in the process. Missing here costs you deals that were nearly closed.

Problem prompts

"How do I stop losing track of client approvals". No product name, just pain. Models answer with a method and then a tool. Being the tool named after the method is a strong position.

Brand prompts

"Is [your brand] any good" and "[your brand] pricing". Here the model describes you directly, so accuracy and sentiment matter more than presence.

How to find which prompts mention your brand

You can do this by hand to learn the shape of it, then automate it once you see the value.

1. Build the prompt set

Take your top 30 commercial keywords and turn each into the question a buyer would ask. Add comparison and problem versions. Pull real phrasing from sales calls, support tickets and community threads. Fifty to one hundred prompts is enough to start.

2. Run it across engines

Run every prompt on ChatGPT, Perplexity, Gemini and Claude. Answers differ by engine, sometimes sharply. A brand that wins on Perplexity can be invisible on Gemini.

3. Log three things per prompt

Were you mentioned. Were you cited as a source. Who else was named. That last column is the one most teams skip, and it is the most useful.

4. Repeat the runs

One run proves little because answers shift between runs. Run each prompt several times and record how often you appear. A mention rate of 40 percent beats a screenshot of one lucky answer.

Reading the results

Sort the sheet three ways and you will see where the work is.

• Prompts where you always appear. Protect these.

• Prompts where you sometimes appear. Cheapest wins.

• Prompts where competitors appear and you never do. Your gap list.

For every gap prompt, check which sources the engine cites. Those pages are your target list. Get listed there, earn a mention there, or publish a clearer answer than the one the model is pulling from.

Why spreadsheets stop working

A 100-prompt set across four engines, run five times each, is 2,000 answers a cycle. Multiply by weekly re-runs and the sheet collapses. This is the point where a dedicated LLM brand visibility tracker earns its place, because it runs the set on a schedule, flags changes and shows share of voice against named competitors. If you want prompt-level data without the manual work, AI visibility tracking is what we built for it.

Three mistakes that distort the data

Testing only branded prompts makes you look great and tells you nothing about discovery. Testing from a logged-in account with chat history skews answers toward what the model already knows about you. And treating one answer as truth ignores that models vary run to run.

What to do this week

Write 25 prompts your buyers would really ask. Run each on two engines. Mark every answer where you are named and every answer where a competitor is. By Friday you will have your first gap list.

Want a new prompt-level breakdown like this every week, with real data from ChatGPT, Perplexity and Gemini? Subscribe to the newsletter below and the next one lands in your inbox.

That blind spot is what a prompt research tool fixes. It shows which questions people ask AI engines in your category, which of those questions name brands, and which ones your competitors already own. If you are a marketer, founder or brand manager deciding where to spend the next quarter, this research comes before content, before outreach and before any AEO budget.

Why keyword research breaks in AI search

A keyword is a compressed guess. "CRM agency" could mean a dozen different intents. A prompt removes the guess. "Which CRM works for a 20-person agency that already lives in Slack and needs client reporting" carries the buyer's whole context. The answer engine reads all of it and returns a shortlist of two or three names.

Two things follow. Volume data is thin or missing, because nobody publishes how often a given prompt gets asked. And one intent shows up in hundreds of phrasings, so a single target keyword stops making sense. You research patterns of questions, not single terms.

This is why LLM visibility work starts with the prompt set. Without one, you are measuring your brand against questions nobody asks.

What a prompt research tool should do

Strip away the marketing and the job has four parts.

It finds prompts you would not have written

Your team writes prompts that sound like your website. Buyers do not. A good tool pulls phrasing from sales calls, support tickets and community threads, then expands each seed into the variations people actually use. Look for tools that show you the messy versions, typos and half-sentences included.

It sorts prompts by intent

"What is prompt research" is a learning question. "Best prompt research tool for an agency" is a buying question. Only the second one moves revenue this quarter. The tool should tag each prompt as informational, comparison or purchase so you can put your effort where shortlists get made.

It shows who wins each prompt

A prompt list without answers is a spreadsheet. The tool should run every prompt across ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews, then record which brands are named and which URLs are cited. Each prompt gets an owner, and you see which ones have none.

It keeps running

One snapshot lies. Ask the same question five times and you can get five different lists. Repeat runs turn an anecdote into a visibility rate, the share of runs where your brand appears at all.

Build your first prompt set in one afternoon

You do not need software to start. You need 40 prompts and two hours.

Start with sales. Ask for the last ten questions prospects raised before they bought, and write them down word for word.

Add support. Pull the ten most common pre-sales tickets. Remove company names, keep the phrasing.

Mine communities. Search Reddit and industry forums for threads in your category and copy the titles of the ones with long replies. Those are prompts in disguise.

Expand every seed three ways. Add a constraint like budget or team size, a comparison against a named rival, and a role such as CFO or agency owner.

Finally, tag each prompt by intent and funnel stage, and delete duplicates.

Signals worth acting on

Once results land, three patterns deserve your attention.

The first is prompts where a competitor appears in every run and you never do. That is your gap list. The second is prompts where you appear but a rival's page gets the citation. The model knows your name yet trusts someone else's evidence, which is nearly always a content problem. The third is prompts where nobody is named at all. That is open ground, and a strong page can claim it faster than any crowded search result.

Together these give you a ranked backlog. It is LLM brand visibility work in practice. Pick the prompts, fix the pages, re-run, and watch the rate move.

Mistakes that waste the budget

Teams treat one run as truth. They chase rank position in a system that reshuffles on every answer. They buy a tool that supplies volume estimates it cannot back up. They research prompts in English when their market buys in Portuguese or German. And they skip competitors, so the report says "you are visible" without saying visible compared to whom. Tracking AI share of voice fixes that last one.

Questions to ask before you buy

Ask where the prompts come from, real user data or model-generated guesses. Ask how many times each prompt runs per engine. Ask whether you can see the raw answer and cited URLs behind every number. Ask whether you can export the full history. A vendor who answers plainly is worth a trial.

Run the check today

Open ChatGPT, Perplexity and Gemini. Type five questions your buyers ask, exactly as they would. Write down which brands appear. If yours is missing from most, you have your first research finding and the business case for tracking it properly.

LLM Search Console turns that manual check into a scheduled prompt set across every major engine, with citations, sentiment and competitor share of voice in one dashboard. Start with a free check of your brand, then subscribe to this Substack for one practical breakdown a week on getting your brand into AI answers.

That row is AI referral traffic. It is small, it is growing every month, and it is almost certainly undercounted. If you run marketing for a company that sells anything researched online, this is the number your board will ask about within a year. Better to own it now.

What counts as AI referral traffic

AI referral traffic is any session that starts when a person clicks a link inside an AI assistant or AI search surface. That means ChatGPT (including its search mode), Perplexity, Gemini, Microsoft Copilot, Claude, Grok, and the citation links inside Google AI Overviews and AI Mode.

It is not the same as being mentioned. A mention is visibility. A referral is a click. Visibility is what you optimise for; referrals are the first hard proof that the optimisation turned into pipeline. You need both numbers, and you need them in the same report.

Why your analytics undercounts it

Three leaks hide most of this traffic.

First, Google. Clicks from AI Overviews and AI Mode carry a google.com referrer, so they land in your organic bucket and look like ordinary search. Search Console now folds AI Mode impressions and clicks into the Web performance report, but it does not break them out. You cannot see them in GA4 at all.

Second, apps. The ChatGPT mobile and desktop apps, Copilot inside Windows and Edge, and in-app browsers often strip the referrer. Those sessions arrive as "direct." Your direct traffic has been quietly rising for eighteen months. Part of that is AI.

Third, naming. Some engines pass a referrer, some pass UTM parameters, and some do both inconsistently. ChatGPT appends utm_source=chatgpt.com to many outbound links; Perplexity passes a clean perplexity.ai referrer; Copilot can show up as copilot.microsoft.com, bing.com, or nothing. If you are not grouping these by hand, GA4 scatters them across four channels.

Build a dedicated AI channel in GA4

This takes about twenty minutes and pays for itself the first time someone asks "is AI sending us anything."

• Open Admin, Data settings, Channel groups, and create a custom channel group. Add a channel named "AI Assistants" with a rule that matches source against a regex such as chatgpt|openai|perplexity|gemini|bard|copilot|claude|anthropic|grok|x\.ai|mistral|you\.com|phind. Place it above Referral and Organic Search so it wins the match.

• Add a second rule for session campaign or source containing chatgpt.com to catch the UTM version.

• Build an exploration with sessions, engaged sessions, key events, and conversion rate, segmented by that channel. Pin it to your weekly dashboard.

• Annotate the date you created the channel so nobody mistakes the jump for growth.

Once it exists, compare the AI channel against organic on two metrics only: conversion rate and pages per session. Volume will look embarrassing for a while. Quality will not.

Attribute the clicks you cannot see

For the leaks above, there is no clean fix, so use proxies that hold up in a board deck.

Track branded search volume alongside AI mention rate. When an assistant names you in a buying question, a share of those people search your brand on Google ten minutes later. A rising branded-search curve that tracks your AI mention curve is the attribution story for the invisible clicks.

Add "How did you hear about us?" as a free-text field on demo and signup forms. Within a quarter you will have a list of answers that say "ChatGPT recommended you." Count them. That is pipeline you can name.

Watch direct traffic to deep pages. Nobody types a 90-character URL to a comparison page. Direct sessions landing on long URLs are, in most cases, copied or clicked from an assistant.

Which prompts send the clicks

Referrals come from citations, and citations come from specific prompts. The question is not "does ChatGPT send us traffic" but "which buyer questions make ChatGPT cite our page instead of a competitor's."

Take the ten pages that already receive AI referrals in your new channel. For each, ask the engines the questions a buyer would type that would plausibly surface that page. Note where you are cited, where a competitor is cited, and where nobody is. Pages that get mentioned but never cited are usually missing one thing: a clear, extractable answer near the top with a stat or a definition an engine can quote.

Re-run that set monthly. Models change their retrieval sources without warning. A page that earned citations in spring can go dark by autumn while your SEO rankings sit exactly where they were.

