How to track whether AI recommends your business.

What AI recommendation tracking means.

AI recommendation tracking is checking, on a schedule, whether AI assistants name your business when someone asks them what to buy. The assistants are ChatGPT, Gemini, Perplexity, Claude and the AI Overviews at the top of Google results. The question is a buyer's question, such as "What's the best invoicing app for freelancers?", and the result is your name in the answer or somebody else's.

The phrase has an older meaning, and it causes trouble. For years "AI recommendations" meant the product suggestions a shop's own recommendation engine shows its visitors. Gemini still reads it that way some of the time. We asked it "How do I track AI recommendations for my business?" once a day for nine days, and in two of those answers it searched for things like "ai recommendation system performance metrics business". Across four related questions, 16 of 42 answers ran at least one search about recommendation engines or personalization.

This article means the first sense: an assistant naming your business to a buyer.

Decide what counts as being recommended.

An answer can include you in three ways, and each one is worth recording on its own.

• Named. Your business appears in the answer text.
• Recommended. It appears as one of the answer's picks: in a list, a heading or bold text, as opposed to a passing aside.
• Cited. A page on your site appears among the sources the assistant links to.

A citation without a mention happens when the assistant reads your page for a fact and then names a competitor. A mention without a citation happens when the model already knows your name. Both matter, and they call for different fixes.

Write the questions your buyers ask.

Track the questions a buyer types before they know your name. A brand question such as "Is Acme any good?" shows how the assistant describes you. It cannot show whether you get recommended to someone who has never heard of you.

Good tracking questions share three traits:

• They name the category and the job, not your product: "What's the best CRM for a two-person agency?"
• They carry the buyer's situation when it changes the answer: "I'm a marketing manager and I want to track AI recommendations. What are the best tools for that?"
• They read the way people type into a chat box, in full sentences, not as a keyword string.

Start with 10 to 20 questions. Below that, one odd answer moves your whole number. Above it, a manual process falls apart in the first week.

Ask each question more than once.

The same question gets a different answer from one day to the next. We tracked four questions about AI visibility tools for ten days, from 17 to 26 September 2026. For "How do I track AI recommendations for my business?", Gemini named 18 different businesses across eight answers, between none and six per answer, and no business appeared in all eight. The other three questions behaved the same way: no business appeared in all the answers to any of them.

So one check in a chat window proves little. If your name shows up once, you might be in one answer out of four. If it does not, you might be in the next one. Tracking means asking the same question on a fixed schedule and reading the share over a window. "Named in 3 of 14 answers over two weeks" is a result. "It mentioned us on Tuesday" is an anecdote.

Chat apps add a second problem. ChatGPT and Gemini can personalize answers with your history and saved memories, so a check from your own account shows what the assistant tells you. Use a clean session, or the API, which answers without your history.

Record the answer, the names and the sources.

For each question and each run, keep:

• the date and time, and which assistant answered;
• the full answer text, not a summary;
• all the businesses the answer named, and whether it recommended them or mentioned them in passing;
• the pages it cited as sources;
• the searches it ran before answering, where the assistant exposes them;
• whether it searched at all.

The full text matters because counts drift away from what a reader sees. If a tool reports that you were "mentioned", open the answer and find the sentence. Names get matched loosely, and a partial match on a common word can count as a mention that no reader would notice.

The last item on the list matters for a different reason. An assistant that answers from memory, without searching, cannot be moved by anything you publish this month. Keep those answers, and leave them out of the rate you calculate from answers that searched.

Read the searches the assistant ran.

An assistant that searches before it answers can run several searches for one question. Google calls this "query fan-out" for AI Overviews and AI Mode (Google Search Central), and the Gemini API returns the searches the model ran alongside a grounded answer (Gemini API documentation).

Those searches show which pages the assistant went looking for. For our four questions Gemini ran 80 different searches across 42 answers, a little over three per answer. The most frequent:

A search that comes up in most answers is one a page of yours can be found on. Most of the pages that won these searches were lists of tools. Zapier's roundup of AI visibility tools was cited in 13 of the 42 answers, and a Marketer Milk list of AI marketing tools in 19. The full breakdown is in who Gemini names when you ask for an AI recommendation tracker.

Track by hand first, then decide.

A spreadsheet works for the first two weeks. One row per question per day, the columns from the list above, and the answer text pasted into a linked document. At 15 questions that is 15 chats a day for each assistant you track.

It breaks down in three places. The copying takes longer than the reading. Searches and citations are hard to get out of a chat window. And a rate you can trust needs more runs than anyone keeps up by hand. At that point a tool earns its price, and we compared the options in the best AI recommendation tracking tools.

Act on what the tracking shows.

Tracking pays off when it tells you what to do next. Sort your questions into four groups:

1. A competitor is named, and the answer cites their page. The most useful group. The assistant searched, found a page that answered the question, and named the business behind it. Write the page that answers the same question better, on your own site.
2. Nobody in your category is named. The assistant answered with general advice. A clear page that answers the question is often the only candidate it finds.
3. You are named but not cited. The assistant knows you. Check that what it says about you is true.
4. The assistant did not search. It answered from training data, and no page published this month changes that answer. Keep tracking it, and spend your effort on the other three groups.

Then measure again the next day. The numbers that tell you whether any of it worked are in how to measure whether AI recommendations are working.