Most marketing sites were built for two readers: a human with a browser and Googlebot. A third reader now shows up every day. It loads your page, strips it to text, checks your pricing against three alternatives and either completes a task or leaves. It doesn't complain. It just recommends someone else.
Learning how to optimize for AI agents is less about new tactics and more about testing. You can't fix what you haven't watched fail. The seven tests below take an afternoon, need no special software and give you a ranked list of fixes for the pages that drive revenue. They also feed directly into your LLM visibility, because the same pages agents struggle with are the ones answer engines misquote.
What "optimizing for AI agents" actually covers
An AI agent is a model that takes actions, not just one that answers. ChatGPT agent, Perplexity Comet, Gemini's agent features and Claude's browsing tools all open pages, follow links, read tables and fill in forms on a user's behalf.
That gives you three jobs.
Let agents in
Make facts readable
Make tasks finishable
Our earlier piece on agentic search optimization laid out the strategy. This one is the hands-on version. Run each test, write down what breaks, then fix in order.
Test 1. The raw HTML test
Open a terminal and fetch your pricing page without a browser.
curl -sL https://yourdomain.com/pricing | grep -i "per month"
If your prices, plan names and feature limits don't appear in the output, many agents and AI crawlers won't see them either. Some agents render JavaScript. Many retrieval crawlers don't, and even the ones that do work faster and more reliably with server-rendered text.
The fix is usually server-side rendering or static generation for pricing, product, comparison and integration pages. Start with the three pages a buyer needs to make a shortlist decision.
Test 2. The door test
Read your robots.txt line by line. Then check what your CDN or bot-protection layer does with AI traffic.
The common mistake is treating every AI bot the same. Vendors now run separate agents for separate jobs. OpenAI, for example, uses GPTBot for training, OAI-SearchBot for search and ChatGPT-User for requests a person triggers. Anthropic and Perplexity follow a similar split.
Blocking training crawlers is a business decision. Blocking user-triggered agents removes you from live answers and agent tasks, which is rarely what anyone intended. Review the rules, decide per bot and write the reasoning down so nobody "fixes" it later.
Then check the firewall. Aggressive challenge pages and CAPTCHA walls stop agent browsers just as well as scrapers. Some agents now sign their requests so you can verify them, which makes allowlisting easier.
Test 3. The accessibility tree test
Agents rarely look at your design. They read the DOM and the accessibility tree, the same structure screen readers use.
Open Chrome DevTools, go to the Accessibility pane and click through your signup or demo flow. Look for these problems.
Unlabeled form fields
Icon-only buttons
Divs pretending to be links
Modals that trap focus
If a screen reader would struggle, so will an agent. An agent that can't tell your "Start trial" button from a decorative icon will move to the next site in its list.
Test 4. The fact check test
Ask three engines the questions your buyers ask before they talk to sales. What does it cost? Which plan includes SSO? Does it integrate with HubSpot? What's the contract minimum?
Compare every answer with your current site. Write down each mismatch and its source if the engine shows one.
Wrong answers usually come from one of three places. Your own page hides the fact in an image or PDF. An old review or roundup still lists last year's pricing. Or a competitor's comparison page describes you on their terms.
Put every buying fact in plain text on one page each. Add Product, Offer, SoftwareApplication or Organization schema so models can map the facts to your brand as an entity. Then chase down the third-party pages that disagree with you.
Test 5. The task test
This is the test most teams skip, and the one that matters most once agents start acting.
Give an agent a real buyer task, not a question. "Find a project management tool under $15 per seat with Gantt charts and book a demo with the best fit." Run it in two or three agent tools and watch.
Note where it stalls. Common failure points look like this.
Demo form wants 11 fields
Calendar widget won't load
Pricing says "Contact us"
Signup needs email verification mid-flow
You don't need to remove every step. You need a path an agent can finish or hand back cleanly to the human. Short forms, predictable URLs like /pricing and /demo, and visible confirmation states do most of the work.
If you sell online, look at agentic checkout. OpenAI's Agentic Commerce Protocol and Google's Agent Payments Protocol both aim to let purchases complete inside the agent. Check whether your commerce platform supports either before building anything custom.
Test 6. The outside-in test
Agents cross-check. Before recommending you, they read G2, Capterra, Reddit threads, analyst lists and "best X for Y" articles.
Search your category the way a buyer would and list the ten sources that come up most. For each one, ask a simple question. Does it describe us accurately and recently?
Fix your own listings first, since you control them. Then reach out to authors of outdated roundups with corrected facts. This is slow work. It is also the work that decides whether an agent trusts your site or trusts someone else's description of it.
Test 7. The baseline test
None of the fixes above mean much without a before-and-after number. Traffic is a poor signal here because agents often finish research without sending a click.
Track these instead.
Inclusion rate across buyer prompts
Accuracy of repeated facts
Share of voice versus named competitors
Agent user-agents in server logs
Run the same prompt set every week. Inclusion tells you whether you make the shortlist. Accuracy tells you whether the shortlist describes you correctly. Logs show which agents already visit and which pages they hit. A tool focused on LLM brand visibility like LLM Search Console runs those prompts across ChatGPT, Gemini, Perplexity and other engines automatically, so you can watch inclusion and competitor share move after each fix instead of checking by hand.
A one-week plan
You don't need a replatform. You need a short, ordered list.
Monday, run tests 1 and 2 on your pricing page. Tuesday, run test 3 on your demo or signup flow. Wednesday, run test 4 across three engines and log every wrong fact. Thursday, run test 5 with one real buyer task. Friday, set your baseline with test 7 and ship the single biggest fix.
Then repeat next month on the next revenue page.
Your next step
Agents are already doing first-round research for your buyers. The brands they pick aren't the loudest. They're the ones whose facts load in plain HTML, whose forms an agent can complete and whose third-party footprint agrees with their own site.
Pick your pricing page. Run test 1 today. It takes two minutes and tells you more about your AI visibility than most dashboards.
Want the next playbook in your inbox? Subscribe to the LLM Search Console newsletter on Substack. One practical breakdown a week on AI visibility, agent readiness and competitor tracking, with no fluff.