Report it like revenue, not like traffic

Three numbers belong in your monthly report: AI referral sessions, AI referral conversion rate against the site average, and the count of high-intent prompts where your brand is cited. The first one will be small. The second one is your argument. The third one is your leading indicator, because citations this month become referrals next month.

Tracking that third number by hand works for a few weeks. LLM Search Console runs your prompt set across ChatGPT, Perplexity, Gemini and the rest on a schedule, records which pages get cited, and lines that up against the referral data so you can see which prompts are actually producing clicks.

Do this before Friday

Create the AI Assistants channel group in GA4 today, backdate nothing, and screenshot the conversion-rate comparison after fourteen days. Then subscribe to the newsletter and we will send you one attribution teardown a week, built on real data from real sites.

Search engines handed you a keyword report. AI engines hand you nothing. No console shows which prompts mention your brand, which ones name a rival instead, and which ones skip your category altogether. Marketers who build that map themselves, even in a spreadsheet, make better calls than teams arguing over averages.

Why prompts decide who gets mentioned

Keywords are short. Prompts are full sentences with context. "Which CRM works best for a 20-person agency that sells retainers?" The model reads the team size, the business model and the use case, then picks a handful of brands. Two prompts about the same category can return two different shortlists. You are named in one and absent from the other.

That is why one overall score misleads. A brand can post a healthy mention rate and still miss every prompt containing "under $500" or "for enterprise." Those are the prompts with a budget attached.

Build your prompt set

Start with 40 to 60 prompts. Fewer than that and one odd answer skews the picture. Pull them from four places.

• Sales call notes

• Support and onboarding tickets

• Search queries rewritten as full questions

• Reddit and community threads in your category

Tag each prompt by funnel stage (learning, comparing, buying) and by buyer persona. Ten minutes of tagging saves hours of confusion later.

Write prompts the way buyers type them

Buyers do not type "best CRM software." They type "we outgrew spreadsheets, need something that connects to Gmail, what do people actually use." Keep the vagueness and the constraints. Polished prompts return polished answers that no real buyer ever sees.

Run them across engines, and run them more than once

ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews and AI Mode each pull from different sources. A brand that wins in Perplexity can be missing in ChatGPT. Test every prompt on every engine your buyers use.

Run each prompt at least three times. Answers are probabilistic. One run is an anecdote. Three runs start to look like a signal. For every answer, log whether you were mentioned, where you sat in the list, whether the model cited your domain, and which competitors appeared.

Sort the results into four groups

Won prompts put you early in the answer, ideally with a citation. Protect them.

Shared prompts name you alongside rivals who sit above you. Work on position.

Lost prompts name a competitor and skip you. This is your priority list.

Blind spots return generic answers with no brand named. Nobody owns them yet, and you can.

Find out why you lose a prompt

Open the sources behind the answer. The pattern is usually one of three. A competitor has a comparison page that mirrors the prompt wording. A review site lists them and not you. Or a community thread recommends them by name. Models quote the pages they can reach. If you are missing from those pages, you are missing from the answer.

Check your own site too. Can AI crawlers reach it? Does the page that should answer the prompt state the answer in its first two sentences? Pricing, use cases and integrations buried in a PDF or behind a form are invisible to the model.

Fix the gaps in order of revenue

Take lost prompts with buying intent first. Publish a page that answers the exact question, with the answer up top, a comparison table and named customer examples. Then earn the third-party mentions the model already trusts by pitching the review sites and list pages it keeps citing. Add structured data and plain entity details (name, category, pricing, who it is for) so models describe you correctly. Re-run the prompt set two to four weeks later and compare.

Make it a weekly habit

Models change. Retrieval sources change. A prompt you won in August can flip by October. Re-run the set weekly, flag every prompt that changes status, and report one number to leadership: the count of high-intent prompts where your brand is named. That figure gets budget approved faster than any impression count.

Doing this by hand works for a month. After that it becomes a part-time job. LLM Search Console runs your prompts across engines on a schedule, shows which ones mention your brand, and flags the ones that move.

Start with five prompts today

Open your CRM, find the last five questions prospects asked before they bought, and run them through ChatGPT and Perplexity right now. Write down who gets named. Then subscribe to the newsletter and get one prompt-tracking teardown every week.

That gap exists because keyword research measures what people type into Google. It says nothing about what they ask an AI. A prompt research tool fills that hole. It shows which questions buyers put to ChatGPT, Perplexity, Gemini and Claude, and which of those questions name your competitors instead of you.

What a prompt research tool does

Think of it as keyword research for AI. You start with a topic, and the tool returns the prompts people use around it. Short ones, long ones, comparisons, "best of" lists, problem statements. Then it runs those prompts against the major engines and records what comes back.

The output is not a ranking. It is a map of answers. Which brands are named, which pages are cited, how the answer is framed. That map is the raw material for LLM brand visibility work.

Why keywords stop working here

Search queries are compressed. "crm small business" is three words. The same buyer asks an AI, "What CRM works for a ten-person agency that already uses HubSpot for email?" Same intent, forty words of context.

Those long prompts are where the shortlist gets built. The model reads every constraint and filters. Brands that fit the constraints get named. Brands that only match the head term do not.

Two more differences matter.

• Volume is hidden. No engine publishes how often a prompt is asked.

• Answers drift. The same prompt can return different brands next week.

So you cannot pull a list once and call it done. Prompt research has to repeat.

How to run prompt research yourself

Start from buyer language

Pull the phrases from sales calls, support tickets and G2 reviews. Buyers describe problems in their own words, and those words are what they paste into an AI. Write down 20 of them before you open any tool.

Expand into four prompt types

For each topic, draft one prompt in each group.

• Category prompts. "What are the best tools for X?"

• Comparison prompts. "X vs Y for a mid-size team."

• Problem prompts. "How do I stop losing deals to Z?"

• Brand prompts. "Is [your brand] any good?"

Fifty to one hundred prompts is enough to see patterns.

Run them across engines

Run every prompt on at least four engines. ChatGPT and Perplexity often disagree. Gemini leans on Google's index. Claude favors cleaner, better-structured sources. A brand that wins on one and vanishes on another has an AI visibility problem you would never see in a single-engine check.

Log three things per prompt

Record whether you were mentioned, whether you were cited, and who else appeared. Those three fields give you a mention rate, a citation rate and a share of voice against competitors.

What to look for in a tool

Skip the demo gloss and test for these.

• Multi-engine coverage

• Scheduled re-runs

• Competitor tracking

• Source-level citation data

• Prompt grouping by intent

The last one is easy to overlook. A flat list of 300 prompts is noise. Grouped by intent, it tells you which funnel stage you are losing.

Also ask where the prompt list comes from. If the tool only tracks prompts you type in, you are measuring your own assumptions. A good one suggests prompts you did not think of.

Turn the data into work

Sort prompts by one rule. Competitors appear, you do not. Those are your gap prompts.

For each gap, open the sources the engines cite. That list is your to-do list. Get your brand onto those pages, or publish something that answers the prompt better than they do, in the first 150 words, in the buyer's own phrasing.

Then re-run the set in two weeks. If the mention rate moved, keep going. If it did not, change the page, not the prompt.

Where to start today

Write ten prompts your best customer would ask. Run them in ChatGPT, Perplexity and Gemini. Count how many name you.

That number is your baseline, and most teams have never seen it. If you want it tracked automatically across engines every week, LLM Search Console does exactly that.

New prompt research tactics land in the Substack newsletter every week. Subscribe below and the next one arrives in your inbox.

Security teams treat it as a sandbox and least-privilege issue. Marketing teams don't treat it as their issue at all. The overlap is where the interesting failures live, and where LLM Search Console earns its keep.

1. Injection is an answer-quality bug, not just a security bug

Classic injection demos exfiltrate secrets. The quieter variant just biases the output. Any page an agent retrieves is untrusted input concatenated into the same context as the system prompt. The model has no hard boundary between "data" and "instructions," so text in a retrieved page can steer the summary, the ranking, or the recommendation.

For a brand, this cuts both ways. Competitors, affiliates, or scrapers can plant steering text on pages that get cited. And your own pages can be misread when a third-party widget, review embed, or comment block injects text you never wrote. If you only monitor your own site, you miss the layer where the answer is actually shaped.

2. Citation-level attribution is your only forensic tool

When an answer says something wrong about you, the useful question is which source produced the claim. Engines that expose citations let you trace a sentence to a URL. That mapping is the difference between "the model hallucinated" and "this specific page, updated last Tuesday, now says something different."

Track three fields per prompt: the claim, the cited domain, and the position of your brand in the answer. When a claim shifts and a new domain appears in the citation set at the same time, you have a candidate poisoning or a candidate competitor content play. Either way you have a URL to act on. A mention rate without citations tells you that something changed, not why.

3. Observability without replay is a dashboard, not evidence

Agent teams learned that traces and replay are the only way to debug non-deterministic systems. The same rule applies to visibility tracking. One run of one prompt is an anecdote. Answers vary by engine, by session, and by retrieval order.

Run each tracked prompt repeatedly, store the full answer and citation list, and diff across days. Now an injection looks like a pattern: a sudden brand swap that appears in a cluster of runs and disappears when the source page changes. Without stored runs you can't distinguish that from ordinary sampling noise, and you can't show a client or a legal team what happened.

4. Governance means owning the tool result, not just the tool

Agent governance guidance focuses on scoping credentials and gating tool calls. Useful, but it says nothing about what comes back. Every retrieved page is a tool result, and tool results are the supply chain. Organizations building agents will filter and sandbox that content over time. Until then, your brand's representation depends on retrieval pipelines you don't control.

What you control is detection speed. If a bad claim lives in answers for six weeks before anyone notices, it gets absorbed into cached summaries, secondary citations, and sales conversations. If you catch it in a day, it's a ticket.

Quick wins for GEO

• Log the claim, not just the mention. Store the sentence about your brand per engine, per prompt, per day, and alert on semantic changes.

• Diff citation sets. A new domain entering the top sources for a money prompt is a lead worth investigating within 24 hours.

• Audit third-party embeds. Review widgets, scripts, and UGC blocks that render text on your pages. Anything you didn't write can be read by an agent as if you did.

