Ask ChatGPT which vendor to choose in your category. Read the answer. Now look at the source links stacked underneath it. That list is the actual ranking. Everything above it is just prose generated from those pages.
Most brand teams stop at the prose. They check whether their name appeared, log a yes or a no, and move on. The names are the scoreboard. The citations are the game.
Competitor citation tracking is the practice of recording which URLs an AI model pulls from when it answers a commercial question in your category — for every brand in the consideration set, not just yours. It turns a vague sense that "AI likes them more" into a list of specific pages you can go compete for.
A mention is not a citation
These get conflated constantly, and the difference decides what you do next.
A mention is your brand name appearing in the generated text. A citation is a source link the model surfaced as the basis for that text. You can be mentioned without being cited — the model recalled you from training data. You can be cited without being mentioned prominently — a comparison page you don't own put you in a table.
Citations are the more actionable of the two because they point at a document. A document has an owner, a publish date, a structure, and a reason it got picked. You can reverse-engineer all four.
This matters more on citation-heavy surfaces. Perplexity and Google AI Overviews attach sources to nearly everything. ChatGPT does it whenever browsing fires. Tracking LLM brand visibility without tracking the underlying source set means you see the outcome and none of the mechanics.
What a competitor's citation set tells you
Run forty commercial prompts across four models, log every source URL, and group by which brand the answer favored. Three things fall out immediately.
The first is source type. Some categories run on review platforms — G2, Capterra, TrustRadius. Others run on Reddit threads and practitioner blogs. Others run on the vendors' own documentation. If your competitor wins on third-party review sites and you've spent two quarters on your own blog, you now know why.
The second is the specific pages. Not "Reddit" but a named thread from eleven months ago that four different models keep reaching for. That thread is a citation asset. It can be matched.
The third is coverage shape. A competitor cited across thirty of forty prompts has broad grounding. A competitor cited heavily on eight prompts and nowhere else owns a niche. Those two positions need different responses.
Running the analysis
You need a prompt set, a logging discipline, and a scoring rule. That's the whole method.
Build the prompt set from real buying language
Pull it from your sales call notes, your support tickets, and your paid search query report. Aim for thirty to sixty prompts covering category definition ("what is X software"), comparison ("best X tools for Y"), direct competitor queries ("alternatives to Competitor A"), and problem-first queries where nobody names a vendor at all. That last group is usually where you're weakest, and it's usually the largest.
Log the sources, not the summary
For every response, record the answer text, every cited URL, the domain, the publish date, and which brands the answer favored. Do it across models — ChatGPT, Perplexity, Gemini, Claude, and Copilot cite differently, and a source that dominates one may be absent from another. Re-run on a schedule. A single snapshot tells you almost nothing, because AI answers move week to week.
Score the gap
For each domain in the set, count how often it appears in competitor-favorable answers versus yours. The domains with a high competitor count and a zero for you are your target list, ranked. That ranked list is the deliverable. Everything before it is data collection.
Turning a citation gap into work
Once you have the list, the response depends on who owns the page.
Sources you own but that never get cited usually have a structure problem, not a content problem. Buried answers, no clear definitions, claims without dates or figures, no schema markup. Models extract cleanly from documents that state a claim and then support it. Rewrite for extractability before you write anything new.
Sources you don't own split into two piles. Review platforms, directories, and roundups can be influenced through the normal route — customer review programs, updated vendor profiles, outreach to the writer with better data. Community threads and independent posts can't be gamed, but they can be earned by being genuinely useful in the places your buyers already ask questions.
Then there are the gaps nobody has filled. Prompts in your set where models cite thin, dated, or generic sources because nothing better exists. Those are the cheapest wins available and they show up plainly in the data. Write the definitive page, get it indexed, and watch the citation set shift. Sustained LLM visibility comes from owning the source layer, not from rewriting your homepage.
Metrics worth putting in a deck
Citation share is the headline number: your cited URLs as a percentage of all cited URLs across the prompt set. It's the citation-layer equivalent of share of voice, and it moves faster than mention rate.
Underneath that, track domain overlap — the percentage of competitor-cited domains where you also appear. Low overlap means you're not even in the same conversation. Track citation freshness too, because if the sources favoring your competitor are two years old, they're vulnerable.
One caution. Resist the urge to report a "rank." Position inside an AI answer is unstable across runs and re-phrasings, and building a KPI on it will make your reporting look broken. Rate and share hold up. Rank doesn't.
Mistakes that waste a quarter
Tracking one model. Coverage differs enough that a single-model view is close to guessing.
Running the prompt set once. You need a trend line, not a photograph.
Logging only brand names. Without the URLs you can't act on anything.
Prompts written in marketing language. Buyers don't talk like that.
Start with ten prompts
Pick ten prompts your buyers genuinely ask. Run them across ChatGPT and Perplexity. Write down every source link. Sort by domain. You'll have your first citation gap list inside an hour, and it will probably surprise you.
If you'd rather not maintain that spreadsheet by hand, LLM Search Console runs the prompt sets, logs the citations across models, and tracks the gap between you and your competitors over time.
Subscribe for the next post — a teardown of what actually moves a citation gap in ninety days, with the prompt set included.