The model remembers you, or it doesn’t
Structure and markup refreshed for current answer engines; the original analysis is preserved.
This month ChatGPT became the fastest-growing consumer app anyone has seen, and Microsoft put ten billion dollars behind it. But it answers only from memory — no live search, no links, no sources. There is no index to climb and no tag to tune. The model either learned your company as an entity from the public record, or it did not — and that changes what preparation even means.
the short answer
ChatGPT answers from memory, not an index: no live crawl, no links, no citations. So there is no ranking to climb and no tag to tune. The model either learned your entity from the public record — present, clear, unambiguously named — or it did not, in which case it hedges, confuses you with a namesake, or invents. The lever is not technical but substantive: be a clear, consistent, well-attested entity in the world. That also feeds search and any future retrieval.
key takeaways
- ChatGPT became the fastest-growing consumer app ever this month, and Microsoft put ten billion dollars behind it — your buyers are already using it.
- But it answers only from memory: no live search, no links, no sources, no way to trace where a claim about you came from.
- So there is no index to climb and no tag to tune. The model either learned your entity from the public record, or it did not.
- What it "knows" of you is a compressed impression from training: sharp if you appeared often, clearly, under an unambiguous name; blurry or invented if not.
- The lever is not technical but substantive — be a clear, consistent, well-attested entity in the world, which also feeds search and any future retrieval.
two ways of being known
The top row is the search machinery you know, with a lever at every step. The bottom row is what this month replaced it with for hundreds of millions of people: a memory of your entity, distilled from training, answering with no source — and nothing technical in the middle for you to adjust.
Why this is unsettling, and why it matters now
The unsettling part is the loss of mechanism. A marketer’s competence in search rests on a chain of things you can measure and move: crawlability, index coverage, rank, click-through. Strip that away and the usual questions have no answer — there is no rank to report, no crawl to fix, no snippet to test. What is left feels uncomfortably out of reach: whether a model trained months ago happened to learn your company well. You cannot audit that the way you audit a site, and you certainly cannot change it by Friday. That discomfort is real, and pretending the old levers still apply would only waste the quarter; the honest first step is to accept that the surface you are used to tuning is simply not present here, and to stop reaching for a dial that the new tool does not have. That acceptance is uncomfortable, but it is also the moment the real work comes into focus.
It matters now rather than later because of the adoption, not the technology. A tool does not have to be finished to change behaviour, and this one crossed into the mainstream faster than any consumer app on record, with millions using it daily by this month. Your buyers are already asking it about your category, and whatever impression it formed of you is already part of how some of them start. Waiting for the product to mature, or for a proper optimisation playbook to appear, means sitting out a period in which the model’s memory of you is being consulted regardless. The sensible response is not to chase a feature set that keeps shifting, but to start improving the one input that will matter under every version of this: how clearly you exist in the record it learns from, which is the one variable that survives every change the product is about to go through.
The shift, in three parts
There is no index to climb; you are remembered as an entity or you are absent; and the only lever is entity clarity rather than a setting. Open each part for where it changes the work.
01 There is no index to climb
Every habit of being found that the last two decades taught us assumes a machine you can work with: a crawler that fetches your pages, an index that stores them, a ranking that orders them, a link that gets clicked — and levers at each stage, from the words in your title tag to the structured data you add. The tool that everyone is suddenly using this month has none of that scaffolding. ChatGPT does not crawl the live web or keep an index of it; it answers from what it absorbed in training, and it returns a paragraph with no link, no ranking, and no citation. The practical consequence is disorienting for anyone trained in search: there is no position to move up, no snippet to rewrite, no setting to change. The question stops being "how do I rank for this" and becomes the blunter "does the model know my company at all" — a question you cannot answer by adjusting anything on your site today, because the thing that would answer it was decided months ago, in training.
02 You are remembered as an entity, or you are absent
What replaces the index is memory, and memory works differently. A model does not keep your homepage on file; during training it learns patterns from an enormous amount of text, and what it ends up holding about a company is a compressed impression — an entity assembled from everywhere that company was mentioned. Two things decide how usable that impression is. The first is presence: a company described often, across many sources, leaves a strong trace, while one that barely appears leaves almost none. The second is clarity: a company described the same way each time, under a name that is not easily confused with a common word or a rival, resolves into a sharp, specific entity, whereas a vague or collision-prone name dissolves into something the model cannot place. So the outcome this month is binary in feel: either the model learned you as a recognisable entity and can speak about you, or it did not, and it will hedge, confuse you with a namesake, or fill the gap with something plausible and untrue.