• Publish machine-readable ground truth. Clear pricing, product facts, and comparison pages with structured data give engines a strong source to contradict a planted claim.

• Replay before you escalate. Confirm a shift across multiple runs and engines before filing a takedown or rewriting content.

Tracking this across ChatGPT, Perplexity, Gemini, and Copilot by hand is not realistic. LLM Search Console stores answers and citations over time, so you can see when the story about your brand changes and which source changed it.

Prompt volume is the count and pattern of real user queries that pull a brand into an AI-generated answer, tracked at scale instead of guessed at from a handful of test questions typed into ChatGPT the night before a board meeting. It works the way search volume worked for Google, except the query never touches your analytics, and the "result" is a paragraph the model wrote instead of a link someone clicked. Teams that used to eyeball ten sample prompts are now expected to show a documented set running into the thousands, refreshed on a schedule, broken out by engine.

That expectation didn't come from nowhere. Semrush's 126-million-prompt sample broke visibility down by sector, and the spread was wide. In News and Media, the top three brands captured 82.9% of all visibility inside AI answers. In Finance, the top three held just 41.4%. A finance brand chasing "top three" the way a media brand does is solving the wrong problem, because the opportunity in a fragmented sector looks nothing like the opportunity in a concentrated one. Buyers now ask vendors to show sector-adjusted numbers before they sign, not a flat visibility score pulled from someone else's industry.

What Prompt Volume Actually Measures, and What It Doesn't

Prompt volume by itself only counts how often a query surfaces your brand somewhere in the answer. It says nothing about whether you were named favorably, cited with a link, or buried under four competitors in the same paragraph. Semrush's data makes that distinction hard to ignore: ChatGPT cites an average of 15 sources per response, Gemini cites roughly 3, and on Gemini the brands mentioned in the answer text and the domains actually cited overlap only 30% of the time. A brand can be named constantly and still be invisible in the citation layer. The reverse happens too.

Treat the raw number as a filter, not a scoreboard. It tells you which prompts are worth digging into further. The digging is where mention rate, citation rate, and real LLM brand visibility benchmarking against named competitors actually live.

Building a Prompt Set That Reflects Real Buyers

Start from language, not keywords

Most teams build their first prompt set the way they built keyword lists a decade ago. Guess ten phrases, run them once, screenshot the results for a slide. That produces a vanity number and nothing repeatable. A usable prompt set starts from the language buyers actually use when they ask an AI model for a recommendation, not the head terms a brand wishes it ranked for. Pull real questions from sales call transcripts, support tickets, and the forums where your category gets discussed. Then sort by buying stage. Awareness prompts, comparison prompts, decision prompts. Each behaves differently across engines and needs its own tracking line.

Track the number by engine, not as one blended score

A blended visibility score hides more than it shows, because ChatGPT, Gemini, Google AI Mode, AI Overviews, Perplexity and Claude pull from different indexes and cite in different ways. Semrush found only 36 global brands held top-100 visibility across all four platforms it studied, every month, for the full period. Most brands that look strong on a single combined number are actually strong on one engine and close to absent on the rest. Report each engine on its own first. Blend only once you know which one is carrying the average.

Re-run it on a schedule, because the answer moves

A single snapshot of prompt volume is out of date by the time it reaches a slide deck. Model updates, retrieval changes and competitor content all shift which prompts surface your brand, and Semrush's own January-to-April comparison showed real movement inside a single quarter. Monthly re-runs, at minimum, are what separate an actual LLM visibility program from a one-time audit somebody ran to answer a question from the board.

Where Teams Get This Wrong

The same mistakes show up in almost every prompt volume rollout. Treating one run as truth. Comparing raw counts across sectors with wildly different concentration instead of a sector-adjusted benchmark. Reporting mentions while ignoring the citation layer, which is where a recommendation is actually earned. Fix the reporting cadence and the sector context first. The tooling matters less than the discipline behind it.

What To Do With This Number Monday Morning

Pull ten real buyer questions from your last twenty sales calls. Run them through ChatGPT, Gemini and Google AI Mode this week, by hand if that's what it takes. Note who gets named, who gets cited, and where your brand is missing entirely. That is your baseline. Everything Semrush's 126-million-prompt study found applies just as much to ten prompts as it does to ten thousand.

Prompt volume isn't a number you buy a tool for and forget about. It moves every time a model updates, and the brands checking it monthly are the ones showing up when their buyers ask an AI which vendor to pick.

Want the sector benchmarks, the engine-by-engine breakdowns, and what's moving in ChatGPT, Gemini and Perplexity before it shows up anywhere else? Subscribe to the newsletter below.

That habit is now baked into how B2B buying starts, and it changes what AI visibility actually has to mean. A brand that ranks well in ChatGPT but disappears from Perplexity isn't halfway visible. It's invisible to whichever slice of the market happened to ask the other engine first.

Most teams that started tracking AI mentions in 2025 built their process around a single engine, usually ChatGPT, because it had the largest user base and the most obvious starting point. That approach made sense when generative search was new and buyers had one default habit. It doesn't hold up anymore.

The question buyers ask before they ask you

Marketers shopping for an LLM Visibility tool almost always open with the same question: how many engines does this actually cover? Not because more is inherently better, but because they've already noticed the gap themselves. They asked ChatGPT about a category, got one set of brands back, asked Perplexity the same question, and got a noticeably different list.

That divergence isn't a bug in any one engine. Each model draws from different training data, different retrieval sources, and different citation logic. ChatGPT leans on its training corpus plus browsing. Perplexity is built around live citations and source transparency. Gemini pulls heavily from Google's index and Search-adjacent signals. Copilot folds in Bing's web graph. Grok pulls from X in ways none of the others do. Ask the same question five different ways across five engines, and you'll get five different answers about who the credible players in a category are.

What single-engine tracking actually misses

A brand watching only ChatGPT can look strong for months while quietly losing ground somewhere else. Three patterns show up constantly once teams start comparing engines side by side.

Coverage gaps by engine. A brand cited consistently in Gemini answers, thanks to strong Search rankings and structured data, can be nearly absent from Perplexity if its content lacks the citable, source-dense format that engine favors.

Sentiment gaps by engine. The same brand gets called an industry leader in one engine's answer and a budget option or legacy player in another, depending on which sources each model weighted.

Competitor gaps by engine. Some engines surface newer, smaller competitors more readily because they draw from fresher web content or social signals. A brand tracking only the incumbent battleground on one engine can miss a competitor quietly winning share of voice somewhere else entirely.

Any one of these, caught late, is a rebrand-level problem. Caught early, it's a content fix.

A simple framework for multi-engine tracking

Teams that get this right treat engine coverage as a checklist, not an afterthought.

Start with the big four: ChatGPT, Perplexity, Gemini, and Copilot. Together they cover the overwhelming majority of AI-assisted search and research traffic right now. Add Grok if your audience skews toward X-native industries like crypto, media, or politics.

Run the same prompt set across every engine on a fixed schedule, not just once. Model updates ship constantly, and an engine that ignored your brand last month may cite it heavily this month after a retraining pass or an index refresh.

Compare, don't just collect. Raw mention counts per engine tell you less than the gap between engines. If your visibility on Gemini is triple your visibility on Perplexity, that gap is the actionable signal, not either number alone.

Track competitors on the same prompt set. A brand's visibility score means little in isolation. What matters is share of voice against the two or three companies buyers are actually comparing you to.

Why this matters more now than it did a year ago

Generative engines have stopped being a novelty channel and become a default research habit for B2B buyers, especially at the top of the funnel, where a weak first impression is expensive to undo. A buyer who gets a lukewarm answer about your brand from their first AI query rarely goes and checks a second source to confirm it. They move on to whichever company the engine described with more confidence.

Multi-engine coverage is also where budget conversations are heading. Teams that can show a CMO exactly where visibility breaks down by engine, and which competitor is winning the gap, are the ones getting next quarter's content budget approved. "We rank well in AI search" is a vague claim. "We're cited in 80% of Gemini answers for this category but only 20% of Perplexity answers, and here's the content gap causing it" is a plan.

Where to start this week

Pick your top ten buyer-intent prompts, the actual questions a prospect would ask about your category. Run each one across ChatGPT, Perplexity, Gemini, and Copilot manually if you have to. Write down which brands show up, in what order, and with what tone. That one afternoon of manual auditing will tell you more about your real LLM Brand Visibility than a single-engine dashboard ever will.

Once the gaps are visible, decide which one costs you the most and fix that first. Multi-engine tracking is a discipline built over quarters, not a dashboard you check once and forget.

If you want this audit automated across every major engine, updated daily instead of once a quarter, subscribe below. We'll walk through the exact prompt sets and tracking setup in the next issue.

Here are three connections between vector-only RAG, single-shot prompting, and brand visibility that rarely get discussed, and what to do about each. Track the results with LLM Search Console instead of guessing.

1. Embeddings blur proper nouns, so you get swapped for a competitor

Dense embeddings are good at meaning and bad at identity. "Acme CRM for agencies" and "Globex CRM for agencies" land almost on top of each other in vector space. The brand token is a rounding error next to the category semantics. A vector-only retriever will happily return your competitor's chunk for a query that named you.

This is why hybrid retrieval exists: a lexical pass (BM25 or similar) keeps exact tokens like brand names, SKUs, version numbers and plan names from dissolving into the category average. The implication for GEO: exact-match consistency matters again. If your product is "Acme CRM" on the homepage, "AcmeCRM" in docs and "Acme Platform" on G2, you split your lexical signal three ways and hand the dense side a fuzzy entity to confuse.

The symptom you will see in monitoring: an engine describes your feature set accurately but attributes it to someone else, or names you in a list with the wrong pricing tier. That is a retrieval identity failure, not a content quality problem.

2. Chunking orphans your brand from its own claims

Retrievers do not read pages. They read chunks, typically a few hundred tokens split on headings or length. Marketing copy loves pronouns and implied subjects: "It integrates with 40 tools. Our platform cuts setup time in half." Once that paragraph is chunked away from the H1, the entity is gone. The chunk says something integrates with 40 tools. The model cannot cite you for a claim that does not contain your name.

