Ask the model what it says about you
Structure and markup refreshed for current answer engines; the original analysis is preserved.
GPT-4 arrived this month and suddenly a capable model sits in everyone’s browser tab. The first useful thing to do about how AI talks about you is not a dashboard or a strategy deck. It is to ask the model about your own company and read the answer slowly — what is right, what is stale, what is invented, what is missing. It costs ten minutes, and almost no one has done it.
the short answer
With a capable model now a tab away, the cheapest first move is to ask it about your own company and read the answer in four buckets: accurate, stale (training cut-off), invented (hallucinated), missing. The invented bucket is the dangerous one — a confident falsehood sounds just as authoritative as a true line. This is a diagnosis, not a metric: the output is non-deterministic and frozen at a cut-off, so there is no stable score to trend. It shows you the shape of the problem, and where to start.
key takeaways
- GPT-4 landed this month and a capable model now sits in everyone’s browser tab — so the cheapest first move on AI visibility is to ask the model about your own company and read the answer.
- Sort what it says into four buckets: accurate, stale (training cut-off), invented (hallucinated), and missing. Each points at a different problem and fix.
- The invented bucket is the most dangerous: a confident falsehood sounds exactly as authoritative as a true statement, and a stranger cannot tell them apart.
- This is a diagnosis, not a metric. The output is non-deterministic and frozen at a cut-off, so there is no stable score to trend — it shows you the shape of the problem, not a number.
- Match action to bucket: protect what is accurate; corroborate the current truth where it is stale; remove ambiguity where it is invented; earn presence and clarity where it is missing.
the ten-minute diagnosis
Ask, get one confident answer, then read it four ways. The exercise is fast and revealing — and the most important thing it reveals is that the invented parts wear the same calm authority as the true parts, which is why a stranger cannot tell them apart and why you need to.
Why this month is the moment to look
Two things changed at once this month. A markedly more capable model became widely available, fluent and convincing enough that people are starting to take its answers at face value rather than as a novelty. And the access barrier effectively vanished — the model is in a tab that hundreds of millions of people, including your customers, already have open. Put those together and the model’s account of your company stops being a curiosity and starts being part of how you are perceived, by people who will not think to doubt a fluent, assured paragraph. That is the reason to look now rather than later: not because the technology is finished, but because the perception is already forming, and you are currently not in the room when it does. The cost of waiting is not abstract: every confident, unmanaged answer is a first impression you did not get to shape.
There is a quieter reason too, specific to measurement. People are about to be sold a great deal of machinery for tracking AI visibility, much of it promising tidy scores and dashboards. Before you buy any of it, you owe yourself the unmediated version — your own eyes on the raw answer — because it inoculates you against over-trusting a number later. Once you have seen how much the response shifts between asks, how confidently it invents, and how stale parts of it are, you will read any future dashboard with the right scepticism. The ten-minute look is not just a starting diagnosis; it is the thing that teaches you what the fancier tools can and cannot honestly claim.
The diagnosis, in three parts
Do the ten-minute thing almost no one has done; read the answer in four buckets rather than as one verdict; and treat the whole exercise as a diagnosis, not a measurement. Open each part for where it changes the work.
01 Do the ten-minute thing almost no one has done
This month a genuinely capable model arrived in a tab everyone already has open, and with it a diagnostic that costs nothing and that almost no company has bothered to run: ask the thing about yourself. Type the questions a prospective customer would type — what is your company, what does it do, who is it for, is it any good, who competes with it — and read what comes back. It feels almost too simple to count as work, which is precisely why it gets skipped in favour of grander plans. But there is no substitute for seeing, in plain words, what the model asserts about you to a stranger who asks, because that assertion is now part of your first impression whether you manage it or not. The instinct to wait for a proper tool is understandable and, here, wrong: the tool does not exist yet, the model does, and ten minutes of looking will teach you more about your starting position than a quarter of speculation.
02 Read the answer in four buckets, not as one verdict
The mistake is to read the response as a single grade — good, bad, flattering, unfair — and react to the whole. The useful move is to sort it. What is accurate tells you what the model has correctly absorbed, and is worth knowing so you can protect it. What is stale was true once and has aged out, a symptom of the training cut-off rather than of ill intent; it tends to cluster around anything that changed recently — a rebrand, a new product, revised pricing. What is invented is the confident falsehood, and it is the bucket that should worry you most, because a hallucinated claim is delivered in exactly the same assured tone as a true one, and the reader has no way to tell which is which. What is missing is the negative space: the things a fair summary would include that the model simply does not seem to hold. Four buckets, four different diagnoses — collapsing them into one number throws away the part that tells you what to fix.
