How to track brand mentions in AI search across four engines

The false premise is that one AI engine can stand in for AI search. Track the same buyer questions across ChatGPT, Gemini, Claude and Perplexity instead, because each engine can produce a different view of your brand.

You need to record the mention, the cited source, the question and the engine behind each answer. A brand can appear in one answer and disappear from another, even when the question is unchanged.

Start with the raw result. Scores can wait.

How do you define a brand mention across engines?

A mention means your product appears in the answer text. A citation means the engine links to a page it used as a source. Those events can overlap, but they are not the same event. An answer can name a product without citing its site, or cite a page without naming the product in the answer itself. If you report only citations, you miss recommendations without a source link. If you report only mentions, you miss the pages that gave the engine evidence. Keep both fields, along with the engine, question, answer and cited page. Use the same buyer questions on every engine, because a changed question set can look like a visibility change. The CiteGraph measurement method records answers and sources by engine, rather than smoothing them into one score.

The unit is each answer and its evidence.

What does each AI engine actually show you?

The sample shows a clear spread. Perplexity named the scanned product in 10.5% of 2,440 sampled answers and cited 5.5 sources per answer on average. Gemini named it in 9.4% of 2,437 answers and cited 4.2 sources. ChatGPT named it in 6.1% of 2,434 answers and cited 3.5 sources. Claude named it in 10.7% of 2,120 answers and cited 4.9 sources. Claude and Perplexity had the highest naming rates in this sample, while ChatGPT had the lowest. Perplexity also returned the deepest source set on average, while ChatGPT returned the shallowest. That does not make one engine better in general. It means an engine-specific report can give you a different brand picture.

Treat each engine as a separate view.

Why do you get different answers from each engine on the same question?

The engines do not have to retrieve the same pages, weigh the same evidence or write the same answer. Even with an identical prompt, their source selection can differ. Across the research sample, we recorded 64,667 citations across 9,791 AI answers from 77 scans of 44 sites in 48 categories, using 507 distinct buyer questions. The engines cited 9,192 distinct domains and 21,465 distinct pages. The page mix was concentrated: listicles and roundups supplied 68.3% of recorded citations, while competitor product pages supplied 24.1%. Together, those two page types supplied 92.4% of citations. Community threads supplied 3%, and every other page type shared the remainder. One engine may retrieve a roundup while another retrieves competitor pages, so the answers can diverge before either model writes a sentence about your product.

Keep the engine label on every result.

How do you run a first pass in under an hour?

You do not need a perfect prompt library to begin. You need a fixed set of buyer questions, consistent product-name rules and a record of what each engine said. Run the same questions through all four engines, then inspect the result rather than hiding it inside a score:

1. Write down the buyer questions that matter to your product, including category, comparison and alternative questions.
2. Run the same questions through ChatGPT, Gemini, Claude and Perplexity.
3. Record whether your product was named, whether the answer recommended it and which pages were cited.
4. Tag each result by engine, question, mention status and cited page type.

Save the wording around each mention. “Included in a list” and “recommended for this use case” are different outcomes, even if both count as a mention. Save the citation URL too. The free AI visibility checker gives you a quick starting point, and the measurement method guide helps when you need to set rules before comparing runs.

Run the same set again and compare each engine with its own earlier results.

What can citations tell you about brand mentions?

Citations show where an engine found usable evidence. They do not prove that the publisher endorses your product or that the cited page caused the mention. They do show which pages are present in the answer's evidence set, which gives you something concrete to inspect. In the recorded sample, listicles, roundups and competitor product pages supplied 92.4% of citations. Comparison pages that earn visibility often make the category, alternatives and selection criteria easy to scan. Review the pages engines already cite for your buyer questions, then decide whether your site needs a clearer comparison, a stronger category explanation or a more direct product page. Your own site is only one part of the evidence set, so include relevant pages you do not control in the review.

Citations explain the route. Mentions show the outcome.

What can you not tell from this sample?

We cannot tell you that a product will be mentioned at a particular rate, that one engine will always outperform another or that a cited page caused a recommendation. We measure samples and report the sample size next to every figure; we do not know what a wider sample would show. The scan data covers the questions, sites, categories and runs included in the sample, not every possible buyer question or answer. We also cannot use these figures as a universal engine league table. The sampled answer counts differ by engine, and question mix, product naming, prompt wording and source availability can move the observed rate.

Use the figures to plan measurement, not to promise an outcome.

This week, do three things:

1. Choose the buyer questions you will run unchanged across all four engines.
2. Record mentions and citations separately, with the engine beside every result.
3. Review the cited roundups and competitor pages before publishing another general product page.