Call it chunk orphaning. It is the quiet reason brands with strong pages still get paraphrased instead of named. The fix is boring and effective: restate the entity inside every section that carries a claim worth citing. "Acme CRM integrates with 40 tools" survives any chunk boundary. "It integrates with 40 tools" does not.

3. Single-shot prompt tracking is the same mistake as single-shot prompting

Engineers dropped single-shot prompt engineering because one phrasing is a sample of one: brittle, non-reproducible, sensitive to wording. Yet many teams still measure AI visibility by typing one query into ChatGPT once and screenshotting the answer. That is single-shot measurement, and it inherits every flaw of single-shot prompting.

Modern engines decompose a prompt into several sub-queries (query fan-out), retrieve for each, and merge. Your brand might win the "best CRM for agencies" branch and lose the "CRM pricing comparison" branch that fires underneath it. One manual check cannot see that. You need the same prompt set run repeatedly across models, markets and phrasings, with visibility scored as a rate rather than a yes/no.

That is exactly what LLM Search Console does: scheduled scans across ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews and AI Mode, a visibility score per prompt and per model, competitor gap analysis, and a full citation log showing which URLs and snippets each engine actually pulled. The citation log is the closest thing you get to seeing the retriever's output from the outside.

4. Why this matters more as engines get "smarter"

Reranking and grounding layers are getting stricter about verifiable, attributable passages. That raises the bar on the three failure modes above. An ambiguous entity loses at the lexical stage. An orphaned chunk loses at the grounding stage because it cannot support a named claim. And a brand measured with one-off screenshots never learns which stage it lost at. Hybrid retrieval did not make GEO easier. It made it more mechanical, which means it is fixable if you instrument it.

Quick wins for GEO

• Lock one canonical brand string and product names across site, docs, schema, review sites and press. Treat variants as bugs.

• De-orphan your claims: every H2 section with a stat, feature or price should name the brand explicitly.

• Keep exact tokens in plain text: SKUs, plan names and prices in HTML, not only in images or JS-rendered tabs.

• Write chunk-sized answers: self-contained 80 to 150 word blocks under question-style headings.

• Track prompt sets, not prompts: 5 to 10 phrasings per intent, scanned on a schedule, per model and market.

• Audit citations weekly: if an engine cites a competitor's page for your feature, you have an identity or orphaning problem.

Stop guessing which retrieval stage drops your brand. Start tracking with LLM Search Console, plans from 49€/month.

That gap is now expensive. DeepSeek's chat app passed 350 million monthly web visits by March 2026 and claims a user base of over 130 million across more than 150 countries, with a footprint that skews toward markets where ChatGPT is expensive, restricted or simply not the default. Every AI visibility tool roundup published this year lists DeepSeek and Mistral coverage as a comparison line, because buyers started asking for it. If your LLM brand visibility tracking has never run a prompt through DeepSeek, you do not know what a large and growing slice of your market is being told about you.

What makes DeepSeek different from the engines you already track

DeepSeek is not ChatGPT with a different logo. Three things change how your brand shows up inside it.

The models are open weight. DeepSeek publishes its V3 and R1 families under permissive licenses, which means the same model that answers questions on deepseek.com also answers questions inside hundreds of third-party apps, enterprise deployments, and hosted variants from Western providers. Perplexity hosted its own version of R1. Cloud providers offer it as a managed endpoint. When you measure your brand in DeepSeek, you are measuring a model family, not a single product, and its answers propagate further than the front-end traffic suggests.

Search is a toggle, not a default. The DeepSeek chat interface has a Search mode that pulls live web results into the answer and a non-search mode that answers from training data alone. Those two modes can describe your brand completely differently. With search on, you compete for citations the way you do in Perplexity. With search off, you are relying on whatever the model learned about you during training, which for a company that was small two years ago may be almost nothing.

The training corpus and the audience are not the same as OpenAI's. DeepSeek's models are trained heavily on Chinese-language and multilingual web content, and the product's usage is concentrated in Asia, Latin America, the Middle East and parts of Europe. A brand with strong English-language authority and thin coverage elsewhere often finds that DeepSeek recommends a regional competitor it has never heard of.

What brand managers get wrong about DeepSeek visibility

The most common mistake is assuming DeepSeek is a niche developer tool. It was, for about three weeks in January 2025. Then it hit the top of the App Store in dozens of countries and stayed in the top tier of consumer AI apps for the year that followed. The second most common mistake is assuming that if you rank well in ChatGPT, you will rank well in DeepSeek. Cross-engine consistency in AI answers is low. Teams that track multiple engines with a proper LLM visibility platform routinely find their share of voice differs by 20 to 40 points between engines for the same prompt set.

A few patterns show up when you compare DeepSeek answers to ChatGPT answers for the same B2B prompts.

• DeepSeek leans harder on documentation. Product docs, GitHub READMEs, technical comparison pages and structured pricing tables get pulled into answers more often than blog posts or press coverage.

• DeepSeek cites fewer sources per answer in search mode, so the competition for each citation slot is tighter.

• DeepSeek is more likely to surface open-source alternatives and lower-cost options when a prompt has a commercial intent, which means premium-positioned brands can disappear from "best tool for X" answers they win comfortably elsewhere.

None of this is a reason to panic. It is a reason to measure.

How to build a DeepSeek visibility baseline in one week

You do not need a new strategy to start. You need the same discipline you apply to ChatGPT tracking, pointed at a different engine.

Start with the prompts you already track. Take your top 30 to 50 buyer prompts, the ones that map to real purchase intent for your category, and run them in DeepSeek with search on and with search off. Log both. The delta between the two tells you whether your problem is training-data awareness or live retrieval.

Run each prompt more than once. AI answers vary between runs. A single prompt that mentions a competitor is an anecdote. A mention rate across ten runs is a number you can act on.

Record three things for every run. Whether your brand was mentioned. Whether a URL you control was cited. Which competitors appeared and in what order. That gives you mention rate, citation rate and share of voice, the same metrics you use everywhere else.

Add the languages your buyers actually use. If you sell into Brazil, Germany or Indonesia, run the prompt set in Portuguese, German and Bahasa. DeepSeek's multilingual behaviour is one of its biggest differences from ChatGPT, and English-only tracking will hide it.

Then compare. Put DeepSeek share of voice next to ChatGPT, Gemini and Perplexity for the same prompts. The engines where you are weakest are where your next quarter of content and PR work should go.

What actually moves DeepSeek visibility

Once you have a baseline, the levers are familiar, but the weighting is different.

Documentation and structured pages carry more weight than they do in ChatGPT. If your pricing, integrations, comparison and technical spec pages are thin or hidden behind forms, DeepSeek has nothing to retrieve. Make them public, make them plain, and make them specific.

Third-party technical coverage matters. GitHub, Stack Overflow, Hacker News, developer-focused review sites and regional tech media get retrieved and cited. A brand that only invests in English-language PR and G2 reviews is under-represented in the corpus DeepSeek draws from.

Localised content is not optional. A Portuguese or Spanish version of your comparison page is not a nice-to-have for DeepSeek visibility. It is often the only reason you show up in a Portuguese or Spanish answer.

Keep AI crawlers allowed. Check your robots.txt and CDN rules. Several brands blocked all AI bots in 2024 during the scraping backlash and never revisited the decision. If DeepSeek's retrieval cannot reach you, search mode will cite whoever it can reach.

Track the open-weight spread. Because DeepSeek's models run inside other products, a change you make that improves your DeepSeek answers tends to improve answers in every app built on the same weights. That is a return most teams never see because they never measured the source.

The multi-engine question your CFO will ask

At some point someone senior will ask why you are spending time on a Chinese AI app when the company's customers are in North America. The honest answer has two parts. First, your customers are not all in North America, and the ones who are not are increasingly asking DeepSeek. Second, the models are open weight, so DeepSeek's view of your brand is already inside tools your North American customers use without knowing it.

The cheaper way to answer the question is with a number. Show the DeepSeek share of voice for your top 20 prompts next to the ChatGPT figure. If they match, you have confirmed you are covered. If they do not, you have found a channel where a competitor is winning by default because nobody on your side was looking. Either outcome is worth the week it takes to find out.

Start with one prompt set, two modes, one engine you have been ignoring

Pick your top 30 buyer prompts. Run them in DeepSeek, search on and search off, three times each. Compare the share of voice to what you already know from ChatGPT. That single exercise tells you whether DeepSeek is a gap or a strength, and it costs an afternoon.

If you would rather not run it by hand, LLM Search Console tracks brand mentions, citations and share of voice across ChatGPT, Gemini, Perplexity, Claude, Copilot and DeepSeek from a single prompt set, so the cross-engine comparison is one view instead of six spreadsheets.

Subscribe to the newsletter to get the next breakdown in your inbox. We cover one AI visibility problem a week, with the data and the fix.

That row is now one of the most contested pieces of real estate in ecommerce, and it is filled by a system that behaves nothing like a search engine results page. Estimates put ChatGPT at roughly 50 million shopping queries per day. Product cards are organic, there is no paid placement, and OpenAI states that recommendations are chosen by relevance to the question rather than by who pays. That sounds like a level playing field. In practice it rewards the brands that treat LLM brand visibility as a measurable channel, and punishes the ones still waiting to see if this matters.

What ChatGPT Shopping is, and what it stopped being

The feature launched in 2025 with two halves. Discovery, where ChatGPT recommends products inside a conversation, and Instant Checkout, where the buyer could complete the purchase without leaving the chat, using the Agentic Commerce Protocol OpenAI built with Stripe.

The checkout half did not last. Merchant adoption was thin, and Walmart reported that in-chat conversion ran about three times lower than sending the buyer to its own site. In March 2026 OpenAI pulled Instant Checkout out of the main experience and moved it into the Apps ecosystem. ChatGPT Shopping today is a discovery-and-redirect model. The model picks products, shows cards, and hands the buyer off to your product page.

That change matters more than it looks. It means the entire commercial value of ChatGPT Shopping sits in one question. Do you get on the card or not. Nothing after the click is any different from a Google Shopping referral. Everything before the click is new.

