Parametric vs retrieval: where an AI knows you from
The structure and markup were refreshed for current answer engines; the original analysis is preserved.
This month Google renamed Bard to Gemini and shipped its most capable model, and the talk was all about which assistant is smarter. It is the wrong thing to watch. The question that decides how an AI talks about your brand is not how clever the model is — it is which of its two memories the answer came from.
the short answer
What an AI knows about you comes from two places: parametric memory (what it learned in training, frozen at a cutoff date) and retrieval (what it fetches live from external sources). You cannot edit the trained-in half — you did not choose it and cannot change it until the next model. You can shape the retrieved half by being a clear, verifiable, current entity an engine can fetch and cite. That is where the work lives.
key takeaways
- In February 2024 Google renamed Bard to Gemini and shipped Ultra 1.0; the debate was which model is smarter. For your brand, that is not the question that matters.
- What an AI knows about you comes from two places: its trained-in memory (parametric — frozen at a cutoff date) and what it retrieves live from external sources (retrieval, or RAG).
- Parametric memory is static: if your brand lives only there, you depend on when the model was trained and whether it learned you well — and anything new is invisible until the next retraining. When it does not know, it guesses.
- Retrieval brings current information and lets the engine cite verifiable sources. It is the half of the AI’s knowledge that moves — and the half you can shape.
- You do not control what went into training. You do control your retrievability: being a clear, verifiable entity with current sources an engine can fetch and cite. That is the GEO play.
two sources, one answer
Both boxes feed the same sentence about you. The top one is set and sealed; the bottom one you build and keep current. The entire practical question of AI visibility is how much of the answer you can move from the frozen box to the one you control.
Why the model news is a distraction
It is natural, in a month when one company renames its assistant and claims the most capable model yet, to read the race as the story — to assume that a smarter model will simply know your brand better. It will not, not in the way that matters, because intelligence is not the bottleneck. A more capable model reasons better over whatever it has, but it still only has two sources, and a brilliant model working from a frozen, incomplete memory of you will produce a more fluent version of the same outdated answer. The leap in quality you actually want does not come from the model getting smarter; it comes from the model retrieving better sources about you — and that depends on you, not on the lab. A sharper model with nothing fresh to retrieve about you is a sharper way of being wrong; the intelligence amplifies whatever it is given, which cuts both ways depending on what you have made findable.
So the right response to model news is not to wonder whether the new one finally knows you. It is to ask whether, when any of these assistants reaches out to retrieve, your brand is among the clear and current sources it can find. That question has the same answer across Gemini, ChatGPT, and Copilot, because they share the same underlying shape: a frozen interior memory and a live retrieval reach. The brands that win attention in this era are not the ones lucky enough to be well baked into one model’s weights; they are the ones retrievable by all of them, on purpose. The model headlines will keep coming. The work that pays off is indifferent to which logo is on top this week.
How it works, in three parts
The two memories behind every answer, why the frozen one works against a brand that changes, and the one you can actually shape. Open each layer for the part that changes where you put your effort.
01 The two memories behind every answer
When an AI answers a question about your company, the words come from one of two very different places, and the difference is not academic. The first is parametric memory: the knowledge baked into the model’s weights during training, a compressed snapshot of everything it read, fixed at a cutoff date. The second is retrieval: information the model fetches at the moment of answering, from sources outside its weights — the idea introduced by the 2020 paper on retrieval-augmented generation, which paired a parametric model with a non-parametric external memory reached at inference time. The two behave nothing alike. Parametric memory is fast and always available but frozen; it knows what it knew when it was trained and not a day more. Retrieval is current but only as good as what it can find. Every answer about you is some blend of the two, and knowing which one is talking is the first step to influencing it. Most arguments about AI visibility go wrong because they treat the model as a single black box; once you see it as two memories with different rules, the question of what to do stops being mysterious and starts being a matter of which memory you can reach. The frozen one is closed to you and the live one is open, so the strategy writes itself once the two are looked at on their own.
