A new study says AI searches for the brands it already remembers. We ran the same test on local businesses and got 3.5×

What geoSurge actually measured

The primary source is geoSurge's own report, “AI Searches What It Remembers”, published 2026-07-28. The design is a two-stage funnel. Memory is the model's top-10 brands for a category, recalled before it touches the web. Searched is whether the model then fires a fan-out query naming that brand — measured on Gemini 3.5 Flash. Nine industries (travel, automotive, finance, business software, education, food and restaurants, luxury, fitness and wellness, fashion), 66 US prompts, each answered 60 times across a 12-day window from 2026-05-29 to 06-09: roughly 3,960 responses firing 13,281 fan-out queries, and 1,416 brand-level observations.

geoSurge is explicit about the limit, and we will repeat it because it applies to our replication too: “this is an association in exploratory data, not a proven cause,” with “brand prominence… the main confound.” Memory and search were measured on different models, which the authors argue makes the gap a conservative estimate rather than one model's quirk.

Why that study does not settle the question for a local business

Three gaps, all of which matter if you own a med spa in Charlotte rather than a national fashion label. The subject is national brands — a category where “does the model remember you” is answerable, because the model has heard of Marriott. The stage is the search step, not the answer: a brand can be searched for and still not be recommended. The engine is Gemini 3.5 Flash, and the engines your buyers use are mostly not that. So the honest reading of the study is directional for anyone below national scale.

We cannot measure a model's pre-search memory. What we can measure is the thing a business actually cares about: once an engine has surfaced you, does it surface you again — on the same question, on a separate run, weeks later, asked on behalf of a different business? That is the answer-step analogue, and our corpus is built out of exactly that shape by accident.

The test we could run, and why it is not circular

Every AskedAbout audit runs three prompts against two engines. Two are unbranded discovery prompts — “I'm looking for <vertical> in <metro>. Who do you recommend and why?” and “What are the best <vertical> in <metro>? Give a short ranked list with a one-line reason each.” The third names the business (“Is X a good choice for…”) and is excluded from every number below, because a prompt that contains the name guarantees the name appears in the answer and would inflate every business identically.

Group the audits by (metro, vertical). Inside a group, every audit was asked the identical two unbranded prompts, on different days, on behalf of a different paying-or-prospective business. So for each business we can ask two independent questions:

• Exposed — does its name appear in at least one unbranded answer belonging to a different audit in its group? (A different business asked; the engine volunteered this name anyway.)
• Named — how often does its name appear in its own unbranded answers?

Those are disjoint sets of API calls, made on different days. That is what makes this a test–retest of the engines' shortlist rather than a tautology. The instrument is in our repo at `ops/brand_reappearance.py`, built on the same smoke-row-free cut (`ops/audit_cut.py`) every one of our published numbers uses.

The result: 27.7% versus 8.0%

A ratio of 3.5× — against geoSurge's 3.2×, at a different stage of the pipeline, on different engines, for businesses roughly six orders of magnitude smaller. The cut is the production corpus as of 2026-08-01: 178 completed audits, 1,054 answers, 703 of them unbranded, of which 163 audits across 53 (metro, vertical) groups in 28 metros had at least one comparable sibling. Pearson χ²(1) is 44.8 at the answer level and 34.6 at the audit level — the gap is not sampling noise at this size, though the audits inside one group are not fully independent of each other and we would not quote a p-value off it.

It holds on both engines, and it is much sharper on Perplexity

Perplexity is the harsher of the two in both directions: if it has surfaced you before it surfaces you a third of the time, and if it has not, it names you in 1 answer in 19. ChatGPT is more willing to reach outside its established set — an unexposed business still gets named in 10.7% of its answers — which lines up with our earlier finding that ChatGPT's shortlist churns roughly twice as fast as Perplexity's on the same question. Two independent measurements, same direction.

Every hit was checked by hand, and two were thrown out

Counting a business name inside answer prose is where this analysis can quietly go wrong, so the method is worth stating. Matching is word-boundary, not substring — the first pass of this script scored “All Electric” against “…resolve all electrical problems…” and “Tampa Smile” against “South Tampa Smiles”, and word boundaries kill both. Word boundaries cannot kill a different business whose name contains ours, so all 99 candidate hits were printed with 40 characters of context and adjudicated by hand. Two were rejected: “Austin Dental Company” matched “South Austin Dental Associates” and “Austin Dental Spa (Dr. Mark Sweeney, DDS)” — different practices. Those two rejections are hard-coded in the script with the reason and the date, so re-running it reproduces the published number rather than the raw one.

One more guard worth naming: two audits of the same business in the same metro are not allowed to count as each other's siblings. Before that rule, this analysis read 11.6× instead of 3.5×, because a duplicate audit of one Atlanta med spa was scoring itself as independent corroboration. The conservative number is the one we published.

What this means if you are the business

• The engines' local shortlist is far more stable than “AI answers change every time” suggests. A business already in the set is named in 64% of its audits; a business outside it, in 18%. Whatever moves you across that line is worth more than any single answer you screenshot on a good day.
• Being invisible is self-reinforcing, and that is the actionable half. geoSurge's framing is that category authority is “earned over time through category authority (mentions, coverage, category association), not won at query time.” Our version is blunter: the engines have a working shortlist for “best med spa in Charlotte,” and if you are not on it, nothing you change on your own website this week will put you there. What puts you there is being on the third-party pages the engines actually read — the directories, the “best X in Y” listicles, the review profiles. We mapped which ones per vertical in where AI looks before it recommends a local business.
• Measure the shortlist, not your ranking. There is no position 7 to climb to here. The unit of visibility is are you in the named set, on the questions your buyers actually ask. That is what the free 60-second check returns, and what the $79 audit measures across four engines with three runs per question.

What this does not say

• It is not causation, and neither is geoSurge's. “Already surfaced” and “surfaced again” are two measurements of the same underlying prominence. The defensible claim is that the shortlist is reproducible, not that appearing once causes appearing again.
• It is not a replication in the strict sense. geoSurge measured memory → search on Gemini 3.5 Flash for national brands. We measured surfaced → surfaced-again on ChatGPT and Perplexity for local SMBs. The two ratios landing at 3.2× and 3.5× is a striking coincidence, not a confirmation; they are different quantities.
• The sample is 163 businesses in four verticals — dentists, med spas, personal-injury law, home services — across 28 US metros, audited between 2026-06-12 and 2026-08-01. Group sizes run from 2 to 9 audits.
• Google's AI Overviews, AI Mode, Gemini and Claude are not in this cut. The comparable-group analysis runs on the two engines every audit uses.

The takeaway

Two independent instruments, one week apart, on different engines and at different stages of the pipeline, both found the same shape: AI answers are far more conservative than they look. The engines reach for what they already have. For a national brand that is a memory problem measured before the first search fires. For a local business it is a shortlist problem measured in the answer itself — and it is worse, because there is no ranking to climb, only a set to be in or out of.

The free 60-second check asks the engines one of your buyers' real questions and shows you whether you are in the named set. The $79 audit runs 25 questions across ChatGPT, Perplexity, Gemini and Claude, three times each, and returns the citation source map — the exact third-party pages the engines pulled from, which is the list you have to get onto. Agencies checking a book of clients can run five at once with the $249 Agency 5-pack.