Where the cards come from

ChatGPT does not crawl your catalog the way Googlebot does and infer what you sell. For the cards themselves, it leans on structured product data, and the single most important source is a merchant product feed. Most brands are pointing ChatGPT at the same feed they already maintain for Google Merchant Center, and OpenAI accepts feed updates as often as every fifteen minutes.

Three things get you dropped before relevance is even considered.

Missing or wrong GTINs. If the model cannot resolve your SKU to a canonical product identity, it cannot match it to the buyer's question with confidence, so it picks a competitor it can match.

Stale price or availability. A card that shows $349 when your site says $399, or "in stock" when you are back-ordered, is a bad experience for the buyer and OpenAI filters aggressively for it.

Thin attributes. Dimensions, materials, compatibility, and use-case fields are what let the model answer "fits a small apartment." A feed with a title, a price, and an image gives the model nothing to reason with.

Feed hygiene is the entry fee. It does not win the card on its own.

How the model decides which four products to show

Once the feed is clean, the model is choosing among hundreds of eligible products for each query. This is where ChatGPT Shopping stops looking like Google Shopping and starts looking like every other AI answer.

The model composes a shortlist from what it knows about your product from training data and what it retrieves at query time: your product page, reviews on third-party sites, comparison articles, Reddit threads, and editorial roundups. Then it explains its pick in a sentence. "Chosen for its compact footprint and strong reviews on stability." That explanation is generated from the sources it found. If those sources do not say anything specific about your product, you are not chosen, because the model has nothing to say.

A few patterns show up consistently when you look at which products win cards.

Products with a clear, repeated positioning across sources. If your site, your Amazon listing, a YouTube review, and a Wirecutter-style roundup all describe the same product the same way, the model has a stable entity to recommend.

Products with specific, extractable claims. "Holds 300 lb, 27-inch deep, assembles in 20 minutes" beats "premium quality, built to last."

Products that appear in independent comparison content for the exact intent the buyer expressed. The model is answering "best X for Y." It looks for pages that already answer "best X for Y."

Notice what is not on that list. Domain authority in the Google sense. Backlink count. Page speed. Those still affect whether your page gets retrieved, but they do not decide whether your product gets the card.

The measurement gap most ecommerce teams have

Here is the uncomfortable part. Almost no ecommerce team can currently answer the question "for our top 50 shopping intents, how often do we appear on a ChatGPT product card, and who appears instead."

Google Merchant Center tells you impressions and clicks. ChatGPT tells you nothing. Referral traffic in GA4 shows you the buyers who clicked, not the queries where you were never shown. So teams either assume they are fine, or they run a handful of prompts by hand, see a competitor once, and reorganize the quarter around a single sample. AI answers vary from run to run and from user to user. One prompt is anecdote. A rate across many runs is data.

Treat this like any other channel and build the baseline first.

Write the prompts buyers actually type. Not "standing desk" but "standing desk under $500 for a small apartment," "quietest electric standing desk," "standing desk that works with a treadmill." Thirty to fifty of these across your top categories.

Run each one repeatedly, in ChatGPT and in the other engines that now show product answers, and log which brands land on the card, in what position, and what reason the model gave.

Score yourself. Visibility rate per prompt. Share of card slots against named competitors. The reasons cited for the winners, because those are the attributes your feed and your content are missing.

This is exactly what LLM visibility tracking exists to automate, and it is the difference between guessing and knowing where the gaps are.

What to fix, in order

Once you can see the gaps, the work is more concrete than most AI search advice.

First, the feed. Complete GTINs on every SKU. Price and availability synced at the tightest interval your platform allows. Every attribute field filled, especially the ones that map to the qualifiers in your buyer prompts. Size, weight, compatibility, material, noise, assembly time, warranty.

Second, the product page. Put the plain facts in the first two hundred words, in sentences a model can lift. Add Product schema with the same values as the feed. Kill the discrepancy between what the page says and what the feed says, because the model sees both.

Third, the third-party layer. Pull the sources ChatGPT cites when it recommends your competitors. That list is your outreach target. Get into the roundups, the community threads, and the review sites that the model is already pulling from for your category. A single credible independent page that names your product for the exact intent will outperform a month of on-site tweaks.

Fourth, re-measure weekly. Feeds change, engines re-weight, competitors publish. A card you won in September can be gone by November if nobody is watching.

Start with one category this week

Pick the product category where a lost sale hurts most. Write twenty buyer prompts for it. Run each one in ChatGPT five times and write down who gets the card and why. That single spreadsheet will show you, in an afternoon, whether ChatGPT Shopping is sending your buyers to someone else.

If you would rather not build the spreadsheet, LLM Search Console runs those prompts continuously across ChatGPT, Perplexity, Gemini and the rest, tracks your share of product recommendations against named competitors, and shows which sources the engines are citing so your feed, content, and PR teams know exactly where to aim.

Subscribe to this newsletter for a weekly breakdown of what is moving in AI search, which shopping tactics are holding up, and what the engines changed since the last issue.

Token economics for agents is usually discussed as an infrastructure problem: prompt caching, model routing, context compaction, per-run cost caps. Read it as a visibility problem instead and three under-discussed links show up between cost controls and which brands get named.

1. Prompt caching freezes the shortlist

Production agents cache the prefix of the prompt - system instructions, tool definitions, and increasingly the retrieved context for a recurring task. A cached prefix is cheaper by an order of magnitude, so pipelines are built to reuse it for as long as it stays valid.

The consequence for brands: once a set of retrieved sources lands in a cached prefix, that set gets reused across many user turns without a fresh retrieval. If you were in the fetch that got cached, you are in every answer until the cache expires. If you were not, no amount of new content moves you in until the next cold retrieval. Caching turns visibility from a per-query lottery into a per-cache-window state. Which is why the same three competitors keep showing up for a week and then the list reshuffles: that is a cache boundary, not an algorithm update.

2. Routing decides whether retrieval happens at all

Cost ceilings push teams toward routers: a cheap classifier looks at the query and decides whether it needs the expensive path (retrieval, tool calls, a larger model) or can be answered from parametric memory. "Best CRM for a 10-person agency" is exactly the kind of question a router will label as answerable without retrieval, because the base model already has opinions.

When the router skips retrieval, only brands already in the weights get named. Your fresh comparison page never enters the picture. When the router does trigger retrieval, tool results are typically summarised before being handed to the answer model, and the summariser is the cheapest model in the stack. A cheap summariser working under a token cap keeps the first three sources with clean, extractable claims and drops the rest. Being fetched is not the same as being kept.

3. Your page has a token price, and expensive pages get truncated

Agents read pages through a fetch tool that returns text up to a limit. A comparison page that costs 14,000 tokens to ingest - navigation, cookie boilerplate, three sticky banners, a 2,000-word preamble before the first product name - gets cut at the limit before the agent reaches the part that mentions you. A 900-token page with the claim in the first paragraph is read in full, summarised cheaply, and cited.

This is the same extractability argument GEO people already make, but with a mechanism attached. It is not that models "prefer" concise content. It is that the fetch budget, the summarisation budget, and the context budget are three separate caps, and a heavy page fails all three. Front-load the claim, strip the chrome, and your cost-per-citation drops for every agent that reads you.

4. Measure the ceiling instead of guessing about it

You cannot see a vendor's router thresholds or cache TTLs. You can see their output. Run the same prompt set on a schedule and watch two things: how often your brand appears at all (retrieval happening or not), and how often it appears alongside the same fixed set of competitors (a cached shortlist). Sudden reshuffles with no content change on your side are cache boundaries. Prompts where you are cited on Perplexity but absent on ChatGPT are routing differences, since one engine retrieves by default and the other does not.

This is what LLM Search Console is built for: scheduled scans of your prompts across ChatGPT, Gemini, Perplexity, Claude, Google AI Overview and AI Mode, with a per-prompt visibility score, the competitor set that appears next to you, and the exact URLs each engine cited. Daily scans on the Pro plan are the resolution you need to see cache windows; the citations log tells you which of your pages are cheap enough for agents to keep.

Quick wins for GEO

• Put the claim in the first 300 tokens. Brand name, category, and the one-line differentiator before any navigation or intro copy.

• Publish a lightweight canonical page per prompt cluster. Under 1,500 words, no interstitials, plain HTML headings. This is the page you want cached.

• Refresh sources on a cadence, not on impulse. If a fixed competitor set persists for days, wait for the reshuffle and make sure your updated page is live before it.

• Track per engine, not "AI". Retrieval-first engines and memory-first engines fail you for different reasons. Fix them separately.

• Scan daily. Weekly sampling cannot distinguish a cache boundary from a real loss of visibility.

Every agent answer is a cost-optimised artifact. Brands that are cheap to fetch, cheap to summarise, and already in the cached shortlist get named. Everyone else pays the ceiling.

That is the problem with the phrase. It has been borrowed from machine learning engineering, where it means shrinking and speeding up a model, and dropped into marketing meetings, where it means something closer to "why does Perplexity recommend our competitor and not us." Two disciplines, one label, and the engineering one owns the search results. This article is about the marketing one. Call it LLM brand visibility optimization if you want to be precise. The goal is the same either way: when a buyer asks an AI assistant a question you should win, your brand shows up in the answer, described accurately, ahead of the alternatives.

Two definitions, one search result

It helps to be blunt about the split, because you will hit it the moment you brief an agency or buy a tool.

The engineering definition covers quantization, pruning, distillation, KV caching, batching, and everything else that reduces the cost of serving a model. It is a real field with real budgets. It is also entirely irrelevant to whether an LLM mentions your company.

The marketing definition covers the inputs an LLM draws on when it composes an answer about your category: your site, third-party pages that describe you, the structured facts an engine can extract, and the live web results a retrieval-augmented engine pulls at query time. Optimizing those inputs so the model surfaces you more often, more favorably, and more accurately is the work. It overlaps with what people call generative engine optimization, answer engine optimization, and LLM SEO. Different names, same job.

One practical consequence. When you hire for this, or search for a vendor, use the marketing terms. "LLM optimization consultant" returns ML engineers. "AI search visibility" returns the people you actually need.

