The three layers of visibility in AI assistants
"How do I show up in ChatGPT?" is a badly posed question. A travel business doesn't appear "in AI" in a single way: it depends on whether the assistant remembers, finds or queries. That is the three-layer visibility framework formulated by Félix Pérez of Mirai in 2026 — a direct-booking vendor, worth noting, but the scheme is the clearest the industry has produced and takes five minutes to grasp.
Layer 1: what the model remembers
The first layer is the model's memory: everything it absorbed during training — websites, OTAs, reviews, guides, articles. When the AI answers "from memory", it answers from here.
It is the layer with the least control and the most inertia. It updates by versions, not in real time: a change of trading name can take months to be reflected. You cannot "edit" a model's memory; you can only influence it indirectly and steadily: a healthy, crawlable website, brand consistency across platforms, authority built over time. This layer dominates the earliest phase of the journey, inspirational exploration — when someone asks "which part of the Costa Brava would you recommend for kids?".
Layer 2: what the assistant finds when searching
The second layer kicks in when the assistant goes online live: it crawls indexed pages, OTAs, media, forums, and builds the answer from what it finds. Control here is moderate and the rules resemble classic SEO: indexable content, protected brand, concrete and up-to-date operational information.
This layer dominates qualified discovery — "spa hotels in Girona that accept dogs" — and carries an uncomfortable nuance: being a source no longer guarantees the click, because users resolve many doubts inside the assistant itself. Even so, it is the layer where a change published today can be reflected in days, not months. It is the fast lane for correcting what layer 1 remembers wrongly.
Layer 3: what the assistant queries in real time
The third layer is live data: the direct connection between the assistant and the business's systems through protocols such as MCP, which in 2026 consolidated as the industry's de facto standard. It is the only layer that can answer an operational question with certainty. The framework's canonical example: faced with "is there a room with a terrace free for my dates?", the model's memory doesn't know, search approximates, and only the dynamic connector answers with confidence.
It is the layer with the most control — the data comes from your systems — but it demands infrastructure: an ordered database, a connected engine, protocols. And it carries a warning the industry keeps repeating: if the direct channel doesn't occupy this layer, the big intermediaries will, and they are already integrated into the main assistants.
The common mistake: investing in the wrong layer
The framework's practical value is diagnostic. Three frequent mismatches:
- "The AI says we're closed" → a layer 1 problem (stale memory), mitigated from layer 2: updated, consistent, crawlable content everywhere.
- "We don't show up when people ask for charming hotels in our area" → a layer 2 problem: missing specific, indexable content answering exactly that, plus distributed reputation to back it.
- "The assistant can't say whether we have availability" → a layer 3 problem: without a real-time data connection, no amount of content fixes it.
Each symptom has its layer, and each layer its tool. Spending on "content for AI" when the problem is connectors — or building an MCP server when what fails is reputation — is budget down the drain.
A dose of honesty about "GEO"
The framework's own author concedes something almost nobody selling these services admits: GEO techniques promising specific effects in ChatGPT, Claude or Gemini still lack solid evidence, and Google is moving towards treating manipulation of generative answers as spam. My criterion, applied on my own website too: work on what is documented — crawlability, structured data, verifiable information, consistent reputation — and distrust anyone selling a "guaranteed position" in a system that doesn't even produce stable rankings.
The idea that sums up the three layers fits in one sentence from the framework itself: in the assistant era, the winner is not the most visible but the most verifiable. The first two layers make you appear; the third makes you checkable. And the machine's trust, like the customer's, is earned with data that holds up when someone checks it.
