You can run this yourself, today, for nothing. It takes about twenty minutes and needs no tool, no trial and no credit card. At the end you will know whether the assistants name your business, who they name instead, and which pages they are reading to decide.
Disclosure: I build a tool that automates this. The manual method below is the whole method, written out. If you only need the answer once a quarter, do it by hand and keep your money.
First, stop asking it about yourself
The instinct is to open ChatGPT and type what do you think of Northgate Plumbing. That tells you nothing useful, for two reasons.
You have named the company, so the model will discuss it whether or not it would ever have recommended it. And you have asked a question no customer asks. Nobody types a brand name they do not know yet. They describe a problem.
The only question worth asking is the one a stranger would ask, phrased the way they would phrase it, with your name nowhere in it.
Set up a clean room
Assistants personalise. If you have been talking to one about your own company for months, it will name you, and that reading is worthless.
- Log out, or use a temporary chat that is excluded from memory.
- Turn off memory and custom instructions if your account has them.
- Check the location the assistant thinks you are in. For anything local this is the single biggest variable. Some tools show it, some infer it from your connection, and a VPN will quietly move you.
- Use one fresh conversation per prompt. Follow-ups inherit context from the answer above them.
Write ten prompts you might lose
Ten is enough to see a pattern and few enough that you will actually finish. Spread them across the ways people arrive:
| Type | Example shape |
|---|---|
| Direct need, local | Who should I call for an emergency plumber in Leeds tonight |
| Direct need, national | Best project management tool for a five person agency |
| Comparison | What are the alternatives to the biggest name in your category |
| Qualified by constraint | Cheapest option that still does X |
| Qualified by audience | What do small law firms use for this |
| Problem first, no category | My rankings look fine but enquiries dropped, what now |
The temptation is to write prompts you are confident about. Resist it. A prompt set you win is a mirror, not a measurement. At least half should be ones where you genuinely do not know the answer.
Run each prompt three times
This is the step most people skip, and skipping it is why so many of these audits are wrong.
These models are not deterministic. Ask the same question three times and you can get three different lists. A single run tells you almost nothing; three runs tell you whether your absence is a pattern or a coin flip. Ten prompts, three runs, three assistants is ninety readings, which is why twenty minutes is the honest floor rather than the estimate.
Run the same set through ChatGPT, Gemini and a plain Google search that returns an AI Overview. They use different retrieval systems and they disagree with each other more than people expect.
Record four things, not one
Open a spreadsheet with a row per run and these columns:
- Named, yes or no. The floor. Everything else is detail on top of it.
- Position in the answer. First, second, third, or buried in a list at the bottom. The first name gets the paragraph; the rest get a comma.
- How you were framed. Write the actual clause. A cheaper option if you can live with fewer features is a mention and it is not a good one.
- Who was cited. Copy the source list. This is the most valuable column and almost nobody fills it in.
At the end, count. If you were named in fewer than three of thirty runs, you have an absence problem, not a positioning problem, and the fix is different.
The citation column is the actual deliverable
Look down the sources you collected and you will almost always find the same handful of pages carrying your entire category. A comparison article on some industry blog. A Reddit thread from two years ago. A directory nobody in your office has heard of.
Those pages are what the model reads before it answers. Which means:
- If you are absent from those pages, you are absent from the answer, no matter how good your own site is.
- Getting added to one of them moves the needle further than a month of on-page work.
- The list is finite, and you now have it written down.
This is closer to digital PR than to SEO. The lever is somebody else’s page, not yours.
What the manual method cannot do
Being straight about this, because the honest limits are the reason tooling exists at all.
You are taking one reading. Doing it once tells you where you stand today, and today is not a trend. Answers shift as models refresh their indexes, and the interesting question is whether last quarter’s work moved anything.
You also cannot realistically hold location steady by hand. For a business with three branches, the honest version of this audit is three location settings times ten prompts times three runs times three assistants, and at that point you are doing data entry rather than marketing.
So: run it manually to find out whether you have a problem. Automate it only once you have decided the problem is worth watching every month.
What to do with a bad result
In rough order of effort against payoff:
- Get onto the cited pages. Pitch the comparison articles that already rank in your category. Unglamorous, effective.
- Publish the comparison nobody wrote. If the model keeps citing a thin listicle from 2023, a better one is a genuine opening.
- Make your own pages answerable. Clear claims, real numbers, plain sentences. Models extract statements, not atmosphere.
- Fix the boring surfaces. Business profile, categories, consistent details. For local prompts these still carry weight.
- Re-run the same ten prompts in ninety days. Same set, same wording, or the comparison means nothing.
Common questions
Do I need a paid tool for this?
Not to find out where you stand. A spreadsheet and twenty minutes gives you the same answer for a single point in time. Tooling earns its cost when you need the trend, multiple locations, or a report a client will read.
Will the assistant give me the same answer as my customer gets?
Close, not identical. Location, account history and the model version all shift it. Logging out and setting the location deliberately gets you most of the way there.
How often should I re-check?
Quarterly is enough for most businesses. Monthly if you are actively working on it. Weekly readings mostly measure the model’s own randomness.
My competitor is named every time. What are they doing?
Usually nothing clever. They appear on the pages the model reads. Check your citation column against their name and the mechanism is normally obvious within ten minutes.
Does this replace checking my rankings?
No. Classic results still carry most commercial intent, especially local. Treat this as a second surface, measured separately.