Originally published at https://thicket.sh/blog/check-if-ai-recommends-your-product.
← Back to Journal
Method
Being cited by an AI and being recommended by one are different standings, and most teams measure neither. Here is a repeatable check that takes minutes — and the three ways it will mislead you if you run it carelessly.
By Thicket AI Team · August 1, 2026 · 7 min read
Ask the model the question your buyer would actually ask — the category question, never your brand name — and record which products it names. Run it three or more times, because output varies between samples. If you never appear, that is a real and useful signal. What it is not is a measurement of how often assistants cite you as a source, which is a separate thing that can move in the opposite direction.
We built this check for a practical reason: we wanted to tell companies truthfully whether an assistant named them, and we were not willing to assert it without running it. What follows is the method and, more usefully, its failure modes.
The check
- Write the buying question, not your brand. “What is the best SQLite editor for macOS?” — not “What is Base?” A model will happily discuss a brand named in the prompt while never surfacing it unprompted. That difference is the entire measurement.
- Ask for specifics. Append something like “list the specific products you would recommend, with a one-line reason for each.” You want a list you can check a name against, not an essay.
- Run it at least three times and count how many runs named you. One run is a coin flip; a count is a position.
- Record who was named. The competitor list is more actionable than your own absence — it tells you which material the model actually learned from.
- Match on brand and domain, with word boundaries. A short brand like “Base” will match inside “database” if you are careless, and you will record a false presence. ## The three ways this misleads
1. It is one model, not “AI”. A result from one assistant answering from training knowledge says nothing about the others, and nothing about live web-grounded search. Absent in one place does not mean absent everywhere. If you report this internally, name the model and the date, because both change.
2. A single run is noise. Sampling variance is real and large enough to flip a verdict. We treat any check that produced zero usable responses as an error rather than an absence — because the failure mode we most wanted to avoid was reporting “they were not named” when the truth was “the request failed.” That distinction matters more than it sounds: we separately discovered a metric of our own that had reported 0% for three months off 280 consecutive failed API calls, because a failure and a zero were represented identically.
3. Citation and recommendation are not the same thing. This is the one that surprises people. Our own portfolio is cited roughly 14,394 times in 28 days by one assistant — and a different model names us for zero of 40 category questions in our own niches. Both are accurate. Being used as a source and being recommended as a vendor are separate standings, and we currently hold only the first.
Validate the checker before you trust it
A test that always returns “absent” would look identical to a working one right up until you acted on it. So run a control: pick a company that obviously is well known in its category and confirm the check finds it.
When we did this, a household-name workspace tool came back named in every run, and a jobs marketplace we expected to be absent turned out to be named as well — correctly excluding it from our own outreach. Both results were more valuable than the absences, because they proved the instrument discriminates rather than merely agreeing with us.
If you are not named
The usual cause is a thin third-party footprint rather than any penalty. Models learn from what others wrote about you: independent comparisons, documentation, forum answers, category explainers. Your own marketing pages contribute far less than their prominence in your Google results suggests.
In our own data, the content that gets picked up shares one property — it answers a question with a specific, checkable, quotable answer. Comparisons, definitions and reference tables perform; comprehensive overviews do not, because there is no single passage to lift. We broke that down in what kind of pages AI assistants actually cite.
Frequently Asked Questions
How do I check if ChatGPT or Gemini recommends my product?
Ask the model the question your buyer would actually ask — the category question, not your brand name — and record which products it names. Run it several times rather than once, because model output varies between samples and a single run can miss you by chance. Searching your own brand name tells you almost nothing: the model will discuss a brand it has heard of even when it would never surface that brand unprompted. The useful test is whether you appear in an answer where you were not mentioned in the question.
How many times should I run the query?
At least three, and treat the result as a count rather than a verdict. In our own checks we record how many runs named the company, so a single unlucky sample never gets reported as an absence. Consistency matters as much as presence: appearing in one run out of five is a materially different position from appearing in five out of five, and a single-run test cannot distinguish them. If you are tracking this over time, hold the question wording and the run count fixed, or you will measure your methodology instead of your visibility.
Does being absent from an AI answer mean I have an SEO problem?
Not necessarily, and conflating the two is the most common mistake here. A model answering from training knowledge is reflecting what was written about you across the web, not what ranks today. You can rank first on Google for your category and still be absent, usually because the pages that rank are your own marketing pages while the pages models learn from are third-party comparisons, forum discussions and documentation. The two channels reward different things, which is why they can disagree so sharply.
Is being cited by an AI the same as being recommended by one?
No, and the distinction is worth holding onto. Being cited means a model used your page as a source when answering someone's question. Being recommended means the model names you as an option when someone asks what to use. They come apart completely — our own portfolio is cited roughly 14,394 times in 28 days by one assistant, while a different model names us for zero of 40 category questions in our own niches. Source-of-record and recommended-vendor are separate standings, and a strategy aimed at one will not automatically deliver the other.
What should you do if an AI does not name your product?
First check whether the material a model could learn from actually exists. Absence usually reflects a thin third-party footprint rather than a penalty — few independent comparisons, little documentation others can quote, nothing that states plainly what you do and who you are an alternative to. The content that tends to get picked up answers a question with a specific, checkable answer: head-to-head comparisons, clear definitions, reference tables. Pages written to be comprehensive rather than answerable do worse, because there is no single passage a model can lift and attribute.
Method and caveats
Our checks run against a single Gemini model answering from training knowledge — not a live web search, and not Copilot, ChatGPT, Claude or Perplexity. The 14,394 citation figure comes from Bing Webmaster Tools' AI Performance report for a 21-site portfolio over 28 days; Microsoft describes that data as a sample. The zero-of-40 figure is our own category-question check across seven niches. Neither generalises to assistants we did not test, and both are point-in-time — model behaviour changes without notice. For how the citation side is measured, see how to measure whether AI assistants cite your site; our full operating numbers, including what failed, are in the Thicket Report.
More from the Journal
Top comments (0)