Why LLMs choose one brand over another

Models do not rank pages. They generate text conditioned on two sources: what they absorbed during training, and what they retrieve at answer time. Optimization has to address both, and they move at different speeds.

Training data is slow. The model's baseline sense of your category was set months ago from the open web, forums, documentation, review sites, and press. If your brand was thin or absent in that corpus, the model's default answer will skip you. You cannot patch that retroactively. You can only make sure the next training run finds more, and better, material.

Retrieval is fast. ChatGPT with browsing, Perplexity, Gemini, Copilot, and Google AI Overviews all fetch live pages for many queries and cite them. A page you publish this week can be cited next week. This is where most of the near-term gains sit, and it rewards a specific kind of content: direct, extractable, entity-rich, and hosted on domains the engine already trusts for that topic.

Three signals show up again and again when you look at which brands get picked.

Consistency across sources. If your homepage, your G2 profile, a Reddit thread, and a comparison article all describe you the same way, the model has a clear entity to work with. If they conflict, it either hedges or drops you.

Third-party corroboration. Engines weight independent mentions over self-description. A listicle on a niche industry site that includes you often outperforms your own product page in citations.

Answer-shaped content. Pages that state a plain answer in the first two sentences, then support it, are far easier for a model to lift than pages that open with three paragraphs of brand story.

A working framework for LLM optimization

Treat this as a measurement loop rather than a checklist you run once. The engines change, the answers change, and a fix that worked in March can be gone by June.

Baseline first. Build a prompt set that mirrors real buyer language, thirty to fifty prompts across awareness, comparison, and purchase intent. Run each one multiple times on each engine you care about. Record whether you appear, where, with what sentiment, and which sources get cited. This is your visibility rate and your citation map. Without it, every later step is a guess.

Fix the entity. Audit every page that describes your company and make the facts match: what you do, who it is for, pricing tier, founding details, key integrations. Add schema markup for Organization, Product, and FAQ. Make sure your About page reads like a reference entry rather than a manifesto. Models are trying to resolve you to a stable set of facts. Help them.

Publish for extraction. For each prompt where a competitor appears and you do not, write or rewrite one page that answers that question directly in the opening lines, uses the same vocabulary the prompt uses, and includes the comparison the buyer is actually making. Keep it plain. A model quoting you does not care about your brand voice, it cares about whether the sentence stands alone.

Earn the citations. Pull the top twenty domains your citation map shows for your category. That is your outreach list. Get into the comparison roundups, the community threads, the documentation ecosystems, and the industry directories that engines are already pulling from. This is PR with a measurable target.

Re-measure. Re-run the baseline set every week. Track visibility rate and share of voice per engine. Note which changes moved which prompts. Kill what did not work. Double down on what did.

The loop matters more than any single tactic. Teams that run it for a quarter typically find that a handful of prompts and a handful of source domains account for most of the movement, and they reallocate effort accordingly.

Mistakes that waste a quarter

Some patterns show up in nearly every team that tries this without a measurement layer.

Optimizing for a single sample. Someone asks ChatGPT one question, sees a competitor, and the whole content plan pivots. AI answers vary run to run. One sample is noise. A rate across many runs is signal.

Treating it as Google SEO with new keywords. Backlinks and rankings still matter for the retrieval layer, but the unit of competition has shifted from a URL in a list to a brand in a sentence. You can rank first on Google and be absent from the AI answer for the same query.

Chasing every engine equally. Perplexity leans on citations. ChatGPT leans more on trained knowledge unless browsing kicks in. Gemini and AI Overviews are tied to Google's index. Your buyers are not evenly distributed across them. Find where your prompts get asked and weight your effort there.

Ignoring accuracy. Being mentioned with the wrong pricing, a discontinued feature, or a competitor's positioning is worse than being absent. Monitoring for hallucinated facts about your product is part of the job, not an afterthought.

Stopping after the first win. A citation you earned can vanish when the source page updates or the engine re-weights. Visibility decays without maintenance.

What to do this week

You do not need a new team or a six-figure budget to start. You need a baseline.

Write down twenty questions your best customers asked before they bought. Run each one on ChatGPT, Perplexity, and Gemini three times. Note who shows up and who gets cited. That single spreadsheet will tell you more about your LLM visibility than any conference talk, and it will make the case for a proper tool if the gaps are as large as they usually are.

If you would rather skip the spreadsheet, LLM Search Console runs that baseline continuously across engines, tracks your share of voice against named competitors, and maps which sources are getting cited so your content and PR teams know where to aim.

Either way, start measuring before you start optimizing. Subscribe to this newsletter for a weekly breakdown of what is moving in AI search, which tactics are holding up, and what the engines changed since the last issue.

Your problem is narrower and more expensive. A buyer types "best [your category] for a 50-person team" into ChatGPT, Gemini, Claude or Perplexity. The answer names three companies. None of them is you. The buyer never opens a search results page, never sees your ad, never lands on the comparison post you spent three weeks on. The deal is shaped before you know it exists.

That is LLM optimization in the marketing sense: influencing how large language models describe, rank and recommend your brand. You will also hear it called generative engine optimization (GEO), answer engine optimization (AEO) or LLM SEO. The labels overlap. The work is the same.

Two meanings, one phrase

Before you brief anyone, settle the vocabulary so your ML lead and your content lead stop talking past each other.

Model optimization is an engineering task. Pruning, distillation, quantization, prompt caching, batching. The goal is throughput per dollar. The owner is infrastructure.

LLM optimization, as used here, is a go-to-market task. Entity clarity, citable content, third-party corroboration, crawler access and ongoing measurement. The goal is a higher rate of accurate brand mentions in AI answers. The owner is marketing, with help from SEO and comms.

Everything below is about the second one.

Why the timing matters

OpenAI put ChatGPT at 900 million weekly active users in early 2026, more than double the figure a year earlier. Gemini ships inside Google Search and Workspace. Perplexity has made itself the default for research-heavy queries. Every one of those surfaces answers the "which vendor should I pick" question without showing ten blue links.

Two things follow from that. First, a growing share of your category demand never touches a results page you can rank on. Second, the answer a model gives is repeatable. Ask the same buying question fifty times and the same three or four brands show up most of the time. Being on that list is a position. Being off it is a leak in your funnel that your analytics cannot see, because there is no click to attribute.

How a model decides who to name

Models pull from two places, and each responds to different levers.

The first is what the model learned during training. Every description of your company across the open web, Wikipedia, review sites, press, directories and forums got compressed into weights. If those descriptions are sparse, inconsistent or outdated, the model either hallucinates your positioning or skips you. You cannot patch training data on demand. You can only make the next crawl of the web more consistent than the last.

The second is live retrieval. ChatGPT search, Perplexity, Gemini grounding and Google AI Mode fetch pages at answer time and quote them. This path behaves more like classic search, with a twist: the model wants a passage it can lift and attribute, not a page it can rank. Direct answers, clean structure and explicit numbers win. Long preambles lose.

Most brands are invisible on the first path and mediocre on the second. Fixing that is the job.

The LLM optimization framework

Five parts. Do them in order.

1. Get your entity straight

Write one sentence that says what you are, who you serve and where you operate. Use that sentence, word for word, on your homepage, your About page, LinkedIn, Crunchbase, G2, your Wikipedia or Wikidata entry if you have one, and every press boilerplate. Models learn by repetition across independent sources. Ten slightly different self-descriptions teach a model nothing. One description repeated in thirty places teaches it exactly what to say.

Test it. Ask each major model "What is [brand]?" and "What does [brand] do?" If the answers are wrong, vague or dated, the entity work is not done.

2. Write answer-shaped pages

Retrieval models extract passages. Give them passages worth extracting.

• Lead with the answer. The first 80 words of any page should resolve the question in the title.

• Use question-form H2s that mirror how buyers actually ask.

• Put two or three specific, sourced numbers on every important page. Models cite numbers.

• Add FAQ and Organization schema. It is not magic, but it removes ambiguity.

• Keep tables and lists real. A comparison table with honest criteria gets quoted. A wall of adjectives does not.

3. Earn corroboration you do not control

This is the part most teams skip, and it is the part that moves the needle. Models weight independent sources above your own site. Run your twenty most important buying prompts through Perplexity and note which domains it cites. Those are the pages that currently define your category. Get on them. That means review profiles with fresh reviews, inclusion in the "best tools for X" roundups that keep getting cited, honest Reddit and community presence, and trade press that describes you the same way you describe yourself.

Think of it as a citation gap analysis. Your competitor is on eight of the ten pages the model trusts. You are on two. Close the gap and the answers change.

4. Open the door technically

Check robots.txt for GPTBot, ClaudeBot, PerplexityBot and Google-Extended. Plenty of sites blocked them in 2023 and forgot. Server-render your key pages; several AI crawlers still do not execute JavaScript well. Keep canonical tags clean so the model does not split your authority across duplicates. An llms.txt file is cheap to add, though evidence that models use it is thin, so treat it as optional.

5. Measure the rate, not the rank

Rank in an AI answer is close to random between runs. Mention rate is not. Build a prompt set of 30 to 100 questions your buyers ask, run it across ChatGPT, Gemini, Claude and Perplexity on a schedule, and track four numbers: how often you are mentioned, your share of voice against named competitors, the sentiment of the mention, and which of your URLs get cited. That is the dashboard. Everything above feeds it.

A 30-day starting plan

Week one, audit. Run the prompt set, record the baseline, list the domains being cited.

Week two, fix the foundations. Entity sentence everywhere, crawler access, schema on the top ten pages.

Week three, rewrite. Take the five pages closest to your buying prompts and restructure them to answer first.

Week four, corroborate. Pitch three roundup authors, refresh two review profiles, publish one comparison page with real criteria. Then re-run the prompt set and compare.

You will not go from invisible to dominant in a month. You will see the mention rate move, and you will know which lever moved it.

Mistakes that waste the quarter

Treating this as a one-off project. Models refresh, competitors publish, answers drift. This is a monitoring discipline, not a launch.

Optimizing for ChatGPT alone. Gemini reaches a different audience through Google. Perplexity cites differently. Claude is increasingly the model behind B2B research assistants. Track all of them or you are guessing.

