AI search · 6 min read
Find out what AI assistants say about your business in half an hour
To find out whether ChatGPT knows your business, ask it ten questions your customers ask, record the answers and repeat monthly. The errors you find are what to fix.
Rather check it yourself? Try our free SEO tools
An AI answer
ExampleWho fixes burst pipes near me tonight?
Your Business Plumbing offers 24-hour emergency call-outs and lists its service area and prices on its site.
- yourbusiness.comCited
- a local directory
- a review site
The short answer
You can’t look this up, so ask it and write it down
There’s no ranking report for AI assistants, because there’s no position in ChatGPT to track. Each answer is generated for that question, in that session, from whatever the system retrieved at the time.
So the only reliable way to find out what an assistant says about your business is to ask it, record the answer, and ask again next month. The first run takes about half an hour; after that, ten minutes.
What comes out isn’t a score. It’s a list of specific things you can fix: a service area you don’t cover, a competitor named in a question you should own, a price nobody at your company has ever quoted.
The routine
The test, in three steps
Do these in order. The value comes from repeating exactly the same thing next month.
Write ten questions a customer would type
Start before anyone knows you exist. "Emergency plumber near me on a Sunday." "Who inspects commercial flat roofs in Calgary." "How much does a fence replacement cost in Ottawa." Then add two that use your company name, to see how you’re described when asked about directly. Eight category questions, two brand questions.
You get: A fixed list of ten prompts you’ll reuse every month
Run them in a clean session, one assistant at a time
Use a logged-out or temporary chat where offered, so your own history doesn’t shape the result. Put the same ten prompts through ChatGPT, a Google search that returns an AI Overview, and one other, usually Perplexity or Gemini. Don’t rewrite a prompt to get a better answer: the prompt a customer types is the one that counts.
You get: Thirty raw answers, pasted in unedited
Record four things about every answer
Were you named? Was anything said about you wrong? Which competitors were named? Which sources were cited, where shown? One row per answer, with the date and the assistant first. Don’t summarise: the raw text is what lets you compare next month.
You get: A dated sheet that can be compared month to month
Reading the sheet
Four results, and the first move for each
Almost every row falls into one of these, and each needs a different first move.
| Dimension | What it usually means | First move |
|---|---|---|
| You aren’t named at all | There’s little about your business that’s easy to find, quote or check against another source. | Give every service you sell a plain page that states what it is, who it’s for and where you do it. |
| Named, but the details are wrong | You were found, and what was found is stale or contradicted somewhere else. | Fix the source before the page. Business profile hours, address and services first, then the same facts on your own site. |
| Competitors named, you aren’t | They have more of the kind of evidence these systems can read and are willing to repeat. | Look at what the answer cited. Work out which of those sources could legitimately mention you, and earn that. |
| A claim you never made | A gap in the record was filled with something plausible, or a third-party listing was taken at face value. | State the real fact on a page of your own, and correct the listing it came from. |
The record
The eight columns worth keeping
Anything beyond this becomes work nobody repeats, and a record nobody repeats is worth nothing.
- Date, and the assistant used, including the model version if it’s shown
- The exact prompt, copied, not retyped
- Named or not named, as a plain yes or no
- Any factual error, quoted word for word
- Competitors named, in the order they appeared
- Sources cited, where the assistant shows them
- The full answer text, pasted into one cell
- One sentence on what you changed as a result, or nothing if you changed nothing
The useful part
The sources column is where the work comes from
Most people read these answers looking for their own name. The column that repays attention lists what the assistant cited: pages the system is willing to believe about your category.
It’s usually not what an agency would predict. Trade association directories, a local newspaper piece from four years ago, a supplier listing, a forum thread, a government register. For most there’s a legitimate way to be present: a membership that comes with a listing, a register you’re eligible for, a journalist who covers your area.
That work is slow, and it’s the part that moves. Publishing another blog post is neither.
Limits
What this test can’t tell you
It can’t tell you how many people asked. There’s no query volume for assistant prompts, so a strong showing on a question nobody asks looks identical to one on a question everybody asks.
It can’t tell you how much business it produced. Some assistant traffic shows in your analytics by referrer; the rest arrives as someone typing your name into a browser a day later, which no system attributes to anything.
It can’t give you a share of voice figure, whatever a tool reports. Counting how often you appear in ten prompts is a sample of ten prompts. Treat single-run differences as noise, look for patterns over three months, and act on the errors, not the score.
Questions
Common questions about testing AI answers
Why does the same prompt give different answers?
The answer is generated fresh each time, not looked up. Temperature, the pages retrieved at that moment, your chat history and the model version all change the output, and any of them can change without notice.
That’s why this test tracks a pattern across months, not a position on one day. Named in eight of ten runs one month and two of ten the next is a signal. One different answer isn’t.
Do I need a tool to track this?
No, and it’s worth doing by hand first. A spreadsheet and thirty minutes gets you the same four facts a paid tool reports, plus the raw answers, which are what you learn from.
Tools become useful when the prompt list grows past about thirty, or when several people need the same record. Before that they mostly add a number that can’t be verified.
Should I ask the assistant about my own brand name?
Include two brand prompts out of ten, no more. A brand prompt shows whether your business is described accurately, but almost nobody who hasn’t heard of you types your name.
The other eight show whether you exist in the category at all, the harder and more valuable question.
What if the assistant states something about us that’s untrue?
Write it down word for word, then find the source. It’s usually one of three things: a stale directory listing, an old page of your own that still says something you stopped doing, or a plausible guess filling a gap where your site says nothing.
The last case is the most common with prices and service areas. If your site doesn’t state them anywhere, something else will.
How often should I repeat the test?
Monthly is enough for most businesses, with the same prompts on roughly the same day. The value is in comparison, so changing the prompt list breaks the record.
Add prompts when you add services. Retire one only when the service is gone, and note in the sheet why it went.
Run the test, then bring us the sheet
If the ten prompts turn up errors you can’t trace, or competitors you can’t explain, send us the sheet. We’ll tell you what’s worth acting on and what’s noise, on a call, before anyone quotes you anything.
Rather check it yourself? Try our free SEO tools
References
Sources
The official documents and research this page is based on. Platform rules change, so check the source before you act on a detail.
Last updated