The 2027 budget line for AI visibility is being spent at the wrong end of the funnel. Most of it optimizes the moment an assistant runs a search. By that moment the assistant has already decided which brands are worth searching for, and the decision was made months earlier, inside a training run.
Research published this week by AI visibility firm GeoSurge, in a guest piece for Marketing Dive, measured the size of the gap. Brands that a model already remembered were searched about three times as often as brands it did not, 55.7% versus 17.4%, and for the five most-remembered brands the rate reached 67%. The pattern held across all nine industries tested.
The method matters for anyone tempted to dismiss the finding as anecdote. Over 12 days the firm ran 66 realistic buyer questions 60 times each, capturing roughly 4,000 responses and 13,281 searches that the model executed 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 five most-remembered brands.
Google calls the behavior query fan-out: the model decomposes a question into a batch of parallel searches, some generic and some branded, and every citation or recommendation flows through those queries. A source that never matches one of them never enters the answer. And a brand the model cannot recall cannot be searched for, in the same way a person cannot look up a restaurant they have never heard of.
The effect is not uniform, and that is the useful part. In automotive, 82% of brand-specific searches named a top-five remembered brand; in finance, 77%. In fitness and wellness the figure was closer to 50%, because models searched beyond their memory far more often. The author's tentative read is that memory-led categories are dominated by a few long-established, heavily covered brands, while search-led categories contain more smaller and newer entrants. The strategic implication is diagnostic: in a search-led category, retrievable and well-structured content can move results within months, while in a memory-led category the same work has a much lower ceiling.
Three budget consequences follow. Diagnose before spending, because the two category types require different investments and one of them does not respond to technical work. Fund earned media as infrastructure, since model memory forms from the training corpus through years of press coverage, analyst reports, reviews, forum discussion and consistent association between a brand and its category, and none of that can be injected at query time. And measure recall rather than only citations, because the scoreboard is downstream of the game. Tracking that cost hundreds of dollars a month now has free options, which frees budget for shaping the answer instead of counting it.
The same week produced a second signal from the other side of the same shift. Reddit began testing ads inside its own search results on Nov. 3, for a subset of eligible advertisers, with placement determined partly by contextual relevance to the query. The company's examples are intent-rich: a search for a home espresso setup might surface ads for machines, beans and mugs, while a home office setup search might surface monitors, desks and chairs.
The financial context explains why. Reddit reported third-quarter advertising revenue of $847 million, up 54%, with daily active uniques flat at 130.3 million. In its Oct. 29 shareholder letter the company said referral traffic from search engines fell sharply during the quarter and accelerated late in the period, with October continuing the decline at more than a million visits a week. Chief executive Steve Huffman pointed to Google's AI Overviews and AI Mode as the leading hypothesis, according to Search Engine Journal, and also raised the possibility that machine-translated Reddit content now ranks lower.
Reddit's response is to stop depending on the referral and start monetizing the intent it still owns. The company is shifting toward a login-centered service and will test that model in select markets outside the United States in the fourth quarter. Old Reddit now requires login, and new applications for the public Data API closed Oct. 31.
Two events, one structure. The layer that used to send buyers to a page is being replaced by a layer that answers them, and the platforms that still hold high-intent search are selling access to that intent directly. For B2B marketers this changes what is actually being bought. Ranking on a results page buys a click. Being cited or recalled buys consideration before a click exists, and consideration is what a specification committee remembers when it writes a shortlist.
For industrial and B2B categories, which behave much like memory-led markets, the work is slower and less technical than the playbook suggests. It means sustained third-party validation through trade press, analyst notes, customer case studies and participation in standards work, plus a disciplined association between the company name and its category in every channel that feeds the corpus. Query-time tactics still matter, especially on generic category prompts where no brand is named, and the falling cost of measurement makes them easier to justify. But no amount of query-time optimization manufactures recall.
For challenger brands the clock is the training cycle. Remembered brands get searched, searched brands get cited, and citations feed the corpus that the next generation of models learns from. Each cycle that consolidates incumbents makes the loop harder to break. The budget question for 2027 is therefore not how much to spend on AI visibility, but how much of it is being spent on memory rather than on the last millisecond.
