TL;DR
- AI search has a third layer beneath crawler access and citation selection: what the model already knows when it answers without looking anything up. No robots.txt setting and no on-page tactic reaches it.
- We measured that layer on a Quebec trade ourselves: twelve French prompts through a panel of nine models, July 2026. Closed-book, the panel named a real local business in 7.1 percent of responses, and named no concrete business of any kind in 69 percent. It is a probe rather than a full study, and we hold it at medium confidence.
- The models are not silent. They substitute: national manufacturer brands stand in for a certification signal, and lead-generation marketplaces stand in for a contractor.
- This reads mostly as a notoriety penalty rather than a Quebec one. Being well known in the training data beats being nearby, and strong Google visibility does not carry over.
- The only lever is earned third-party coverage accumulated over years. It is slow, it cannot be bought as a guarantee, and anyone promising to install you in a model’s memory is describing something no vendor has ever offered.
Ask ChatGPT to recommend a roofer in Trois-Rivières without letting it search the web, and watch what comes back. Most of the time it is not a local company. That gap is the third layer of AI search, and it is the one no amount of technical work on your own site will close.
We have walked the first two layers already. Eligibility is whether the engine’s crawler can fetch you, which every vendor documents. Citation selection is which fetched page gets quoted, which no vendor documents at all. Both of those assume the model went and looked. The third layer is parametric memory: what the model retained from training, before any of that. It is what you are left with when the model never looks.
What is the memory layer, and why does no tactic reach it?
When a model answers from memory, your visibility equals your presence in its training data at the cutoff date. Nothing else is in play. The model is not fetching your site, so it does not matter whether OAI-SearchBot is allowed through your firewall. It is not comparing eligible pages either, so the statistics and attributed quotes you added last quarter do not change anything. It is reproducing what it absorbed about your market before it was ever deployed.
Closed-book answering is not an edge case. It is what happens in plenty of ordinary situations: a quick question with no browsing enabled, a model tier that does not retrieve, an app calling an API without a search tool attached. Every one of those is a moment where the entire GEO toolkit is switched off and only training-data presence decides who gets named.
So the work splits in two. Crawler eligibility and on-page work govern the browsing case, which is real and worth doing well. The memory case is governed by how much the wider web wrote about you before the cutoff, and no configuration change touches it.
What we found when we measured it on a Quebec trade
Closed-book, only 7.1 percent of the model responses we tested about Quebec roofing named a local roofer we could verify, and 69 percent named no concrete business of any kind. Rather than cite someone else’s numbers, we ran the measurement ourselves. In July 2026 we put twelve prompts through a panel of nine models, closed-book, asking for roofing contractors in Quebec markets. The prompts were French, phrased the way a Quebec customer would ask. This was a probe rather than a full study, and we hold the results at medium confidence, but the direction is not subtle.
One note before the numbers: Perplexity’s models do not fully honour a closed-book instruction, since retrieval is native to how they answer. That is why the two naming rates below are reported separately instead of blended into one figure.
| JPL closed-book probe, July 2026 (12 prompts x 9 models, medium confidence) | Result |
|---|---|
| Responses naming a local roofing business we could verify automatically | 7.1 percent (5 of 70 closed-book cells) |
| Same, counting Perplexity’s live retrieval as closed-book | 18.9 percent (17 of 90 cells) |
| Responses naming no concrete business of any kind | 69 percent (48 of 70 cells) |
| Responses naming some concrete commercial entity (any kind) | 31.4 percent (22 of 70 cells) |
Read the first and third rows together. The models are not refusing to answer, and they are not naming your competitor down the street either. For a small local contractor, the closed-book model is very close to a blank page. Three caveats we hold to: these are aggregate rates across the panel and not a ranking of individual businesses, our business-name verification was automated best-effort rather than checked against the RBQ and REQ registers by a human, and at least one name the panel produced closed-book was garbled or invented outright.
A model with nothing specific to recall does not always say so. Sometimes it produces a plausible-sounding company that does not exist.
What AI names instead of you
When a model has no local business to recall, it substitutes: national manufacturer brands stand in for a certification signal, and lead-generation marketplaces stand in for a contractor. The blank-page reading is only half the story, and the less useful half. What the models put in the gap is more diagnostic than what they leave out, because it shows what the training data taught them a good answer looks like.