03 Treat it as a diagnosis, not a measurement
It is tempting, especially this month, to turn the exercise into a scorecard — to count the right answers and call it a visibility metric. Resist that, because the output will not hold still. The model is non-deterministic: ask the same question an hour later and the wording, the emphasis, even the facts can shift, so there is no stable figure to write down and trend. It is also frozen at a training cut-off and prone to confabulation, which means any single reading is a snapshot of one roll of the dice, not a measurement. That does not make the exercise worthless; it makes it a different kind of valuable. A diagnosis shows you the shape of the problem — does the model know you at all, is what it knows current, where is it confidently wrong — and the shape is what you act on. Real measurement, the kind you can trend and trust, is a separate and harder craft, and pretending this casual reading is that craft would only mislead you. Honest about its limits, it is still the best first hour you can spend, the cheapest hour with the highest information return, and exactly the kind of plain, truthful groundwork the AC Group has built on for 27 years.
The bucket that should worry you most
Of the four, the invented bucket deserves a closer look, because it is the one people underestimate. When the model is stale or silent about you, the harm is mostly missed opportunity — a customer learns less than they could. When the model invents, the harm is active: it tells a customer something false about you in a tone indistinguishable from the truth. A made-up feature, an imagined limitation, a competitor wrongly described as a partner, a price that was never yours — each is delivered with the same calm assurance as the accurate lines around it, and the reader has no signal that one sentence is solid and the next is fiction. That is what makes confident fabrication more dangerous than an honest gap: a gap leaves the reader uninformed, but an invention leaves them misinformed and certain.
The practical response is not to argue with the model, which you cannot do, but to ask why the invention was plausible enough to generate. Often the answer is that your own public information left room for the guess: a claim stated vaguely, a detail buried, a fact that no independent source confirms, so the model filled the space with something that sounded right. Seen that way, an invented answer is also a diagnosis of where your own account of yourself is thin or ambiguous. You close the gap not by correcting the model but by making the real thing clear, specific, and corroborated enough that the plausible guess and the truth are the same sentence.
What to do with this
Spend the hour. Ask the model the handful of questions a customer would, in a few phrasings, and write down the answers verbatim before they change. Sort each claim into accurate, stale, invented, or missing, and resist the urge to total it into a score — the value is in the sort, not the sum. Then match the action to the bucket: protect what is accurate, make the current truth clear and well-corroborated where the answer is stale, remove the ambiguity that invited the invention where it is wrong, and earn plainer, more widely confirmed presence where you are missing. None of these are toggles; they are the ordinary work of being a clear and well-attested source, now aimed by an actual diagnosis instead of a guess.
And keep it in proportion. This is the first hour, not the last word — a qualitative look that tells you where to begin, not a measurement you can trend or a strategy you can run on autopilot. The model will keep changing, your own information will keep needing care, and proper measurement remains a separate, harder discipline for when you are ready. But as a place to start, on the month a capable model became something everyone can ask, there is nothing cheaper or more honest than looking with your own eyes — which is the plain, unglamorous habit the AC Group has trusted for ' + years + ' years: see the thing clearly before you try to manage it.
The questions worth asking
Ask the questions a real prospect would, not the ones you wish they would. Start with the plain identity question — "what is [your company]" — because that reveals whether the model knows you at all and how it frames you in one breath. Follow with the job-to-be-done question — "what does [your company] do and who is it for" — which exposes whether the model has your actual positioning or a vague approximation. Then ask the evaluative question a buyer always asks privately — "is [your company] any good" or "what are the downsides of [your company]" — because that is where flattering summaries and quiet hallucinations both tend to surface. Finish with the comparison question — "who are the alternatives to [your company]" — which tells you the competitive set the model places you in, and sometimes reveals that it does not place you in one at all.
Ask each in two or three phrasings, because the model’s answer is sensitive to wording and you want the pattern, not a single take. Where the answers agree across phrasings, you are seeing something the model holds firmly; where they wobble, you are seeing the edge of what it actually knows. Keep the questions in a customer’s voice rather than an insider’s — a real prospect does not type your internal product names or your tagline, so neither should your test. The aim is to stand where your customer stands and read what they would read, which is both the simplest framing and the one most likely to surface the gaps that matter commercially.
