x
Most founders have never asked an AI assistant what it says about their own business. Nine questions, two assistants, fifteen minutes, one honest list.
Ask two AI assistants what they say about your business, write down what they get wrong, and stop there.
Fifteen minutes, no tools, no spend. The stopping is part of the method — the temptation is to start fixing at question three, and that costs you the only clean read you'll ever get. After this you will know too much to be surprised again.
Here's the sequence.
Ask by name. What do you know about [business name]? Then: what does [business name] do, and who for? You're checking two different things — whether it knows you exist, and whether what it knows is right.
Ask by need. Who does [the actual job you do] for [the actual kind of client]? Phrase it the way a buyer who doesn't know the jargon would phrase it. Then: what should someone look for when hiring for this?
Ask the one that stings. Who are the best options for [your service] in [your area or niche]? And: what are the alternatives to [business name]?
Run all of it in two different assistants, and once more in one of them in a fresh session. They will not agree, and the disagreement is where the information is.
Three columns. Nothing more elaborate, because elaborate logs don't survive month two.
What it got wrong. Wrong services, an out-of-date location, a merger with a similarly named company, a confident sentence about work you don't do. Note the specific error, not "it was vague" — vague isn't actionable, "it thinks we only do payroll" is.
Who it named instead of you. This is the most useful column and the least comfortable. If it named three competitors, look at what their pages have that yours doesn't. Usually it isn't budget. Usually it's a plain page that answers a question directly, which is Wednesday's post.
Where it says it got that. Ask directly: what sources are you drawing on? Treat the answer as a lead, not a fact — models are unreliable about their own provenance and will sometimes produce a plausible-looking source that doesn't say what they claim. Click through and check. A cited page that doesn't support the claim is still useful information: it tells you which page is being read.
There's a fourth thing worth noting in the margin: whether it named you at all. Not being mentioned is the common result and it is not a verdict on the business. It usually means there is very little describable public text about what you do — which is a fixable problem and, unlike ranking, not a competitive one.
Because the manual version teaches you something the dashboard doesn't, and because at this scale the dashboard is solving a problem you don't have yet.
Monitoring products will run hundreds of prompts and chart it. Useful at volume. But the first time through, you want to read the actual sentences — the tone, the hedging, the specific wrong claim — and a score out of 100 launders exactly that detail away.
There's also a methodological trap worth naming. Generated answers are non-deterministic: ask the same question twice and you can get different names. A single run is close to meaningless, which is why this is a look-and-list exercise now and a ten-question monthly check later, with fixed wording so you're measuring the world rather than your own edits.
And be careful with the "what do you know about me" question specifically. It is very easy to read a flattering hallucination as recognition. If an assistant produces a warm paragraph about your business containing two facts you can't verify, the correct reaction is suspicion, not delight. Retrieved text is not evidence about you — it's a claim that happens to be about you, and it earns the same scepticism as any other unsourced claim. That instinct is the same one behind making a model show its working.
Sit on the list for a day. Then pick exactly one thing.
Most lists produce the same shortlist: one factual error to correct on your own site, one question your pages never actually answer, and one competitor whose page is simply clearer than yours. That's a good week's work, and it's the week this whole theme lays out.
What you should not do is start publishing volume. The reflex is to conclude you need more content. The GEO research points elsewhere — the levers that moved visibility most were statistics, source citations and quotations, which are properties of a page, not a quantity of pages. One page rewritten to be quotable beats five more pages nobody can lift from, and it's Friday's post.
Fifteen minutes. Two assistants. One honest list. Then close the tab.
For the prompt set this check belongs to — free, 5-min read: 7 AI Prompts That Run a One-Person Business.
Related Decodes
AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.
TL;DR: Most founders have never asked an AI assistant what it says about their own business, and it takes fifteen minutes to find out. Ask by name (what do you know about us, what do we do and for whom), by need (who does this kind of work, what should someone look for), and the uncomfortable one (who are the best options, what are the alternatives) — across two assistants plus one fresh session, because answers are non-deterministic and the disagreement is the signal. Log three columns: what it got wrong specifically, who it named instead of you, and where it claims it got that, checking any cited source because models are unreliable about their own provenance. Not being mentioned at all is the common result and usually means there's little describable public text about you, not that you're losing a competition. Treat a flattering unverifiable paragraph as suspicious rather than as recognition. Then stop — don't fix anything for a day, and when you do, rewrite one page rather than publishing five.