Chasing rank. Two runs of the same prompt can reorder the list. Report mention rate and share of voice instead, or you will be explaining noise to your CFO.

Pasting "AI-friendly" boilerplate onto pages. Models are trained on the whole web. They recognize filler. Specifics beat signals every time.

Where to start today

Open ChatGPT, Gemini and Perplexity. Ask each one the question your best customer asked before they found you. Write down who gets named. If you are not on the list, you now have a baseline and a reason.

If you would rather not do that by hand every week, LLM Search Console runs your prompt set across the major models on a schedule and reports mention rate, share of voice, sentiment and citations in one place. Start with a free audit and see what the models are already saying about you.

Most GEO advice still assumes a model reads your page. In agentic products it does not. It reads a tool result: a JSON payload returned by a function call, often through the Model Context Protocol (MCP), trimmed to fit a token budget, then summarised again when the conversation runs long. Your brand has to survive three handoffs it never sees. Here are three connections between MCP, function calling, and context windows that decide whether it does.

Connection 1: The tool description is the new meta description

With function calling, the model does not decide to "search". It decides to call web_search, get_reviews, or compare_products based on a few hundred tokens of schema text the MCP server exposes. That schema is the only thing between the user's question and the retrieval path that surfaces you.

Different tools return different shapes of data. A web_search tool returns snippets, so your title tag and first 160 characters do the work. A get_reviews tool hits a review aggregator's API and never touches your site. A compare_products tool pulls structured fields: price, rating, category, region. If your brand exists only as prose on a landing page, two of those three paths return nothing about you. The model then says "I could not find pricing for X" and moves on. That sentence is a negative mention, and it is measurable.

Connection 2: Function-call results are truncated from the bottom

MCP servers and orchestration frameworks cap tool output. A common default is a few thousand tokens per result. When the search tool returns ten results and the cap cuts at six, results seven through ten do not exist for the model. There is no "page two" of a tool result.

This changes what "ranking" means. Position in the raw API response matters more than position on a SERP, because the truncation is deterministic and the model never learns it happened. Grounding is only as good as the slice that survived the cap. The practical effect is that brands sitting at result eight in a retrieval call are invisible across every prompt that uses that tool, and no amount of on-page work fixes a cutoff that happens in the middleware.

Connection 3: Compaction erases you mid-conversation

Long agentic sessions run out of context. Frameworks handle this by compacting earlier turns: older tool results get replaced by a short summary the model writes for itself. That summary keeps whatever the model thought was salient at the time. If your brand appeared as one name in a list of five, the summary usually keeps "several options were found" and drops the names.

So the brand that was retrieved in turn three is gone by turn nine, when the user finally asks "which one should I pick?". The competitor that got a full sentence of explanation in turn three survives compaction because there was something to summarise. Token efficiency is not just a cost line for the vendor. It is a filter applied to your brand, repeatedly, inside a single conversation.

What this means for measurement

None of this shows up in a rank tracker. It shows up as variance: the same prompt, same model, same day, with your brand present in some runs and absent in others. Teams read that as noise. It is usually retrieval path selection, result truncation, and compaction doing exactly what they were built to do.

LLM Search Console runs the same prompts on a daily or weekly schedule across ChatGPT, Gemini, Perplexity, Claude, Google AI Overview and AI Mode, across markets, and logs every citation with its source URL and the snippet the model actually used. That citation log is how you find out which retrieval path is surfacing you, which one is truncating you, and which competitor is getting the full sentence that survives compaction.

Quick wins for GEO

Publish structured fields. Price, category, region, ratings, and a one-line description in JSON-LD and in plain HTML tables. Tools that pull structured data need something to pull.

Front-load the snippet. Brand name, category, and differentiator in the first sentence of every page. Truncation is bottom-up; win the top.

Get into the aggregators. Review sites, comparison APIs, and directories are the endpoints get_reviews style tools actually call. A missing listing is a missing tool result.

Earn a sentence, not a name. Pages that explain why a brand fits a use case give the model something to summarise. Lists of names get compacted to "several options".

Track variance, not rank. Run each prompt on a schedule and watch presence rate per model. A brand at 40% presence has a retrieval problem, not a content problem.

Add your five most important buyer prompts to LLM Search Console, run the first scan, and read the citation log. You will see which tool path is dropping you within the first week.

That gap is where the phrase "AI SEO tools" gets confusing. Search the term and you'll get a wall of products with "AI" in the name. Most of them are content generators, keyword clusterers, or classic SEO suites that bolted a chatbot onto the sidebar. They help you produce pages for Google. They do not tell you whether ChatGPT, Perplexity, Gemini, or Claude recommend your brand when a buyer asks. Those are different jobs, and the second one is the one your CMO is going to ask about next quarter.

This article sorts the category, gives you a checklist for evaluating tools built for LLM visibility, and shows how to run a 30-day pilot that produces a number your leadership will accept.

Two products hide under one label

The market uses "AI SEO tools" for both of these, and they solve opposite problems.

The first group is AI-assisted SEO. These tools use language models to speed up traditional work: drafting briefs, generating outlines, rewriting meta descriptions, clustering keywords, auditing technical issues. The output is content and recommendations for Google's ten blue links. Surfer, Clearscope, Jasper, and the AI features inside Semrush and Ahrefs mostly live here. They're useful. They're also solving 2019's problem with 2026's technology.

The second group is AI search visibility tools. These measure what generative engines say. They run a defined set of prompts against ChatGPT, Perplexity, Gemini, Claude, Copilot, and Google AI Overviews, record whether your brand appears, what position it takes, which sources get cited, and how the sentiment reads. The output is a visibility rate, a share of voice, and a citation map. This is the category LLM Search Console belongs to.

Here's the practical test. Ask the vendor one question: "Can your tool tell me how often my brand appeared in ChatGPT answers last week compared to my top competitor?" If the answer involves the word "roadmap," you're looking at a content tool.

Why your existing SEO stack misses this

Traditional SEO tooling is built on a shared assumption: there is a results page, it has positions, and you can scrape it. AI answers break all three parts.

There's no fixed results page. The same prompt returns different answers on different days, sometimes different answers in the same minute. Position is fuzzy. A brand can be mentioned first in one run and omitted entirely in the next. And the engines don't publish an API for "show me what you said to everyone today," so measurement requires running prompts at scale and sampling repeatedly.

This is why "rank tracking for ChatGPT" is a misleading promise. Any tool that shows you a single rank number for an AI answer is hiding the variance. What you need is a visibility rate: across N runs of this prompt over this period, your brand appeared in X% of answers. That's a stable metric. Rank is not.

A second blind spot is citations. Perplexity and Google AI Overviews lean heavily on cited sources, and the pages getting cited are frequently not the pages ranking #1 in classic search. Reddit threads, niche comparison sites, and documentation pages show up constantly. Your rank tracker sees none of it. A tool built for AI search visibility has to track those citations by URL and domain, because that list is your actual link-building target now.

The evaluation checklist for LLM-visibility tools

Once you've filtered out the content generators, the remaining vendors look similar on the surface. These are the criteria that separate them in practice.

Engine coverage. ChatGPT and Perplexity are table stakes. Gemini and Google AI Overviews matter for anyone with a Google-heavy audience. Claude and Copilot are where several tools go quiet, and they're growing in enterprise and B2B usage. Ask for the full list and ask how often each engine is queried.

Prompt methodology. The tool should let you define your own prompt set, not just track a handful of generic keywords. Real buyers ask "best CRM for a 20-person agency with HubSpot fatigue," not "CRM." Look for the ability to build prompts by funnel stage, persona, and use case, and to run each one multiple times to get a real rate rather than a single sample.

Competitor benchmarking. You need the same prompt set run against your named competitors, with share of voice calculated per prompt, per engine, and in aggregate. A visibility rate with no comparison is a vanity metric.

Citation and source tracking. Which URLs are being cited when your brand appears, and which are cited when it doesn't. This is the data that tells your content team what to build and your PR team where to get placed.

Sentiment and accuracy. Being mentioned isn't the goal. Being mentioned correctly and favorably is. The tool should flag negative framing and factual errors about your product, because hallucinated pricing or a wrong feature claim is a reputation problem you can't see without monitoring.

Historical trend data. Weekly snapshots at minimum. You're going to be asked "did the campaign work," and the only honest answer is a before-and-after visibility rate.

Exportable reporting. Agencies and in-house teams both need to drop this into a deck. Check for CSV exports, scheduled reports, and multi-brand or multi-client workspaces if you manage more than one.

Three things you can safely ignore: AI content generation inside the visibility tool (use a dedicated writer if you want one), any "AI rank" number presented without variance, and daily-frequency claims that don't specify how many runs per prompt.

A 30-day pilot that produces a defensible number

You don't need a year-long contract to find out whether this matters for your brand. Run it like this.

Week one is baseline. Build 30 to 50 prompts that mirror real buyer questions across awareness, comparison, and purchase intent. Include your three closest competitors. Run the full set across every engine the tool supports. Record your visibility rate, share of voice, and the top 20 cited domains.

Weeks two and three are the intervention. Pick the five prompts where competitors appear and you don't. Look at what's being cited for those answers. Then do two things: publish or update one page per prompt that directly answers the question in plain, extractable language with clear entity signals (who you are, what you do, who it's for, what it costs), and pursue a mention or citation on two of the third-party domains the engines already trust for that topic.

Week four is measurement. Re-run the full prompt set. Compare visibility rate and share of voice against the baseline. Note which engines moved and which didn't.

Early movers report that Perplexity and AI Overviews shift fastest because they re-crawl cited sources frequently; ChatGPT with browsing moves next; model-only answers without retrieval move slowest and depend on training data you can't accelerate. The pattern will differ for your category, and that's the point of measuring instead of guessing.

What to tell your CMO

Frame it as a coverage problem, not a tooling problem. Your current SEO stack measures 100% of Google's classic results and 0% of AI answers, and AI answers are now where a meaningful share of high-intent research happens. The ask isn't "another SEO tool." It's instrumenting a channel you're currently flying blind in.

