Is it possible to track brand mentions in AI search? Yes, with a margin

You can track brand mentions in AI search. The useful result is a measured rate with a visible margin of error, not a single reassuring screenshot.

The rest of this post shows what to record, how engine samples differ, and how to compare results without treating one assistant as the whole market.

Can you track AI mentions when the answers keep changing?

Yes. Record the questions asked, the engine used, whether the brand appeared, and which pages or domains the answer cited. Repeat that process across a defined sample and you have a measurement of brand mentions, rather than a collection of screenshots gathered when the answer happened to look encouraging.

The variation is the point, not a nuisance to hide. CiteGraph's research data contains 64,667 citations across 9,791 AI answers, from 77 scans of 44 sites in 48 categories. Those scans used 507 distinct buyer questions. This is a sample of answers, not a permanent property of any brand.

A sample gives you a range, not a guarantee.

How do you measure whether an AI engine names your brand?

Start with a fixed set of buyer questions. Keep the wording, category, target product and engine recorded. Then mark each answer for the outcome you care about: named or not named, cited or not cited, and, where useful, which page supplied the citation. An answer can name a product without citing its website, or cite a page without naming the product in the prose.

The engine samples show why one engine cannot stand in for all of them. Perplexity named the scanned product in 10.5% of 2,440 sampled answers. Gemini named it in 9.4% of 2,437 answers. ChatGPT named it in 6.1% of 2,434 answers, while Claude named it in 10.7% of 2,120 answers. The methodology page records these figures and explains the measurement approach.

The choice of engine changes the number you see.

Why does your AI mention rate have a margin of error?

A mention rate is an estimate based on the answers you sampled. If you asked one set of questions today and another set later, the proportion could move because the questions, retrieved sources, model behaviour or answer wording changed. The observed percentage is therefore a measurement of that sample, not proof that the next answer will follow it.

The size of the sample helps set the statistical margin, but it does not capture every source of uncertainty in AI testing. Changes in question mix, retrieval and model behaviour still matter, even when the arithmetic interval looks tidy.

The margin belongs beside the rate.

How much does sample size change what you can trust?

Sample size affects precision, but it does not repair a badly chosen sample. If you measure only branded questions, only one category, or only one engine, a larger collection can give you a precise answer to a narrow question. It will not tell you how the brand appears across other buyer questions or engines.

The current engine samples range from 2,120 answers for Claude to 2,440 for Perplexity. The full scan data covers 9,791 answers, 21,465 distinct pages and 9,192 distinct domains. That broader set gives more room to inspect patterns, but it still describes the questions and sites included in the scans.

A larger sample cannot widen the question set for you.

How should you compare AI mentions across engines?

Compare like with like first. Use the same question set where possible, keep the site and product definition fixed, and report each engine separately before combining anything. The recorded rates show that cross-engine comparisons can produce different rankings, so a result from one assistant should not be treated as a universal score for the brand.

A lower rate may reflect the engine, the question mix, the way the answer is written, or the sources it retrieved. Keep those factors visible in the report instead of folding them into one unexplained score.

Keep the engine in the result.

What does separating mentions from citations let you diagnose?

Separating the outcomes tells you where to inspect next. A low mention rate with strong citation coverage points towards a recognition or wording problem. A healthy mention rate with weak citation coverage points towards source selection, page relevance or retrieval. If both are low, the question set, product positioning and source material all deserve inspection.

The recorded samples show why these measures can diverge. Perplexity averaged 5.5 sources per answer, Gemini 4.2, ChatGPT 3.5 and Claude 4.9. More sources do not automatically mean more brand mentions. They describe how an answer supports itself, not whether it recognises the product in its wording.

Separate the measures so the diagnosis has somewhere to go.

How do you improve a page and report the result?

First, increase the sample in a controlled way. Add buyer questions that represent the decisions you care about, rather than adding random prompts until the number looks calmer. Keep a record of the question set and engine, then compare later runs against the same baseline. Report the observed rate, sample size, engine, question scope and margin of error together. If the sample design does not support a defensible interval, say so instead of printing a spurious exact boundary.

Before, a product page opens with “a flexible platform for modern teams”, follows with feature claims, and ends with a contact button. After, it states who the product is for, names the buyer problem, explains where it fits against alternatives, and puts the relevant facts in plain text. A page that states the buyer problem gives an answer something to quote; a page that does not gives it nothing.

We didn't find a vendor quote worth holding up here, so we're skipping that device rather than manufacturing one.

The source data covers 44 sites and 48 categories, not every brand, question or answer an assistant may produce. We measure samples; we print margins; we do not know what we have not measured. For a practical starting point, use the measurement method guide alongside how CiteGraph measures, or try the free AI visibility checker on a site.

Yes, you can track brand mentions in AI search, provided you measure a defined sample and report its margin.