Most companies have never asked an assistant what it says about them.
The ones who do are rarely reassured. The common outcome is not silence, it is a confident, fluent description that is wrong: an outdated product line, a competitor’s positioning, a market you left two years ago.
You can find out in about an hour. Here is the method we use.
Before you start
Two rules, or the results are useless.
Use a logged-out session. Assistants personalise. If you are signed in and have spent months discussing your own company, you will get a flattering answer nobody else sees. Use a private window, or a fresh account.
Do not lead the witness. “Why is Acme the best CSPM tool?” tells you nothing. Ask what your buyers ask, in their words, without your name in it.
Step one: write the questions your buyers actually ask
Not your keywords. The sentences a buyer types when they do not yet know you exist.
Aim for twelve to fifteen, spread across four types:
- Category discovery. “What tools help with X?”
- Problem-first. “How do I stop Y happening in Z environment?”
- Comparison. “Best alternatives to [the market leader].”
- Direct. “What does [your company] do?” and “Is [your company] any good?”
Your sales team knows these better than your marketing team does. Ask them what buyers say on first calls, and use those words.
Step two: run them across the assistants that matter
At minimum: ChatGPT, Perplexity, Google’s AI Overviews and Gemini. Add Claude and Copilot if your buyers are enterprise.
Run every question in every assistant. Yes, it is tedious. It is also the whole point: the answers differ more than people expect, and a company that is invisible in one is often prominent in another.
Step three: record four things, not one
This is where most audits go wrong. People record whether they were mentioned and stop.
Record all of these:
- Mentioned or not. The obvious one.
- Position in the answer. Named first, named among five, or mentioned in passing at the end. These are very different outcomes.
- Accuracy. Is the description right? Score it: correct, partly wrong, or wrong. Then note exactly what is wrong.
- Sources cited. Which pages did it draw on? This is the most actionable column on the sheet and almost nobody fills it in.
Put it in a spreadsheet, one row per question per assistant. It will be ugly and it will be the most useful document your marketing team produces this quarter.
Step four: do the same for two competitors
Without this, you have numbers and no meaning. Being named in four of fifteen answers sounds poor until you learn the category leader manages six.
Pick one competitor above you and one at your level. Run the same questions. The gap tells you whether you have a visibility problem or the whole category is under-represented, and those need completely different responses.
Step five: read the sources column
This is where the audit turns into a plan.
Look at what the assistants actually cited. You will usually find one of three patterns:
- They cite your competitors’ pages to describe your category. Whoever wrote the definitional content owns the framing, and right now it is not you.
- They cite third parties, not you. Directories, review sites, forums. Your own site is not being treated as authoritative about you, which is usually an entity clarity problem.
- They cite something outdated. An old press release, a stale listing, a conference bio from three years ago. This is the easiest and most satisfying thing to fix.
What to do with the results
In priority order, because they are not equally urgent:
- Fix wrong descriptions first. Being described incorrectly costs more than being absent. Trace each error to its source and correct it there.
- Fix your entity data. Make your own site the clearest, most consistent description of what you do. Same wording everywhere you appear.
- Write the definitional content you are missing. If competitors own the “what is X” answer, that is a page you can write better.
- Build corroboration. Get third parties to describe you the way you describe yourself.
Then do it again
Once is a snapshot. The models change underneath you, sometimes dramatically, and so does what they have read.
Re-run the same questions quarterly with the same spreadsheet. The trend line matters more than any single result, and it is the only way to know whether anything you did worked.
If that sounds like more discipline than your team has spare, it is part of what we do in GEO Strategy. But the method above is yours either way, and running it once will tell you more than any article can.