These are the substitutions we saw across the same probe, not a general observation about AI.
| What the answer needs | What the model names instead | Why that hurts the contractor |
|---|---|---|
| A quality or certification signal | National manufacturer brands (BP, GAF, IKO, Soprema) | The trust cue attaches to a product line, not to the contractor who installs it |
| A place to find a contractor | Lead-generation marketplaces (RénoAssistance, Soumission Rénovation) | The customer is routed into a paid-referral funnel your competitors also sit in |
| A specific local name | Usually nothing, occasionally the province’s largest firms | Recall tracks size and coverage, so the small operator is close to absent |
This is the pattern to bring to your own market. The demand did not vanish, it got redirected: to a manufacturer whose product you install, or to an aggregator that will sell your prospect back to you as a lead. You are paying for that redirection whether or not you have noticed it.
Is this a Quebec problem or a small-business problem?
It is mostly a small-business problem, which is worse news than a Quebec-specific one. The broadest Quebec measurement published to date (Observatoire Noos, étude #1), whose closed-book design we reimplemented with our own prompts for the roofing run above, queried a panel of models closed-book across twelve Quebec sector markets and found local SMEs capturing 18 percent or less of mentions in nine of them, with presence concentrated in sectors where artisanal producers dominate the written record and near-absent in professional and industrial ones. It also reported that Google visibility barely predicts presence in model answers, and that the pattern reads as a bias toward well-known entities rather than a bias against Quebec. It did measure a residual Quebec gap at equal Google visibility, so “mostly notoriety” is the honest reading, not “notoriety only.”
Both findings land the same way for an operator. If the driver were geography, there would be nothing to do. If the driver is notoriety, then the mechanism is the ordinary one: models recall what got written about, repeatedly, in the sources they trained on. That is a hard problem rather than an impossible one, and it moves over years, not quarters.
The second finding is the one that should reset expectations. A first page of Google is not evidence that a model knows you exist. They are different systems reading different corpora, and treating a strong SEO position as proof of AI visibility is the most common mistake we see on this topic.
What actually moves the memory layer
What moves the memory layer is being written about somewhere other than your own website, consistently, over years. As far as we can tell, nothing else does. The citation research already pointed that way. When AI answers do cite sources, the overwhelming majority of those citations point at third-party sites the brand does not own, which we covered in the citation-selection post. The same asymmetry holds one layer down: the corpus a model trained on is mostly other people writing about you.
In practice that means a short and unglamorous list. Local and trade press that gets indexed and syndicated. Association and registry pages that carry your name and category. Supplier and partner pages that describe what you actually do. Accurate, well-sourced reference entries where your category is documented. Client work written up publicly by the client. It is earned rather than installed, and it accrues slowly. Cutoffs move with every model generation, which is why coverage earned this year is what the next generation gets to absorb.
What it does not mean is manufacturing mentions. Seeding fake references across the web to look talked-about is on Google’s own list of things not to do, and it puts a real business at risk to chase a layer that rewards durability above all else.
How to check your own exposure in ten minutes
You do not need a research budget to see where you stand. Open a model, turn web search off, and ask it the question a customer would ask: who should I call for your service in your city. Run it a few times, and run it on more than one model, because coverage varies between them. Then read the answer for three things.
- Does any local business get named, yours or anyone’s? Usually the answer is no, and that is the baseline.
- What fills the gap instead? Manufacturers, directories, marketplaces, national chains. That list is your real competition on this layer.
- If a name does come back, is it a real company? An invented one tells you the model is reaching.
Do the same exercise with browsing turned on and compare. The gap between the two is the part of your visibility that depends on your site being reachable and citable, which is the part you can work on this quarter. What survives with browsing off is the part that took years to build.
What to ask before you hire anyone on this
The memory layer is where AI-visibility pitches get least honest, because it is the hardest layer to affect and the easiest one to make promises about. A useful screening question: ask what they would do about the case where the model answers without searching. A competent answer separates the three layers, says plainly that this one moves on earned coverage over years, and does not attach a timeline to it. An answer that folds it into the same package as a technical audit, or that offers to get you into the model’s memory, is describing a mechanism nobody sells.
Our own position is the same one we bring to a GEO engagement: fix the documented layer completely, hedge the inferred layer deliberately, and treat the memory layer as a long game measured rather than promised. That is also why we still push local SEO hard alongside it, because the browsing case is where near-term work actually pays. If you want a read on where your business currently sits across all three, that is what an audit or a consultant engagement is for.
The uncomfortable part is that the layer that matters most for a small Quebec business is the one that responds slowest. The businesses that will be named in five years are accumulating the coverage now, quietly, while the market argues about schema markup.