If you have five minutes and want to know whether AI answer engines will put your company in front of a buyer, run one prompt: ask for the best tools for the job your product does, without mentioning your brand at all. Whether your name appears in the answer, and where, tells you more in thirty seconds than any dashboard, because it is the exact question a buyer asks before they know you exist.
The reason this single prompt works so well is that it is unaided. You are not asking the engine to describe a company you named. You are asking it to choose, which is the only version of the question with commercial consequences.
The prompt, and the rules for running it
Replace the bracketed part with the job your product does for a customer, phrased the way a buyer would say it rather than the way your positioning deck says it.
What are the best tools for [the job your product does]?
# worked examples
What are the best tools for reviewing contracts before signature?
What are the best tools for tracking infrastructure spend across cloud accounts?
What are the best tools for onboarding new engineers in the first week?Four rules, and skipping any one of them gives you a result you cannot trust.
- Do not use your brand name. Mentioning it turns an unaided question into a prompted one and guarantees your name appears. That measures nothing.
- Use a logged-out or temporary session. Assistants with memory and chat history personalise answers. If you have been discussing your own company for six months, your account is the worst possible place to test this.
- Do not lead. Asking whether your product is good at something tells you how agreeable the model is, not whether you are visible.
- Run it in more than one engine. Each maintains a separate corpus, so a result from one tells you very little about the others.
The second rule catches almost everybody. The first time most founders run this properly, in a clean session, the result is noticeably worse than what they had been seeing in their own account, and that gap was the whole problem with their previous testing.
One prompt, four engines, and a clean session each time.
That is the minimum version of this test, and it takes about five minutes if you have the tabs open. It is also the version most people skip straight past, which is why so much internal reporting on AI visibility is measuring a personalised session rather than the market.
If you would rather have it run properly across engines with the raw answers handed back, that is the first thing our free AI visibility check does.
Reading the result
There are five outcomes and each points somewhere different. Find yours before drawing any conclusion.
| What you got back | What it means | What to do next |
|---|---|---|
| Named in the first few options | You are in the consideration set and the machinery is working | Move on to whether the description is accurate |
| Named far down a long list | Present but not preferred, usually a corroboration gap | Look at reviews and independent comparisons |
| Not named, competitors named | The most common result. Presence or extractability | Check crawler access before touching content |
| Nobody named, generic advice only | The question is too broad for the engine to answer with brands | Narrow to a buyer type and run it again |
| Named, but described wrongly | Not a visibility problem at all | Correct the record, this is a representation issue |
One more read that people miss. Run the prompt twice on different days before believing any of it. The retrieved documents shift as indexes update and the model samples probabilistically as it writes, so two runs that disagree is normal rather than a sign something broke. If a single run is all your reporting rests on, the number is noise.
The four ways this test gets run wrong
We see the same mistakes repeatedly, and each one produces a falsely reassuring answer.
Testing in a logged-in account. Covered above and worth repeating, because it is the single most common contaminant and it always flatters you.
Using your category name rather than the job. Buyers who do not know you exist rarely know your category name either. They describe a problem. Asking about contract lifecycle management software finds vendors who have already won the naming argument, which is not the test you wanted.
Running it once and screenshotting the good one. Understandable and useless. If you run five times and share the best result internally, you have built a reporting system that cannot detect bad news.
Stopping at one engine. Absence in one and presence in another is a genuinely useful finding, because it points at corpus presence rather than content quality. You only get that finding by checking more than one.
Why unaided is the only version worth running
Two findings explain why this specific prompt earns its place over the more comfortable alternatives.
The first is about how people actually behave. Nielsen Norman Group ran a qualitative usability study on how AI is changing search behaviour and found participants bringing their own real research tasks to assistants alongside traditional search rather than instead of it. Nine participants is a small qualitative sample and not a market estimate, and the texture is what matters. People arrive describing a problem, not naming a vendor, which is exactly the shape of the prompt above.
The second is about what position in the answer is worth. Research testing brand recommendation behaviour across three major models found that when products had identical specifications the established brand was recommended 100 percent of the time, with that dominance collapsing once a competitor held even a tenth of a star advantage in rating. That study covered consumer goods rather than B2B software, so treat the mechanism rather than the number as transferable.
Put together they say something useful about your result. Being named late in a list is not a small version of being named first. It is a different position entirely, and it usually reflects corroboration rather than anything on your own site.
Recording it so the answer is worth something next quarter
The test takes five minutes. Recording it properly takes another three and is the difference between a check and a baseline. Write down six things every time.
| Record | Why it matters |
|---|---|
| The exact prompt text | Rephrasing between runs makes two results incomparable |
| The date and the engine | Corpora update constantly, so an undated result decays into an anecdote |
| Session state, logged out or otherwise | A personalised session is the single most common contaminant |
| Every company named, in order | Order is the finding, not a detail |
| Whether you were named at all, as yes or no | The count you will actually track over time |
| Anything factually wrong in the description | Catches representation problems the visibility count misses |
Do that twice a month against a frozen prompt and within a quarter you have something no dashboard sold to you can provide, which is your own longitudinal record with the raw answers behind it. Then when somebody claims a change moved your AI visibility, there is a before to check it against.
When five minutes is not enough
This prompt tells you whether you have a problem. It does not tell you which of several possible causes you have, because absence looks identical whether you are blocked at the crawler, invisible to rendering, ambiguous as an entity, or simply not discussed anywhere.
If the answer came back badly, the next step is the longer diagnostic rather than a content project. The full walkthrough works ten prompts and five causes in the order that isolates which one is actually binding, and the first two checks in it are free and frequently decisive.
Resist one specific temptation in the meantime. The instinct after a bad result is to commission content, because content is the thing marketing teams know how to buy. If the cause turns out to be a crawler that has been blocked since 2023, every article you commission in the interim is invisible, and you will have spent a quarter proving that writing more does not help.
Run the free checks first. They cost an afternoon and they decide whether the next conversation is with an engineer, an editor or nobody at all.
One last thing worth saying, because founders take this result personally more often than any other metric we discuss. Not being named for a category question is the normal starting position for almost every company that has not deliberately worked on it, including good products with happy customers. It is a measurement of machine visibility rather than a verdict on the business, and the first two causes behind it are usually things nobody chose.
One prompt will not tell you everything, and it will tell you whether you have a problem worth spending money on. For most companies that is the only decision that needs making this week.
Related reading
- What a knowledge graph is. What one actually holds, and why an engine consults it before it will name anybody.
- AEO vs SEO. Why the unit changed from a page to a passage, and the prize from a click to a mention.
- The AEO glossary. Every term defined, each marked by whether it comes from a real standard or from marketing.
Five minutes now beats a quarter of assuming.
Run it logged out, in four engines, twice. Whatever comes back is the most honest read you have had on this, and it costs nothing but the tabs.
If it comes back badly and you want the cause rather than the symptom, send us the domain for a free AI visibility check. You get the raw answers, the log finding and a named binding constraint.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.