To find out whether AI assistants mention your product, ask each one the same set of real buyer questions every week and record four things per answer: whether you are mentioned, which competitors are, which sources the answer drew on, and which it cited. One answer proves little; a fixed question set read over weeks shows a trend you can act on.
Why one answer tells you almost nothing
Ask ChatGPT the same question twice and you can get two different answers. Models change, web retrieval changes, and consumer apps personalize by account, location and history. A screenshot of one answer that mentions you is an anecdote, not a measurement.
What holds still is the method. Keep the questions, the assistant, the model and the market fixed, run them on a schedule, and compare like with like. That is the difference between "ChatGPT mentioned us once" and "we are mentioned in 3 of 10 buyer questions, up from 1 a month ago".
Step 1: Choose ten questions your buyers actually ask
Start from evidence, not guesses. The best sources are the search queries that already bring people to you in Google Search Console, questions from sales calls, and what visitors type into your support chat. Pick questions that describe a problem, not your brand name: an assistant that only mentions you when asked about you is not helping anyone find you.
| Weak question | Better question | Why |
|---|---|---|
| What is Acme Analytics? | Which web analytics tools work without cookies? | Problem-shaped, asked before the buyer knows you |
| Best analytics tool | Which analytics tool shows which pages lead to sign-ups for a small SaaS? | Specific audience and outcome, so answers are comparable |
| Acme vs everything | Is there a Plausible alternative that also tracks SEO and AI answers? | Names the comparison buyers actually make |
Give each question one market and one language. "Best analytics tool" asked in Germany in German is a different question from the same words in the US.
Step 2: Ask every assistant the same way, every week
You can run questions by hand in each app, but you will not keep it up, and app answers are personalized to your own account. Asking through each provider's API is repeatable: the same model, the same settings, no personal history. Write down which model and whether web search was on, because both change the answer.
Run the full set weekly. Monthly is too slow to see whether a page you published helped; daily mostly measures noise.
Step 3: Record four things per answer, separately
These four observations answer different questions. Merging them into one "visibility score" hides exactly what you need to fix.
| Observation | What it tells you | What it does not tell you |
|---|---|---|
| Your brand is mentioned | The answer text names you | That anyone clicked, or that every app gives this answer |
| Competitors are mentioned | Who you are being compared with | Why the model prefers them |
| Sources retrieved | Which pages the assistant read for this answer | That it relied on them in the final answer |
| Sources cited | Which URLs the final answer links as its sources | A visit or a customer |
A citation matters more than a mention: it is the only one of the four that can send a reader to your page. Keep a receipt of each run (date, model, question, full answer) so you can check a surprising result later.
Step 4: Read the trend, not the run
Your mention rate is the number of questions whose answer mentions you, divided by the number of questions that were answered. Leave out runs that failed or were skipped: a timeout is not a miss, and counting it as one makes you look worse for a reason that has nothing to do with your content.
Compare only runs with the same questions. If you add five new questions this week, your rate will move for that reason alone. Add questions deliberately, and start a new comparison when you do.
What to change when you are missing
When the trend says you are absent, look at what the answers cite instead of guessing. The cited sources show you what kind of page earns a place.
- Answer the question on one page, in the first two sentences. Assistants quote short, self-contained passages. A page that answers in paragraph six is hard to quote.
- Publish the comparison buyers ask for, honestly. If answers keep naming two competitors, a page that compares you with them, including where they are the better choice, is far more citable than a features list.
- Show something nobody else has. Your own data, a worked example or a precise definition gives an answer engine a reason to cite you rather than a summary site.
- Get onto the pages that are cited. If answers lean on a review site or a community thread, being accurately represented there can matter more than another post on your own blog.
- Check that crawlers can reach the page. Pages that are blocked, slow or not indexed cannot be read. Your analytics can show whether AI crawlers request your pages at all.
For Google's AI Overviews and AI Mode, Google says ordinary SEO practice applies: a page must be indexed and eligible for a snippet, and no special AI markup or extra AI text file is required. That is a statement about Google, not every assistant, but it is a good default: build pages that are easy for people to read, and they tend to be easy for machines to quote.
How we do this in TractionScout
Search › AI answers in the TractionScout demo workspace on sample data: which buyer questions name the product, in which engine, and which sources the answers cite.
We run this exact method in TractionScout's AI answers tab. You keep a reviewed list of questions, each with a topic, market and language, and can draft them from the queries Search Console already shows. TractionScout asks them through the OpenAI, Anthropic, Google and Perplexity APIs with your own API keys, and records mentions, competitor mentions, retrieved sources and citations, with a receipt for every run. Failed and skipped questions never count as misses.
Because the same workspace holds your Search Console data, web analytics and sign-ups, you can follow a cited page through to the visitors and customers it brings, instead of stopping at "we were mentioned". See SEO & AI search for the feature, or SEO and AI answer visibility for the full method.
Questions
Is asking through the API the same as what users see in ChatGPT?
Not exactly. Consumer apps personalize answers by account, location and history, and may use different models or settings. API answers are a repeatable baseline you can compare week to week, not a copy of every user's screen.
How many questions should I track?
Ten to twenty is enough to start. Fewer and one question swings your rate; many more and you will not read the answers. Grow the list only when a new question reflects a real buyer problem.
Does an AI mention bring customers?
Not by itself. A mention can shape a buyer's shortlist, but only a citation links to you, and only a recorded visit and sign-up show that it worked. Track all three separately.
Track your own buyer questions in AI answers
See which questions name you, in which assistant, and which pages the answers cite.