Your Brand Isn't Invisible to AI. It Just Can't Remember You.

Somebody on your team asks a chatbot about your company, mid-meeting, half as a joke. It hedges. It gets a detail wrong. Maybe it names a competitor instead. The room goes quiet, and the conclusion everyone reaches is the same one: the model has never heard of us. New research on AI brand recall suggests that conclusion is usually wrong, and the actual explanation is both weirder and more fixable than 'we don't exist to AI.'

The panic, and what the research actually found

You've seen the screenshots by now. Someone runs their brand name through a chatbot, it doesn't know the founding date, it mixes up the product line, and the post gets captioned something like 'we're invisible to AI.' It's a compelling story, because it matches the fear everyone in marketing already has: that a new gatekeeper showed up and it never got the memo about you. It's also, according to the people who actually opened up these models and measured what's inside them, not quite what's happening.

The paper (titled 'Thinking to Recall: How Reasoning Frees Up Parametric Knowledge in LLMs' in the researchers' own words, produced by teams at Google Research alongside Technion and Tel Aviv University) did something most brand-visibility discourse skips entirely. It didn't just ask models questions and grade the answers. It went in and checked, directly, whether the fact was represented in the model's parametric memory at all: the internal weights formed during training, the model's actual encoded knowledge, as opposed to whatever it says out loud when you ask it something. For frontier models like Gemini-3-Pro and GPT-5, the researchers found the fact was in there 95 to 98 percent of the time. The model had seen it. It had absorbed it. And then, a startling chunk of the time, it still got the direct question wrong.

“The knowledge was almost always present. The failure was in getting the model to surface it on demand, which is a retrieval problem, not a knowledge gap.”

AI brand recall isn't the same as AI knowledge

This is the distinction worth sitting with, because it changes what you're actually solving for. Everyone treats a bad chatbot answer as evidence of an empty file, as if the model has a folder for every company and yours is missing. But encoding and recall are two separate mechanisms, and the research is explicit that they don't move in lockstep. A model can encode a fact during training and still fail to produce it correctly when asked a plain question later, because generating an answer is a different operation than storing one. Storage happened during training, largely out of your control at this point. Retrieval happens at answer time, in response to how the question is framed and how the model chooses to process it, and that part is very much a live variable.

Put another way: your brand almost certainly isn't a stranger to the model. It's more like an acquaintance whose name is on the tip of the model's tongue. The gap between those two situations (total absence versus unreliable retrieval) is the gap between 'start from zero' and 'fix a specific, describable failure mode.' One of those is an enormous, open-ended content problem. The other is closer to an engineering problem, with levers you can actually pull. That reframing alone should lower the temperature on a lot of the current AI citation anxiety, because a missed citation on one prompt doesn't mean the underlying fact isn't in there.

Why turning on reasoning fixes most of the AI brand recall gap

Here's the part of the paper that should reshape how you read a bad chatbot answer. When the researchers let the model reason before answering (the 'thinking' or extended-reasoning mode now standard on frontier systems), the failure rate on those same facts dropped from the 26-34% range down to roughly 11-12%. That's not a marginal improvement. That's the gap closing by more than half, without a single new fact being added to the model. Nothing changed about what the model knew. What changed was whether it was given room to check its own memory before committing to an answer.

This matters practically because not every AI surface your buyers touch is running the same mode. A fast default response in a chat window and a reasoning-enabled pass in a tool like Google's AI Mode, Claude with extended thinking switched on, or a reasoning-model tier of ChatGPT are, functionally, two different retrieval systems answering the same question. If your brand gets misrepresented on the fast surface but shows up correctly the moment reasoning is engaged, that's not evidence your content strategy failed. It's evidence you got unlucky with which retrieval mode happened to run that query. Worth knowing before you rewrite a page in a panic over one bad screenshot.

Search Engine Land's coverage of this line of research has framed it as a shift in how practitioners should think about AI 'understanding' a brand at all: less like a static lookup and more like a search over the model's own memory, one that can succeed or fail depending on effort spent. Search Engine Land and outlets tracking the underlying study both landed on the same practical point: reasoning-enabled answers are simply more reliable, and that reliability gap is now something you can plan around instead of just hoping it lands in your favor.

The direction problem: how you phrase the fact matters

There's a second finding, from a related piece of Google's own research, that explains a specific and very common flavor of AI brand-recall failure: direction. Facts get harder to recall when a question reverses the subject and object from how the model most often encountered that fact during training. If most of the web describes your company as '[Brand] was founded by [Person],' the model recalls that phrasing well. Ask it the mirror image ('what company did [Person] found?') and recall gets noticeably shakier, even though it's logically the identical fact stated the other way around.

Dan Petrovic at DEJAN, a technical research firm focused on generative engines, has been describing this under the term generative self-retrieval: the idea that a model doesn't just store facts, it stores them with a directional bias baked in from whichever phrasing dominated its training data. If your entire public footprint states relationships in one direction only, you've effectively built a one-way door in the model's memory. It can walk through going one direction and gets stuck coming back the other way, and your buyer's actual question might be phrased either way depending on what they're trying to figure out.

What this means for generative engine optimization

Take the two findings together and the strategy shift is fairly specific. A meaningful share of what used to be framed as pure content volume (publish more, so the model eventually learns about you) was already accomplished the moment your facts got encoded in training, which the research suggests is happening at a 95-98% clip for anything a frontier model has plausibly seen. The unclaimed value isn't in teaching the model something new. It's in making what it already knows easy to pull out, on any phrasing, on any surface, whether or not the user happened to trigger reasoning mode that day.

In practice that means restating your core facts in both directions rather than only the direction that reads naturally to a human copywriter. It means being explicit about relationships instead of trusting inference: say plainly who makes what, who acquired whom, who founded what, rather than leaving it implied by context two sentences earlier. It means structured markup that states these relationships in machine-parseable form alongside the prose, since schema is one of the clearest signals AI engines read directly when they're trying to resolve exactly this kind of entity relationship without guessing. None of that is exotic work. It's closer to redundancy engineering than content marketing, and it's the kind of structural groundwork our generative engine optimization practice builds into every account, because a single well-phrased sentence buried on one page is a much weaker bet than the same fact stated three different ways across the properties a model actually pulls from.

This is the same instinct that shows up in how we handled it for Zenity, where the fix wasn't a bigger content calendar so much as making a narrower set of facts unmistakably retrievable, stated plainly and redundantly enough that no phrasing of a buyer's question could dodge them. For companies in regulated or technical categories, the kind we work with across AI and machine-learning industries, the redundancy matters even more, because the facts buyers ask about are more likely to be precise, technical, and phrased in a dozen different ways depending on who's asking.

None of this promises a 100% recall rate. The research itself shows even reasoning-enabled models still miss roughly one in nine encoded facts, and a fast-mode answer will keep getting things wrong at a noticeably higher rate no matter how well-structured your site is. That's not a reason to give up on the fix. It's a reason to stop treating every missed mention as proof of nonexistence and start treating it as a retrieval failure with a known set of causes: shallow mode, unfamiliar phrasing, one-directional sourcing. Each of those has a countermeasure. None of them require convincing a model you exist from scratch.

So the next time your name doesn't come out of a chatbot cleanly, resist the urge to conclude you're starting from zero. You almost certainly aren't. The fact is probably already sitting in the weights, encoded the day the model finished training on whatever the internet had said about you at the time. Your job now isn't to reintroduce yourself. It's to make sure that fact can be found no matter which door the question walks in through.

See where you are cited today

A free snapshot audit of your rankings and AI citations before we ever talk.