02 Why the frozen half works against you
The parametric half has a quiet problem for any brand that changes: it is a photograph, not a live feed. If you rebranded, repriced, launched, or repositioned after the model’s cutoff, none of that is in the weights — the model will describe a version of you that no longer exists, with complete confidence. Worse, if your brand was thinly represented in the training data to begin with, the model may have learned almost nothing accurate about you, and a model that does not know tends not to say so; it fills the gap with something plausible and wrong. That is the mechanism behind a lot of the confident errors people notice — the famous early Bard demo getting a fact about the James Webb telescope wrong is the same failure mode, scaled. You cannot patch any of this directly. The weights are set, you did not choose their contents, and the next training run is on someone else’s calendar. Relying on the parametric half to carry your brand is relying on a stranger’s memory of you from an unknown date.
03 The half you can actually shape
The retrieved half is the opposite in every way that counts: current, correctable, and yours. When an engine retrieves, it goes looking for clear, credible, up-to-date sources and grounds its answer in what it finds — which is precisely why retrieval reduces the confident-but-wrong problem and lets the engine cite something verifiable. The opening that gives you is the whole game. You cannot rewrite the weights, but you can be the source the retrieval step reaches for: a clean, unambiguous entity a machine can resolve, current pages that answer your category’s real questions, and a presence in the credible third-party places engines trust. Do that and the live layer carries an accurate, current picture of you no matter what the frozen layer happens to hold. This is why entity and source work is not cosmetic — it is the only durable lever on what an AI says about you, and it is the one the AC Group has built for 27 years, long before the layer had this name. The vocabulary is new; the discipline of being a clear, current, verifiable source is not, and that continuity is the reassuring part — the work that earned trust from editors and researchers is the same work that earns it from a retrieval step.
What to do with this, concretely
The framing is only useful if it changes where your effort goes, so here is the concrete version. Stop trying to influence what a model already learned — you cannot, and the attempts waste the budget that the retrievable layer would reward. Instead, make your brand the source a retrieval step prefers. Resolve yourself into one clean entity: a single, well-defined identity that an engine connects every mention back to, rather than a smear it confuses with similarly named things. Keep your most important pages current and factual, written so a clear answer can be lifted from them without guesswork. And earn a presence in the third-party sources engines already trust, so your accuracy does not depend solely on your own domain.
If you do only one thing, make it the entity work, because it is the multiplier the other two depend on. A page is only retrievable for you if the engine can tell it is about you; a third-party mention only helps if the engine connects it back to the same entity as your own site. Resolve the ambiguity first — one name, one definition, one consistent identity across every place you appear — and the freshness and the earned presence start compounding instead of scattering. Skip it, and you can publish current pages all year while the engine files half of them under a company that is not quite you.
None of that is exotic, and that is the point. Retrievability is not a growth hack; it is the unglamorous sum of being recognizable, current, and verifiable — the same hygiene that has always separated sources worth citing from sources worth ignoring. What is new is the stakes. In the parametric era of search, a stale or thin presence cost you a ranking. In the retrieval era, it costs you the answer itself, because the engine will simply reach for whoever did the work instead. The brands that treat the retrieved layer as their real product — and the frozen layer as a bonus they cannot control — are the ones an AI will keep describing correctly, long after this month’s model is yesterday’s news. There is a discipline in that posture worth naming: it means making peace with the part you cannot control and pouring your attention into the part you can, rather than the more common and more wasteful instinct of fretting over the model and neglecting the sources. The brands that get this right are not chasing the news; they are quietly making themselves the easiest correct answer to find, and then letting every new model, on its own schedule, discover what was true all along.
What this looks like in practice
Picture a company that repositioned a year ago — new pricing, a sharper category, a product that does more than it used to. Ask a model trained before that change to describe them, working from parametric memory alone, and you get a confident, fluent, year-out-of-date answer: the old price, the old positioning, sometimes a feature that no longer exists. The model is not broken; it is reporting accurately from a snapshot taken before the change. Nothing the company does to its own site moves that answer, because the answer never touched the site — it came from the weights. This is the trap brands fall into when they assume that being "out there" is enough: out there is not the same as in the frozen memory, and the frozen memory is the only thing a non-retrieving answer can see.
