Why AI Still Thinks You're Last Year's Company

Ask a model about your company and you’re really talking to two informants at once. One of them read the internet a while ago, formed impressions, and had those impressions frozen in place. The other is checking the web right now, mid-answer, skimming whatever it can reach. The reply you get is a negotiation between the two. And when they disagree, the older voice usually wins the tone even when the newer one wins the facts.

This is the part of AI visibility almost nobody manages, because it never shows up on a dashboard. You can watch your live citations climb. You can publish something sharp on Tuesday and see it quoted by Thursday. And you can still lose, quietly, because the model’s frozen memory—the version of you it absorbed during training—is describing a company you stopped being months ago.

Every model carries a snapshot. During training it read an enormous slice of the web and compressed what it saw into a set of associations: this company does that, sells to those people, competes with them, is known for this one thing. That snapshot doesn’t update when you rebrand. It doesn’t notice when you sunset a product, move upmarket, or replace your founder. It just sits there, quietly confident, and colors the first draft of every answer about you.

I keep meeting companies who’ve lived through this. A team pivots from a scrappy freelancer tool to a serious platform for finance departments, spends a year rebuilding the product and the message, and then watches a model recommend them to freelancers—warmly, fluently, and completely wrong. Another company discontinues a product line, and the model keeps suggesting it, because in the frozen memory that product is still the thing they’re famous for. The facts on their website say one thing. The model’s gut says another. The gut goes first.

The uncomfortable truth is that in the frozen layer you are not who you say you are. You are the average of what the internet said about you back when the snapshot was taken. If that average was thin, outdated, or written by other people, that’s the you the model reaches for.

The live layer is the one everyone has learned to chase. It’s retrieval: the model runs a search mid-answer, pulls a few pages, and grounds its response in what it finds. This is where fresh content earns its keep, where a well-placed comparison page or a strong recent article can show up in a citation within days. It’s fast, it’s measurable, and it rewards effort quickly. Naturally, it gets all the attention.

But the live layer arrives late to the conversation. By the time retrieval kicks in, the frozen memory has already set the frame—the assumptions about what you are and who you’re for. Retrieval mostly fills in details inside that frame. It’s very good at correcting a specific fact. It’s much worse at overturning a general impression. So a brand can be current in every retrievable source and still get described through the lens of a version of itself that no longer exists.

Managing one layer and ignoring the other is the most common mistake I see. Teams pour energy into publishing, then wonder why the model keeps mischaracterizing them at the level of identity. They’re patching the live memory while the frozen one keeps narrating.

Here’s the part that sounds like bad news and isn’t. You cannot reach into a model and edit its frozen memory. That snapshot is fixed until the next training run, and you don’t control the schedule. What you can do is change what the next snapshot will absorb—and shore up the live layer hard enough to overrule the old impression in the meantime.

The frozen memory is built from consensus: the same description of you, repeated across enough independent places that it reads as fact. So the long game is consistency. Pick the sentence you want a model to believe about your company, and make sure the same claim appears, in roughly the same shape, everywhere you’re discussed—your own pages, your profiles, the write-ups and directories and conversations you don’t control. Not clever variations. The same claim. Models learn what the web agrees on, and the web only agrees when it repeats itself.

The short game is retrieval pressure. If the frozen memory is stale, you want fresh, unambiguous, quotable material sitting wherever the model looks when it goes to check its work—stating plainly what you are now, who you serve now, what you no longer do. You’re not trying to be subtle. You’re trying to give the live layer enough clean evidence to talk the frozen layer out of its assumptions.

And you have to watch both. Ask the models about yourself on a regular schedule, and read the answer for its frame, not just its facts. A reply can get every detail right and still be built on last year’s premise. When you notice the premise slipping, that’s the signal your consensus needs reinforcing—long before it shows up as a lost customer who “read somewhere” that you do a thing you dropped a year ago.

The reason this matters more each quarter is that fewer people are visiting your carefully updated homepage first. They’re asking a model, and the model answers from memory before it ever checks your site. You spent the year becoming a new company. The internet, and the snapshot it fed, is still describing the old one.

So there are two of you now. There’s the version that’s live—the one you update, ship, and refine. And there’s the version that’s remembered—the one the model reaches for first, assembled from everything that was ever said about you and never corrected. You’ve been managing the first one. The second one is the one doing the talking.