There's a deeply uncomfortable idea doing the rounds, and I've been sitting with it for a few days now trying to find the angle where it doesn't sting. I haven't found one yet.

Here it is: when you ask an AI assistant a question about your industry — "which CRM should a mid-market B2B company use," "what's the best project management tool for remote teams," whatever — you'd assume the model does the decent, democratic thing. Searches the web. Reads what's out there. Reports back like a diligent intern with good Wi-Fi.

According to a new report from geoSurge, that's... not quite the sequence. What actually happens is closer to your diligent intern already having a favourite before they open the browser. They'll still search. But the search itself is shaped — constrained, steered, arguably rigged — by which brands the model already carries in its memory from training.

One example from the report: when asked which buy-now-pay-later provider a business should offer, the model fired six searches. Four of them named a specific provider — Affirm, Klarna, Afterpay, PayPal — each slotted into the same "merchant fees US" query template. Those four names weren't discovered during the research. They were recalled from training data and then searched in the order the model remembered them.

The consideration set was pre-loaded. The search was a formality.

(I mean, we've all had that colleague who "does the research" but clearly walked in with a decision already made. Turns out language models have the same energy.)

To understand why this matters, you need to see the mechanic the report describes. When a model answers a question, it doesn't fire one search. It fans the question out into a spread of related queries — we've covered fan-out mechanics before — then reads the results and cites a subset. The report breaks this into two stages that most of the industry has been treating as one.

Stage one: memory. The brands the model already associates with a category, ranked internally before it touches the web.

Stage two: search. Whether the model actually fires a query naming a given brand.

Memory sits upstream of search. That positioning is the entire story.

The numbers are blunt. Across nine industries, sixty-six US buyer questions, and nearly four thousand responses, a brand the model remembered was searched at 3.2 times the rate of one it didn't. 55.7 percent versus 17.4 percent. Among brands sitting in the model's top five recalled names, 67 percent appeared in at least one search query. For brands it didn't remember at all? Seventeen percent.

Here's the part that really got me: roughly seven in ten of the model's queries were generic category searches with no brand named at all. But in the three in ten cases where it did reach for a specific brand, it overwhelmingly reached for one it already knew. Sixty-three percent of brand-specific queries named a top-five recalled brand.

The fan-out — the thing we've been treating as this neutral, web-scanning mechanism — is biased before it begins. By what the model already believes belongs in the category.

(If this feels like the AI equivalent of only going to restaurants you've already been to, that's because it basically is.)

I can hear the objections forming, and so could the researchers — who, to their credit, are unusually candid about the limitations. This is exploratory data, not demonstrated causation. Brand prominence is an obvious confound: famous brands are both more likely to be recalled and more likely to be searched, so part of the gap might just be fame doing what fame always does. The study window was twelve days. The prompts were US-only. A few industries rested on as few as six questions.

But two things make the direction hard to dismiss, even if the exact magnitude is soft. First, the gap held across every single one of the nine industries — not-remembered search rates ranged from 9 to 23 percent, remembered from 41 to 82 percent, with zero exceptions. Second, memory and search were measured on two different models. A brand had to clear both independently, which means you can't wave this away as one system's quirk.

Whatever the true size of the effect, the direction is consistent: remembered brands get searched more. If you're allocating budget against AI visibility, the direction is the part you can act on.

Here's where this gets genuinely useful, because the report reframes AI visibility in a way that explains why a lot of the current advice feels like running on a treadmill.

There are two clocks. Most of the industry is watching the wrong one.

The fast clock is everything that operates at query time. Live content, documentation, freshness signals, on-page structure, schema markup. It works — the report is careful to show that. In one example, a model searching for payment providers reached past its recalled set to pick up Lemon Squeezy by name — a brand it never remembered at all. Strong live presence pulled it into the fan-out before recall had formed. You're not locked out if you're not yet in memory.

