How to check what AI says about your business

Open a fresh chat, type your trade and your city the way a customer would, and read what comes back. Do that twice in ChatGPT and twice in Claude, three runs each, and you have twelve answers, which is the same depth our own probe stores for every business we measure. It takes about ten minutes. What follows is how to phrase it, why the repetition matters, and how to read the result.

How our probe asks

Every business in a city cohort gets twelve stored answers. That is two questions for its trade, put to two engines, three times each. The engines are the API models behind ChatGPT (OpenAI's gpt-4o-mini) and Claude (Anthropic's Claude Haiku 4.5). Every answer is kept in full, so any number we publish can be traced back to the text it came from.

The questions are phrased the way a customer phrases them, not the way a marketer would. In the nine wedding cities we have published, they were: best wedding venue in the city, where should I get married there, best wedding planner, who should plan my wedding, best wedding officiant, who should officiate my wedding. Short, ordinary, and none of them mention any business by name. Between 21 July and 11 August 2026 that produced 5,904 stored answers.

The manual version

You can run the same shape by hand.

1. Write down two questions in your customer's words. One evaluative: best electrician in Tempe, AZ. One urgent or situational: my circuit breaker keeps tripping in Tempe, AZ, who should I call. Use the city as a customer would name it.
2. Open a new chat for each run. A chat that already knows what you asked a minute ago is not a fresh customer. If your assistant has memory or custom instructions turned on, turn them off for this, or use a logged-out or temporary chat. An assistant that knows who you are will name you, and that tells you nothing.
3. Ask each question three times, in ChatGPT and in Claude. Twelve runs.
4. Keep a note as you go: the date, which assistant, the exact question, and every name in the answer. That note is what makes a re-check in a month comparable to today.

Why the trade and the city, never your business name

Because the customer you are worried about does not know your name yet. Someone who already knows it searches for it directly and finds you. The question that decides whether you get new work is the one that names a job and a place.

Typing your own business name into an assistant tests something else entirely: whether it can describe you when told who you are. That is a nice thing to know, and it is not the same as being recommended. Our own counts put the gap in plain numbers. Of 2,232 stored answers to who-to-hire questions across eight cities, six named the business the question was asked about.

Why one answer proves nothing

Ask the same question twice and you will often get two different lists. That is normal, and it is why we set a floor before publishing anything: no business is counted as absent on fewer than three stored answers per engine. Mixed evidence gets reported as mixed rather than rounded in the direction we would prefer.

So treat a single run as one sample, not as the answer. What you are looking for across twelve runs is a rate. Named in nine of twelve is a different situation from named in one of twelve, and neither of those is visible if you only ask once.

Named once is not recommended

When you count your hits, count them the way a customer would experience them. Being mentioned in passing at the bottom of a list, in one run out of twelve, is not the same as being the name the answer opens with. In our own counts an answer scores once for a name however many times it repeats it, precisely so that a chatty answer cannot look like an endorsement.

How to read what came back

Three outcomes, and each means something different.

Your name appears. Note in how many runs and on which engine. Engines disagree, and a name that shows up on one and never on the other is a real result worth writing down.

Nothing local is named, and a directory is. This is the common one. Across those 2,232 hire answers, 1,407 named a directory. The Knot appeared in 1,371 of them and WeddingWire in 1,332. Read that as a fact about the answers rather than a complaint about the directories: they publish their information in a form a machine can read, and the answers reflect that.

Big, famous places are named instead. When we asked where to get married rather than who to hire, only 383 of 3,672 answers named a directory, and the answers filled up with landmarks and large venues. If your result looks like that, the assistant is not short of local knowledge. It just does not know you well enough to put your name behind.

What asking cannot tell you

Asking measures an output on one day, from one model version, and it moves for reasons that have nothing to do with you. It also gives you nothing to act on. You cannot fix an answer.

The other half is what your website makes readable, and that half you control. Our readiness audit fetches a site the way a plain AI crawler does, without running JavaScript, and scores four things: crawl access at 30, structured data at 25, legibility at 30, contactability at 15. It is deterministic, so the same site scored twice returns the same number, and every deduction quotes the exact words it read on your page. When your score moves month to month, the movement is your site changing rather than the tool changing its mind.

We keep the two measurements apart on purpose. Whether an assistant names you is a visibility probe, reported on its own and never folded into a readiness score, because blending them would hide which one actually moved.

Run both. The twelve chats tell you where you stand in the answers today. The free audit on our front page tells you what an AI reader can actually see when it arrives at your site, with the line behind every deduction quoted, so your web person can start on it the same afternoon.