How to check whether AI recommends your business

You do not need a tool to find out whether AI recommends your business. You need a spreadsheet, about three hours, and the discipline to record what you actually see rather than what you hoped to. Here is the whole method.

Step 1 — Write ten buyer questions

A buyer question is one your potential customer would type without knowing your name. This is the part people get wrong, and getting it wrong invalidates everything downstream.

Two rules:

1. 01No brand names. Not yours, not a competitor’s. “Acme vs Beta” tests whether the model can compare two things it was handed. It cannot tell you whether you would have come up.
2. 02Somebody has to get chosen. If a perfect answer could be written without naming a single business, the question measures nothing. “How does inventory software work?” fails. “Best inventory software for a small distributor?” passes.

Ten is enough to see a pattern and few enough to finish. Spread them across the things you actually sell — if you have three service lines, do not put all ten on one.

Step 2 — Choose your engines

These are not interchangeable, and recording them together is the second most common way to produce a meaningless number.

• ChatGPT, no browsing. What the model absorbed during training. A newer or smaller business is often absent here by definition — that is information, not a verdict.
• ChatGPT with search and Perplexity. Live lookups that cite their sources. Closest to what a buyer doing research actually experiences.
• Gemini. Grounded in Google Search, so it often surfaces different sources again.
• Google AI Overviews. You can look at it in a browser, but there is no API, so you cannot do this systematically or repeatably. Note it separately if you check it at all.

Step 3 — Ask, and keep the answers

Ten questions across four engines is forty answers. Paste each one into a spreadsheet in full — not a summary, not your impression of it. You will want the exact wording in step four, and again in three months when you repeat this and want to know what changed.

One row per answer, with these columns:

• Question · Engine · Date
• The full answer text
• Were you named? (yes / no)
• How were you named? (recommended / mentioned only)
• Every other business named, in the order they appear
• Every URL the answer cited

Step 4 — Classify each answer

Four outcomes. Keep them apart — this is the distinction that makes the whole exercise worth doing.

Do not average these. “Appeared in 60% of answers” is a number that hides whether you were the recommendation or the eleventh name on a list.

Step 5 — Count questions, not answers

Here is the counting rule that stops you flattering yourself: a question counts once. If four engines all name you for the same question, that is one question where you are visible — not four wins.

So your headline figures are:

• X of 10 buyer questions named you on at least one engine
• Y of 10 named a competitor instead
• Z of 10 recommended you, rather than merely listing you

Counting per answer inflates everything and makes engines look like evidence when they are mostly agreeing with each other.

Step 6 — Read who got named instead

Now put your own result aside and list every business the answers named. This is usually the most uncomfortable and most useful part.

Three things fall out of it:

• Competitors you track, appearing where you do not. That is a direct, addressable gap: look at what the answer cited about them.
• Businesses you have never heard of. Either a real competitor you were not watching, or a sign the question is pointed at the wrong market.
• Names that make no sense at all. If a twelve-person software firm sees household-name consultancies in its results, the question is measuring their market, not yours. Drop it and rewrite it narrower.

What the result means

Most businesses doing this for the first time get a worse number than they expect, especially on the search-backed engines.

Two interpretation rules:

• A zero on browsing engines is a content problem, not a reputation problem. The engine found nothing to cite. That is fixable and fairly directly.
• A zero on model knowledge and a decent browsing score is normal for a young business. Training data lags. Do not read it as a verdict on your brand.

When to stop doing it by hand

Once is genuinely useful — it tells you where you stand today. The trouble starts when you want to know whether anything changed, because a comparison is only valid if the question set and the engine list were identical both times. One edited question invalidates the comparison, and by the third run nobody remembers which questions were edited.

Do it by hand if you want a snapshot. Automate when:

• You want to compare runs rather than read one
• You are tracking more than about ten questions
• You need the answers stored verbatim to prove what changed
• You are doing this for clients and need it to be repeatable

And do not run it daily. We tested that: repeating the same 31 questions back to back produced zero changes in whether the business was mentioned. Weekly is the shortest interval where a moving number means something actually moved.