GEO Research #05: The Five Layers of Existing Within an AI

TL;DR — The fifth edition of the GEO Metrics Research Lab was born in a room in Granada. We organized Alhambra Search, the first SEO and GEO event in the city, and what came off the stage was not a list of tricks: it was a sequence. Five layers, in a specific order, that explain why most GEO strategies fail by starting at step four. This edition reconstructs that sequence and, layer by layer, places our own dataset on top of it — 400,000 responses across 7 engines over 14 months.

Around 300 people between the room and the livestream. Five talks. A full afternoon at the Escuela Internacional de Gerencia in Granada.

On paper, another SEO event. In practice, the talks were ordered with a logic almost nobody perceives in real time: each one solves a problem the next one takes as already solved.

If you build AI visibility before controlling what the AI says about you, you are amplifying your own errors. If you automate before defining your entities, you scale fast on sand. If you go for local without a fixed identity, the model confuses you with the business next door.

The order mattered more than the talks.

How to read this edition: there are two types of number. Those presented from the stage are always attributed to whoever presented them — they are their evidence, not ours, and we reproduce them without validating them. Those marked as GEO Metrics data come from our editions #01 to #04, on a representative sample of the platform database and prompts monitored in real production.

The Thread: The Full Sequence

Nico Bignu opened the day with the framework. The metrics that fall out of play are traffic, impressions and clicks, because in a synthesized response there is no click to count. What survives is share of model: the frequency with which a model includes you in its response. It is not a new metric. It is the same share of voice as always, moved to an environment where the impression is no longer measured in results pages.

From there, five layers:

Layer 1 · Defensive: Before Chasing Visibility, Stop Feeding the Invention

Alicia Puga (Aumenta) opened with the layer almost everyone skips.

Her thesis: hallucination about a brand is not an inevitable model failure — it is a consequence of losing control of your own sources. When the AI finds your address on three different sites, it does not abstain. It fills the gap by plausibility.

She distinguished two types. Passive hallucination: when the model invents opening hours, phone numbers or services when responding about you. And active hallucination, more silent: AI-generated content published without verification that returns to the ecosystem as a source and feeds the error back in. According to the figures she presented, the hallucination rate is around 20% in local and SMB queries, versus 3-10% in global informational queries.

Hence her 20/80 rule: only 20% of what the model knows about you is your website. The remaining 80% is external conversation — press, directories, Reddit, forums, YouTube, social profiles. If that 80% does not confirm what your 20% says, the model lowers confidence or fills in on its own.

Her mitigation protocol: hard data sheets on the website, cross-auditing between models and stress prompt batteries about the brand.

The data layer — where you are most likely to be fabricated:

Our citation study (edition #03) measures the percentage of responses that arrive with no source cited. A response without evidence is the natural terrain of hallucination, and that terrain is not distributed equally across sectors:

In Real Estate and Marketing, almost 1 in 3 responses arrives with not a single source cited. These are exactly the sectors where the talk placed the hallucination peak. The less the engine documents, the more room it has to fill by plausibility — and the more profitable it is to feed it verifiable hard data.

A detail that reinforces the point from our edition #04: only around 1 in 7 responses includes a concrete price figure, and 65.7% of those that do present it with a "from." The hard data Alicia asks you to publish is, literally, what the models are not finding.

Layer 2 · Identity: There Is No Page Two — Either You Are in the Response or You Do Not Exist

Pau Llambí took on the layer that defines who you are for the model.

His starting point: in the classic SERP, there was a click distribution across positions. In a conversational interface, visibility is binary. And models do not rank by repeated keywords — they rank by entities and relationships within a graph.

Hence the canonical sentence: a dense, unambiguous declaration in the first visible paragraph of the homepage, with the formula brand + exact category + value proposition + audience + geographic scope + founders + evidence. Then satellite nodes that develop each segment of that sentence, and an authority circuit to back it: real team with names, author profiles linked to LinkedIn or ORCID, and JSON-LD markup with Organization, Person and sameAs.

The most actionable part was source engineering: mapping sector prompts, tracking which URLs the model consults in the fan-out phase, classifying those domains and doing digital PR only where the AI already looks. And the Wikidata hack as an entry point for brands without enough notability to survive on Wikipedia.