03 The lever is entity clarity, not a setting
If there is no tag to tune, the work moves to the only thing that actually shapes the memory: how clearly and consistently your company exists in the public record the model learns from. That is a substantive task, not a technical one. It means stating plainly and repeatedly who you are and what you do, in the same terms, on your own pages and anywhere else you appear; making your name and identity unambiguous so a model can tell you apart from things you are not; and earning enough independent mention that the impression is built from many corroborating sources rather than one thin one. None of that is configured in a dashboard, and none of it pays off overnight, which is exactly why it is durable: it is hard to fake and hard for a competitor to copy. It is also the same entity discipline that has always helped a careful reader, a search engine, and now a model alike know precisely who you are — the unglamorous identity work the AC Group has done for clients across 27 years.
The namesake problem, made concrete
It helps to see how the blur happens, because it is rarely dramatic. Suppose your company shares its name with a common word, a town, or another firm in a different field. In search, that ambiguity is manageable — links carry context, and a user scanning results disambiguates in a glance. A memory-based model has no such moment: asked about the name, it blends everything it absorbed under that string into one answer, and your specific identity can be diluted by the louder or more frequent meanings. The result is not an obvious error but a soft one — a description that is partly you and partly something else, delivered with the same confidence as a correct answer. That softness is the characteristic failure of this period, and it traces directly back to how distinctly your entity was established in the text the model read.
The fix is identity work, not keyword work. The more consistently your name is tied to a clear description of what you specifically are — the same way, in many places others trust — the easier it is for any system, model or otherwise, to resolve the name to you rather than to the noise around it. This is the quiet heart of what will later be formalised with structured data and knowledge graphs, but the underlying need predates any of that machinery: be unmistakably one thing in the record. A model can only remember you accurately if the record gave it a clear you to remember.
What to do with this
Stop looking for the meta tag that does not exist, and start strengthening the entity the model learns from. Say plainly and repeatedly, on your own pages, who you are and what you do, in consistent terms; make your name and identity unambiguous wherever you appear, so a model can tell you apart from anything you are not; and earn enough independent, corroborating mention that the impression is built from many clear sources rather than one faint one. Where your presence is thin or your name collides with something else, those are the gaps a memory-based model will fill with guesswork, so close them first, before any of the more visible polish. A gap is not a neutral absence here; it is an open invitation to be described by whatever the model finds nearby. None of this is a setting; it is the patient work of being a distinct, well-described entity in the world.
And keep the moment in proportion. ChatGPT cannot search or cite today, and that will change — likely soon, at this pace — so do not over-fit to the current limitation. Build instead for the constant underneath every version: whether a system remembers you from training or retrieves you live later, it serves you best when your entity is clear, consistent, and well attested — the one input that pays under every version of the tool. That same clarity helps the human reading your page and the search engine indexing it right now, so the effort is never stranded, whichever way the technology happens to turn, next month or next year or the one after. You cannot climb a ranking that is not there, but you can make yourself unmistakable in the record — which is the slow, unglamorous, durable identity work the AC Group has done, patiently, name by name and source by source, for ' + years + ' years.
Two ways of being unknown, and which is worse
Not being learned by the model shows up in two different ways, and they call for the same fix but carry different costs. The first is silence: asked about you, the model simply has little to say, or admits it does not know. That is a missed opportunity, and an honest one — the reader learns nothing, but they are not misled, and they may go look elsewhere. The second is worse: the model, reluctant to leave a blank, produces a confident answer assembled from fragments and plausible guesses. Now the reader is not uninformed but misinformed, and has no way to tell, because the invented description arrives in the same assured voice as a true one. A memory-based system with thin material on you does not usually go quiet; it improvises, and the improvisation is indistinguishable from knowledge, and the reader walks away believing a confident sentence that no one verified.
That is why presence and clarity are not vanity metrics here but the difference between being unrepresented and being misrepresented. The more clearly and frequently your real description sits in the record, the less room there is for the model to fill a gap with something that merely sounds like you. You are not chasing a flattering paragraph; you are crowding out the plausible-but-wrong one by making the true one the easiest thing to assemble from the material at hand. Crowding out the wrong answer is a more honest goal than chasing a flattering one, and it is the goal you can actually act on. The work is the same either way — be present and be clear — but the stakes are higher than they look, because the alternative to being known is not neutral silence so much as confident fiction.