Bring one number: your share of voice versus your top competitor across your core prompt set. If it's lower than your Google share, you have a gap that's costing you deals you never see in analytics. If it's higher, you have an advantage worth protecting before competitors notice.

Either way, you can't manage what nobody in the building is measuring.

Start measuring this week

Run the baseline. Fifty prompts, four engines, three competitors, one afternoon. LLM Search Console is built for exactly this workflow and covers ChatGPT, Perplexity, Gemini, Claude, and Google AI Overviews in one dashboard.

If you'd rather read before you run, subscribe to this newsletter. Every week we publish one piece on how brands get chosen inside AI answers, with the prompts, the data, and the methodology included. No filler, one email, unsubscribe whenever you like.

The term is messy. Search "AI SEO" and half the results are about using AI to write meta descriptions faster. That is not what this article covers. This is about the other side of the equation, the one where the AI is the search engine and your brand either shows up in the answer or it doesn't.

Why AI SEO matters right now

Three shifts happened at once. Google put AI Overviews on the majority of informational queries, which compresses clicks even when you rank first. ChatGPT, Perplexity, Gemini and Claude became a default research step for buyers, especially in B2B, where a single "what are the best tools for X" prompt replaces an hour of tab-hopping. And the answers those models give are opinionated. They name three or four brands, not ten. If you are not in that shortlist, you are not in the consideration set.

The painful part is that you cannot see any of this in Google Search Console. There is no impressions report for ChatGPT. No average position in Perplexity. Which is why LLM visibility has become its own discipline, with its own metrics and its own tooling.

What AI SEO actually is

AI SEO (you will also see it called generative engine optimization, answer engine optimization, or LLM SEO) is the practice of increasing how often and how favorably your brand appears in answers generated by large language models. It sits on top of traditional SEO but optimizes for a different outcome.

Traditional SEO optimizes for a ranking. AI SEO optimizes for a mention, a citation, or a recommendation.

That distinction changes what you do. A page can rank third on Google and never be cited by an AI engine. A page can rank on page four and still be the source Perplexity pulls from, because it answered the question cleanly in one paragraph that the model could lift.

The three ways a brand shows up in AI answers

• Mentioned. The model names your brand in the answer text.

• Cited. The model links to your page as a source (Perplexity, Google AI Overviews and ChatGPT search all do this).

• Recommended. The model puts you in a shortlist in response to a "best" or "which should I choose" prompt.

Each one is tracked differently and each one is earned differently. Most brands are chasing the first without measuring any of them.

How AI engines decide who to name

Nobody outside the model labs knows the full ranking logic, and it shifts with every model update. But the patterns from tracking thousands of prompts are consistent enough to act on.

Entity clarity beats keyword density

Models reason about entities, not strings. If the web disagrees about what your company does, who it serves, or what category it belongs to, the model hedges and picks a competitor it understands better. Your About page, your LinkedIn description, your Crunchbase entry and your G2 category should all say the same thing in roughly the same words.

Third-party corroboration carries more weight than your own site

Your homepage saying you are "the leading platform" does nothing. A review site, a comparison article, a Reddit thread and an industry newsletter all saying you exist and describing you the same way does a lot. AI engines lean on consensus. Build it.

Extractable content gets cited

Perplexity and AI Overviews pull passages, not pages. A 2,000-word article with the answer buried in paragraph nine loses to a 400-word page that opens with a direct definition, a short list and a clear claim. Structure for lifting, not for dwell time.

Freshness matters more than it did

Retrieval-augmented engines (Perplexity, ChatGPT with browsing, Gemini) favor recently updated sources on fast-moving topics. A page last touched in 2023 is a page the model treats as stale.

An AI SEO framework you can run this quarter

Skip the theory. Here is the operating loop.

1. Build a prompt set, not a keyword list

Take your top 30 commercial keywords and rewrite each one as the question a buyer would actually type into ChatGPT. "ai seo tools" becomes "What are the best tools for tracking brand visibility in ChatGPT?" Add category prompts, comparison prompts ("X vs Y"), and problem prompts ("how do I know if my brand shows up in AI answers"). Fifty to one hundred prompts is enough to start.

2. Measure your baseline

Run the prompt set across ChatGPT, Perplexity, Gemini and Claude. Record three things per prompt: were you mentioned, were you cited, and who else was named. That gives you a mention rate, a citation rate and an AI share of voice against competitors. This is the number that replaces "average position." A dedicated LLM brand visibility tracker does this automatically and re-runs it weekly, which matters because answers drift.

3. Find the gaps

Sort prompts by where competitors appear and you do not. Those are your gap prompts. For each one, look at which sources the engines are citing. That is your target list of pages to get onto, earn mentions from, or outcompete.

4. Fix the entity layer

Standardize your brand description everywhere it lives. Add Organization, Product and FAQ schema. Make sure your category is unambiguous. Get listed in the directories and review sites the engines are already citing for your gap prompts.

5. Publish extractable answers

For each gap prompt, publish or rewrite a page that answers it in the first 150 words, uses the exact phrasing a buyer would use, and includes a short comparison or list the model can lift. Keep the H2s as questions.

6. Re-measure and repeat

Re-run the prompt set. Compare mention rate and share of voice to baseline. Keep what moved the number. Cut what didn't.

The mistakes that waste the first six months

• Chasing "rank" in AI answers. Position inside an AI answer is close to random between runs. Track visibility rate over many runs instead.

• Only checking ChatGPT. Perplexity is citation-driven, Gemini feeds AI Overviews, Claude is heavily used by technical buyers. Your visibility differs by engine, sometimes wildly.

• Publishing more without fixing the entity problem. Twenty new blog posts do not help if the model still is not sure what you sell.

• Treating it as a one-time audit. Model updates reshuffle answers. What you measured in June is not what is true in September.

Where this goes next

AI SEO is going to fold into regular marketing operations the way social did: a dashboard, a weekly number, an owner. The brands that set up measurement now get two quarters of signal before their competitors know what to look at. That is the whole advantage, and it is a temporary one.

Start with the prompt set. Measure your mention rate this week. Then decide what to build.

If you want this kind of breakdown every week, with real prompt data and what is moving in ChatGPT, Perplexity and Gemini, subscribe to the newsletter below.

1. Synthetic data is the new crawl, and it only keeps what the teacher explains

Frontier labs hit the data wall. The fix is synthetic data: a strong teacher model generates reasoning paths, comparisons, and worked answers, and the student is trained on those. The under-discussed consequence is that your brand no longer enters the weights by being on the web. It enters by being in the teacher's explanations. If the teacher answers "best rank-tracking alternatives for AI search" and reasons through three vendors, those three get millions of synthetic examples. The fourth vendor, the one it paraphrased as "other tools," gets nothing.

That means today's citation share is tomorrow's training distribution. A brand that is consistently named in current frontier answers compounds; a brand that is consistently summarized without a name gets distilled out. This is why raw "mention count" is a lagging metric. The leading metric is named vs. paraphrased mentions across models, which is exactly what LLM Search Console separates in its competitor tracking. A paraphrase is not a win. It is a signal that you are about to be deleted from the next generation.

2. Fine-tuning is now about format, and format is a schema you either fit or don't

Fine-tuning has stopped being how models learn facts (RAG handles that) and started being how they learn shape: comparison tables, three-item shortlists, pros/cons blocks, JSON for downstream agents. Those formats have implicit schemas. A shortlist row needs a category, a price band, a primary use case, and a differentiator. If a model retrieves your page and cannot fill those four slots from it, the format fine-tune does not produce a half-empty row. It drops the row and picks a competitor whose page filled the slots.

The hidden link: format adherence is a retrieval filter. Your content is not competing on quality. It is competing on whether its attributes can be extracted into the answer template with zero inference. Pages that state "Pricing: from $X/month. Category: AI visibility monitoring. Best for: mid-market SEO teams" survive the template. Pages that say "flexible plans for teams of any size" do not.

3. The 7B model with great CoT beats the 70B one, and it has no room for you

The parameter-count lesson of 2026 is that a small model with strong chain-of-thought often outperforms a large "dumb" one. So agent stacks moved the cheap, high-frequency steps to small models: query classification, source reranking, tool selection. A 7B reranker running quantized on a single GPU now decides which ten of fifty retrieved chunks the big model gets to read.

Small models have small parametric memory. A 7B model has essentially no long-tail brand knowledge. It reranks on lexical and structural signals in the chunk itself: does this passage directly answer the sub-question, does it name entities, is it self-contained. If your brand only appears in a header and a footer while the body chunk talks about "our platform," the reranker scores the chunk as an anonymous passage and cuts it. The frontier model never sees you, not because it forgot you, but because a model that never knew you did the pruning.

4. Measuring the gap between the demo model and the deployed stack

Put these together and you get the GEO failure mode that no screenshot catches: a brand that appears in a direct ChatGPT test, is paraphrased in Perplexity, and is absent from an agentic product built on a small-model router, all for the same query on the same day. Testing one model tells you nothing about the pipeline.

The fix is to run the same prompt set across models on a schedule, track named vs. paraphrased mentions per model, and watch the divergence. When a small or mid-tier model drops you that a frontier model keeps, you have a chunk-level structure problem, not a content problem. When a frontier model paraphrases you, you have a training-distribution problem coming in six to twelve months. LLM Search Console runs exactly this loop: prompts across ChatGPT, Perplexity, Gemini, and Claude, competitor share of voice, and citation source tracking, so the divergence is visible before it becomes permanent.

Quick wins for GEO

• Make every body chunk self-identifying. Name the brand and the category inside the paragraph, not just the H1. Rerankers score chunks, not pages.

• Publish the schema, not the vibe. Category, pricing band, best-for, and one differentiator, stated as plain attribute sentences on your product page.

• Track named vs. paraphrased. A paraphrase is a leading indicator of removal from synthetic training data.

• Test small models, not just flagships. If a 7B or 8B open model cannot place you, the agentic stacks built on it won't either.

• Ship comparison content that fills the template. Third-party comparisons with the same four attributes for you and your competitors feed the teacher model exactly what it needs to name you.

• Schedule it. One-off checks measure a moment; the distributions you are fighting shift monthly.