Ask ChatGPT to recommend a provider in your category and read the answer carefully. Somebody is being recommended. If it is not you, that recommendation is happening every day, silently, to buyers you will never see in any analytics report. Tracking this is now as basic as tracking rankings.
Build a prompt set from real buyer language
Start with twenty prompts a real buyer would type: recommendation requests, comparisons, and problem descriptions. Pull the wording from sales calls, enquiry forms and the People Also Ask boxes for your money keywords. Write each prompt per market if you sell in more than one country, because answers change with the location framing of the question.
Run the set monthly, the same way every time
Consistency beats scale here. Run every prompt in the same engines each month: ChatGPT, Perplexity, Gemini and Google AI Overviews at minimum. Record three things per answer: whether your brand appears, whether it is cited as a source link, and the exact sentence used to describe you. A simple spreadsheet works at first; the discipline matters more than the tooling.
Score what you find
Use a simple scale per prompt: 3 when you are the recommended answer, 2 when you appear among options, 1 when you are cited as a source but not named as a choice, 0 when you are absent. Sum it per engine and per market. That number moving month over month is your AI visibility trend, and it is defensible in a board meeting.
Read the competitor column closely
For every prompt where you score 0, note who owns the answer and open the pages the engine cites. The pattern is usually visible within an hour: clearer answer formatting, stronger entity signals, or third party coverage the model trusts. That pattern is your fix list, ranked by which prompts carry buying intent.
Connect movement to the work
When a page is restructured into a quotable answer or new schema ships, log the date beside the prompt scores. After a few cycles you can show which changes moved citations, which is exactly the evidence most teams are missing when they ask whether AI search work pays.
The next step
Write ten buyer prompts today and run them in two engines. The first run takes an hour and usually produces at least one uncomfortable finding worth fixing this month.