Now ask an engine that retrieves before answering. It reaches out, finds the company’s current pages and a few credible third-party sources, and grounds its answer in those — the new price, the current positioning, the product as it is today, with something it can cite. Same company, same week, opposite answer, and the only difference is whether the engine read the frozen half or the live half. That is the whole argument in one comparison: the brand cannot change what the first engine remembered, but it can make absolutely sure that when the second engine reaches out, the clear and current truth is sitting right there to be found. Everything we call entity and source work is, in the end, making the second answer the one that happens. And the gap between the two answers is not a rounding error — it is the difference between a prospect hearing your current pitch and hearing a competitor’s, because the engine that could not retrieve you accurately will happily retrieve someone else who did the work.
Parametric, retrieval, and your brand: quick answers
What is the difference between parametric and retrieved knowledge?
Parametric knowledge is what a model learned during training and stored in its weights — a fixed snapshot of the world up to a cutoff date. Retrieved knowledge is information the model pulls in at the moment of answering, from external sources outside its weights. The distinction comes from the 2020 paper that introduced retrieval-augmented generation, which paired a parametric language model with a non-parametric external memory accessed by retrieval at inference time. The practical difference is motion: parametric knowledge is frozen until the next retraining, so anything that changed after the cutoff — or was never well represented in the first place — is simply not there; retrieved knowledge can be current, because it is fetched fresh each time. For a brand, that difference decides whether an AI describes you as you are now or as you were whenever the model was last trained.
Can I change what an AI already learned about my brand?
Not directly, and that is the uncomfortable part. The parametric memory is set at training time; you did not choose what went in, you cannot edit it, and you will not know exactly what it contains. If the model learned an outdated fact about you, or never learned much at all, that does not change until the next model is trained — on a schedule you have no say in. What you can change is the other half: your retrievability. When an engine retrieves live, it reaches for clear, current, verifiable sources, and whether yours are among them is entirely within your control. So the honest answer is that you do not fix the parametric memory; you make the retrieved layer so good and so current that it carries the answer regardless of what the weights happen to hold.
How do I make my brand more retrievable?
Be the kind of source a retrieval step can find, trust, and lift. That means being a clean, unambiguous entity — one well-defined thing a machine can resolve every mention back to, rather than something it confuses with a similarly named company. It means keeping current, factual pages on the questions that matter in your category, structured so an engine can extract a clear answer. And it means being present in the credible third-party sources engines reach for, not only on your own site. Retrievability is not a trick; it is the sum of being recognizable as an entity, being current, and being verifiable. The work reads like good information hygiene because that is exactly what it is — and it is the half of the AI’s knowledge you actually own.
Does this mean training data does not matter?
It matters, but you cannot manage it, so it is the wrong thing to obsess over. Being well represented in the data a model trains on is genuinely helpful — it gives the parametric memory a better starting picture of you — but you do not control the training set, its cutoff, or when the next one runs, and a single misremembered fact can sit frozen in the weights for a long time. Spending your effort where you have pull is simply the rational move: the retrieved layer is current, fixable, and yours. Treat strong parametric representation as a welcome bonus you cannot directly engineer, and treat retrievability as the discipline you actually invest in. One you hope for; the other you build.
A note on sources and timing
This is written in February 2024, the week Google renamed Bard to Gemini and launched Gemini Advanced with Ultra 1.0 — the news that prompted it. The parametric-versus-retrieval distinction is not new: it comes from the 2020 paper that introduced retrieval-augmented generation, pairing a model’s trained-in weights with an external memory reached at the moment of answering, and it has been a standard way to describe how these systems know things ever since. We have used the early Bard demo’s mistake about the James Webb telescope only as a well-known illustration of how a confident model fills a gap it cannot retrieve around. We have not reported on capabilities or launches that did not exist as of this writing. The durable point stands on its own: an AI’s knowledge of you has a frozen half you cannot touch and a live half you can build, and the work that pays off goes into the second. It is the work the AC Group has done for 27 years, under one name or another.