The identity work that predates the model
Here is the part that should steady anyone alarmed by a new acronym. Being a clear, consistent, well-attested entity is not a novel demand invented by language models; it is the oldest discipline in being found at all. Long before any model, a careful reader, a librarian, a journalist, and a search engine all needed the same thing from you: a recognisable identity, described the same way across enough credible places that they could tell who you were and trust it. The years just behind us spent their effort on exactly this under other names — expertise, authority, the signals that a source is genuine and reliable. A memory-based model is, in this one respect, a familiar reader with an enormous memory: it rewards the same clarity and credibility, and punishes the same vagueness, that quality work has always concerned itself with. The vocabulary is new; the underlying test of whether a source is clear and trustworthy is not.
That continuity is reassuring because it means the new thing does not ask you to abandon the old work, only to recognise that it now has one more audience. The entity you make legible to a model is the same entity that earns a human’s trust and a search engine’s confidence; the clarity is fungible across all of them. So the sensible posture in the face of this month’s upheaval is not to invent a separate “AI identity” project but to do the durable identity work properly and let it serve every reader at once — the unglamorous, compounding craft of being unmistakably and credibly one thing, which is what the AC Group has built for clients across ' + years + ' years and what the rest of this archive, looking back, was always about.
Being remembered by a model: quick answers
How is appearing in ChatGPT different from ranking in search?
In search there is a machine you can work with: your pages are crawled and indexed, ranked against queries, shown as a link, and clicked — and you have technical levers at every step, from the title tag to the schema markup to the page speed. ChatGPT, as it stands this month, has none of that. It does not crawl or index the live web; it answers from what it absorbed during training, with no link, no ranking, and no click. So there is no position to climb and no tag to tune. Whether the model can say anything sensible about your company depends entirely on whether your company was present, clearly and often enough, in the text it learned from. It is a different kind of presence: not a page that ranks, but an entity the model did or did not learn.
So can I optimise for ChatGPT the way I optimise for Google?
Not in the familiar sense, and it is worth being blunt about it because the instinct to look for the equivalent of meta tags is strong. There is no field the model reads from your site in the moment, because in the moment it does not read your site at all — it is recalling, not retrieving. You cannot submit a correction, buy a placement, or tweak a snippet to change what it says. The only thing that moves what the model knows is what it was trained on: the breadth, clarity, and consistency of how your company is described across the public record. That is slower and less mechanical than SEO, and it is the actual lever. Optimising here means being a clear, well-attested entity in the world, not configuring a setting in a tool.
What does it mean to be "an entity the model learned"?
A model does not store a copy of your homepage; it learns patterns, and what it ends up "knowing" about a company is a kind of compressed impression assembled from everywhere that company appeared in its training data. If you appeared often, described the same clear way, with a name that is not easily confused with something else, the impression is sharp and the model can speak about you accurately. If you appeared rarely, vaguely, or under a name that collides with other things, the impression is blurry — the model may hedge, mix you up with a namesake, or quietly invent. Being "an entity the model learned" means having a recognisable, consistent presence in the record, which is exactly the entity clarity that lets a system tell who you specifically are.
Is this worth acting on while ChatGPT can’t even browse the web?
Yes, because the work is the same regardless of how the tools evolve, and the adoption this month means your buyers are already here. The model will surely gain the ability to search and cite later, and when it does, a clear public record will feed that too — but even today, a company described consistently and credibly across the web is the one a memory-based model can represent without guessing. You are not betting on the current limitation lasting; you are building the entity clarity that helps whether the model remembers you, retrieves you, or both. That clarity also serves your human readers and your ordinary search presence now, so none of the effort is stranded if the product changes next month, which at this pace it will.
A note on sources and timing
This is written in January 2023, the month Microsoft announced a ten-billion-dollar investment in OpenAI and ChatGPT became, by every account, the fastest-growing consumer application yet, with millions using it daily. We have described only what was true of the tool as of this writing: it answers from training, without live search, links, or citations, so there is no index, ranking, or tag involved in whether it can speak about you. We have not assumed it will gain those abilities on any particular timeline. The durable point does not depend on the product’s current shape: a model can only represent you as well as the public record taught it to, so the lever is the clarity and consistency of your entity in that record — the identity work the AC Group has done for 27 years.