Why you can’t just correct it
The instinct, once you spot a stale or invented claim, is to fix it at the source — to tell the model it is wrong and have it remember. You cannot, and understanding why saves a lot of wasted effort. What the model says about you is not stored in an editable record you can log into; it is the product of patterns learned across an enormous body of text, frozen at a training cut-off. Correcting it in a chat does not persist, and there is no form to submit. The honest mechanism of change is indirect and slower: you make the true, current version of your story clear and abundant across the public web, so that the next time the model is trained — or, where it retrieves live, the next time it looks — there is a better, more corroborated answer for it to absorb or find.
That is frustrating if you wanted a button, but it is also clarifying, because it points the work at something you actually control: your own presence. You cannot edit the model, but you can make sure the real answer is stated plainly on your own pages, repeated consistently wherever you appear, and backed by independent sources the model is likely to weigh. A stale or invented claim is, in that light, less an attack to rebut than a signal that the public record it learned from was thin, ambiguous, or out of date — and the durable response is to improve that record rather than to argue with its reflection.
Asking the model about yourself: quick answers
What is the very first thing to do about AI visibility?
Open the model and ask it about your own company — plainly, the way a customer might: "What is [your company]?", "What does [your company] do and who is it for?", "Is [your company] any good?", "Who are [your company]’s competitors?" Then read the answers slowly. You are not measuring anything yet; you are looking. The point of this month, with a capable model suddenly sitting in everyone’s browser tab, is that this diagnostic costs nothing and takes ten minutes, and almost no company has actually done it. Before any tooling, before any strategy, you want to see with your own eyes what the model asserts about you when a stranger asks — because that is roughly what a stranger will now see.
How should I read the answer it gives?
Sort what it says into four buckets rather than reacting to the whole as right or wrong. First, what is accurate — the parts that match reality, which tell you what the model has absorbed correctly. Second, what is stale — claims that were true once but are out of date, a symptom of the model’s training cut-off rather than of malice. Third, what is invented — confident statements that are simply false, the hallucinations, which are the most dangerous because they sound exactly as authoritative as the true parts. Fourth, what is missing — the things you would expect a good summary to include that the model does not seem to know at all. Each bucket points at a different problem and a different fix, which is why sorting beats a single overall verdict.
Is this a way to measure my AI visibility?
No, and it matters to be honest about that. This is a diagnosis, not a metric. The model’s output is non-deterministic — ask the same question twice and you may get different answers — so there is no stable number to record, no score, no rank. It is also frozen at a training cut-off and prone to inventing, so a single reading is a snapshot of one moment, not a measurement you can trend. What it gives you is qualitative and genuinely valuable in a different way: it shows you the shape of the problem — whether the model knows you at all, whether what it knows is current, where it is confidently wrong. Treat it as the doctor’s first look, not the lab result. Proper measurement, when you get to it, is a separate and harder discipline.
What do I actually do with what I find?
Match the action to the bucket. Where the model is accurate, note it and protect it — that is signal you do not want to lose. Where it is stale, the fix is rarely instant, because you cannot edit the model’s memory directly; what you can do is make the current truth abundantly clear and well-corroborated across the web, so the next training round and any live retrieval have something better to find. Where it is invented, look at whether your own information is ambiguous or thin enough to invite the guess, and make the real answer easy to find and hard to misread. Where it is missing, the issue is usually presence and clarity — you may simply not be stated plainly or corroborated widely enough to register. None of these are quick toggles; they are the ordinary work of being a clear, current, well-attested source, now with a concrete diagnosis pointing at where to start.
A note on sources and timing
This is written in March 2023, the month GPT-4 arrived and a capable model became something nearly anyone could open in a browser tab and question. We have described only what was usable as of this writing — asking a general-purpose model about your own company and reading the result — and we have been deliberate that the model is non-deterministic, frozen at a training cut-off, and prone to confident invention, so this is a diagnosis rather than a metric. We have not promised that the exercise measures anything trendable. The durable point does not depend on any one model: when a system speaks about you to strangers, the first sane response is to listen to what it says and read it critically, which is the plain, truthful groundwork the AC Group has worked in for 27 years.