The slow clock is memory itself. And memory is formed at training time, not at query time. That's the crucial distinction. A brand carried in the model's recall doesn't have to be re-discovered every time someone asks the question. It arrives already in the consideration set. Live content has to earn its place again with every query. Memory, once formed, is simply there.

Nearly all of the AEO and GEO tactics being marketed right now — prompt optimisation, on-page tweaks, chasing this week's citation — are fast-clock work. None of that is wasted. In some categories it's the only route available. But it's the route you take when you haven't yet earned memory. The position that compounds — the one that stops you sprinting for the same shortlist every single time — is slow-clock. And by its nature, you cannot buy it at query time.

(Which is annoying, because "buy it at query time" is essentially the entire pitch of the AEO services market right now. Awkward.)

If memory is earned over training cycles, the obvious question is what earns it. The report's answer is the least surprising and most demanding thing it could be: sustained category authority.

Not a campaign. An accumulation.

The mentions. The analyst coverage. The press. The partnership signals. And above all, the consistent association between a brand and its category across the comparison content, the listicle roundups, and the editorial coverage that models learn from. It's the slow, unsexy work of becoming so obviously associated with a category that a model simply can't learn the category without learning your name alongside it.

This should reframe several threads we've been following separately. The PR renaissance, where earned media turned out to drive the overwhelming share of AI citations. Original research and its outsized citation density. Founder-led expertise as an authority signal. Read individually, each looked like a tactic for getting cited today. Read against this report, they look like something else entirely.

They were all memory-building. All along.

The citation you earn this quarter is useful. The category association it leaves behind — folded into the next training run — is the asset.

This isn't uniform, and the variation is worth understanding because it changes what you should do about it.

In some verticals, brand-specific searching is tightly bound to memory. In Automotive, 82 percent of brand-led queries named a top-five recalled brand. In Finance, 77 percent. In those memory-bound categories, getting into the recalled set is close to a precondition for being searched at all. If you're outside the top-five and you're in one of these categories, building memory is the primary job and live content is the stopgap.

In memory-loose categories — Fitness and Wellness sat at 50 percent — the model searched past its memory far more frequently. Plausibly because its category recall is thinner on more specialised consumer topics, so it leans harder on live results. If you're here, strong live content can still carry you into the fan-out while recall is forming.

But recall remains the sturdier position in both cases. Because it doesn't depend on being rediscovered every single time someone asks.

One last nuance, because it's easy to overclaim and I'd rather we didn't.

Being remembered decides whether the model reaches for you at all. It does not decide whether the model then describes you correctly. Those are two separate gates. A brand can clear the first — enter the fan-out on the strength of recall — and still lose at the second if the content the model reads about it is contradictory, outdated, or just messy enough that it hedges, conflates, or quietly drops the name.

Getting into memory is the slow game. Being coherent enough to survive retrieval once you're in it is the more immediate one. You need both. The uncomfortable part of this report is that the first gate is the one you can't rush.

There's a longer shadow here worth naming. If memory is formed from the coverage and comparison content models train on, and a rising share of that content is itself AI-generated, then the brands already sitting in memory have a structural advantage in staying there. The next model learns the current consensus — including the part of the consensus that's machines repeating machines.

Which raises the question the report stops just short of asking: as more of the training corpus becomes a reflection of the last model's answers, does memory become easier to enter — or does it just harden around the names already in it?

On present evidence there's no clean answer. But it's the right question for anyone deciding, right now, how much to invest in being remembered.

And whether "later" is already too late.

Jo Lambadjieva is an entrepreneur and AI expert in the e-commerce industry. She is the founder and CEO of Amazing Wave, an agency specializing in AI-driven solutions for e-commerce businesses. With over 13 years of experience in digital marketing, agency work, and e-commerce, Joanna has established herself as a thought leader in integrating AI technologies for business growth.

😍It was great, all I needed

🙂Good, not great

🙄It sucked