The following is a guest piece written by Francisco Vigo, co-founder and CEO of AI visibility company GeoSurge. Opinions are the author’s own.
Most of what marketers will spend on artificial intelligence visibility in 2027 is aimed at the wrong end of the process. The standard playbook tunes content for the moment that an assistant searches the web, assuming live search gives every brand a fair shot at being mentioned in AI results. By the time the model searches, though, it has already decided which brands are worth looking up, and it makes that call from memory.
We measured this at GeoSurge. Brands an AI model already remembered were searched about three times as often as brands it didn't, 55.7% versus 17.4%, and for its five most-remembered brands, the rate reached 67%. The pattern held in all nine industries we tested. If memory decides who gets searched, the budget belongs in what builds memory: the coverage, reviews and category association that models learn from between training runs.
Search didn't fix the memory problem
When AI assistants first started answering buying questions, brands worried that the model's knowledge was frozen at training time. Live web search seemed to solve that, because whatever the model doesn't know, it can look up in real time. We call that RAG - Retrieval-Augmented Generation. However, before a model can retrieve anything, it has to decide what to search for, and it makes that decision using its own memory, somewhat like a human would.
Modern AI search doesn't run a single query. Ask "what's the best business checking account for a small company" and the model splits it into a batch of parallel searches. Some are generic and some name specific companies. Google calls this “query fan-out.” Every citation and recommendation flows through those queries, and a source that never matches one rarely enters the answer.
To see how memory shapes those queries, we first recorded the brands a model could recall with no web access, across nine industries. Then, over 12 days, we ran 66 realistic buyer questions 60 times each, capturing about 4,000 responses and 13,281 of the searches the model ran behind the scenes. About 69% of those searches were generic category queries. Of the 31% that named a brand, nearly two-thirds named one of the model's top-five remembered brands.
This makes sense mechanically. A model writing search queries draws on the same learned associations it uses to write anything else. It can't search for a brand it never learned, any more than you can Google a restaurant you've never heard of.
Know which game your category is playing
The effect isn't uniform, and that matters for marketers who are planning.
For example, in automotive, 82% of brand-specific searches named one of the model's top-five remembered brands; in finance, 77% did. In fitness and wellness, the figure was closer to 50%, because models searched beyond their memory far more often.
Our tentative read, which the study didn't test, is that memory-led categories are dominated by a few long-established, heavily covered brands. Search-led categories, by contrast, have many smaller, newer ones. Each industry rests on six to 12 prompts, so treat the splits as indicative. Still, the distinction is useful. It tells you whether your problem is retrievability, which content and technical work can fix relatively quickly, or memorability, which it can't.
Three budget implications
First, diagnose before you spend. In a search-led category, well-structured and highly retrievable content can move results relatively quickly. In a memory-led category, the same work has a much lower ceiling.
Second, fund earned media as infrastructure. Model memory forms from the training corpus: years of press coverage, analyst reports, reviews, forum discussions and consistent association between a brand and its category. These disciplines were long justified in the soft language of “reputation.” Now they have a mechanical function, because they are how a brand gets written into the systems that decide what to search for. They can't be injected at query time; they accrue between training runs, or not at all.
Third, measure recall, not just citations. Counting citations measures the scoreboard, but the game is decided upstream, in what models remember before a single query fires. That scoreboard is also getting cheaper. Tracking that typically costs hundreds of dollars a month now has free options. If seeing the answer costs next to nothing, more of the budget can go toward shaping it.
The window is the training run
None of this makes query-time optimization worthless. It works in search-led categories, and even in memory-led ones the generic queries leave room to compete.
But the industry has the priority order backward. The dominant playbook optimizes the last millisecond of a process that is substantially decided months earlier, inside a training run.
For challenger brands, this is urgent. Remembered brands get searched, searched brands get cited and those citations feed the corpus the next generation of models learns from. Each training cycle dominated by incumbents makes that loop harder to break.
You can't win at query time what you lost at training time; however, training time comes around again, and what the models remember next is being written now.