Four measures of disagreement — GEO Metrics data, editions #01 and #03:

Our data supports the thesis from the empirical side. Only 5.6% of monitored queries have the same brand at position #1 across all engines. The average Top-5 brand overlap between two engines is 18.4%. And the cited domain overlap is 10.4%: for the same question, two engines share on average 1 source out of every 10.

The fourth measure is the one that interests us most: the same engine, compared with itself between two consecutive snapshots, retains only 14.9% of its recommendations. There is not even internal agreement.

That is why the canonical sentence is not cosmetic. In an ecosystem where every engine builds a different reality and rewrites it every few weeks, the only variable you control is the coherence of your own definition repeated across every node the model looks at. You do not optimize the response: you optimize the input.

Layer 3 · Operational: MCP — Connect the Model to the Stack Instead of Exporting More CSVs

Raúl Solana brought the strategy down to execution with the four classic pillars — business understanding, technical audit, content and authority — threaded through with Model Context Protocol.

The central idea: stop moving files by hand and connect the model directly to DataForSEO, Screaming Frog, Search Console, GA4 and GEO Metrics' MCP.

Two technical warnings worth keeping: critical content rendered only on the client side remains invisible to the crawlers that matter, and generating hundreds of generic articles destroys authority rather than building it. His alternative to mass production is curation using Search Console data on stalled URLs, plus injection of own UGC — real reviews and ratings exported to CSV — into writing prompts.

The close was a complete outreach loop: extract from the MCP the sources the AI consults for the client's prompts, pull contacts from those sources and launch the sequence from within the conversational environment itself.

The data layer — why speed is not optional:

Our time series explains better than any argument why you need to automate. ChatGPT went from an average of 2.2 sources per response in November 2025 to 20.5 in December and 33.2 in January 2026, before stabilizing at 11.8. Copilot grew 140% in the period, ChatGPT 104%, Perplexity 75%. AI Mode went the other way, down 35%.

A citation audit more than a quarter old is already archaeology: the engine you measured does not cite the same way anymore, and sometimes the difference is three times over.

Layer 4 · Geographic: Local AI Relies on a Map That Is Not Its Own

Nacho Chambó transferred the entity logic to the physical world.

His opening observation is simple and under-acknowledged: ChatGPT, Gemini or Perplexity do not have an up-to-date map of their own. When they resolve a "where to eat nearby," they rely on external geolocation databases, with Google Business Profile as the primary source. Local SEO does not disappear: it becomes the geographic RAG of the models.

Hence the inflexible NAP rule: name, address and phone number identical on the GBP listing, footer, LocalBusiness schema, social profiles and directories. Any discrepancy is exactly the crack through which Layer 1 hallucination enters.

The tactical: hyperlocal landing pages with /service-locality/ structure connected by a semantic internal link ring, 100% response to reviews in under 24 hours with service and locality in the owner reply text, and zero-cost hyperlocal mentions — neighborhood associations, clubs, local blogs.

His case study: a burger joint with no prior reputation outperformed competitors with years of accumulated reviews in AI recommendations.

Layer 5 · Multimodal: YouTube Carries a Lot of Weight — But Not in Every Engine

Emilio García (Campamento Web) closed with the layer almost nobody works: the format that is not text.

His argument: the most advanced engines ingest automatic transcription and analyze video natively. YouTube is the second most cited source by the models.

The actionable: the first three lines of the description carry the heaviest extraction weight, you need to pronounce entities and key terms in the narration because the transcript is the indexable text, and the outlier methodology — finding in small channels the videos that multiply their average by five or ten and replicating the angle and thumbnail — remains the cheapest way to find format that works.

% of responses citing at least one YouTube video — GEO Metrics data, edition #03:

Our data confirms the thesis and gives it a surname. Video is not a universal bet: it is a specific bet on social-profile engines. Grok takes that profile to the extreme, with 80.9% of responses citing social networks, while ChatGPT leans on Wikipedia (15.8%) and news media (13.6%).

In other words: the YouTube channel that gets you into Grok and AI Mode gives you not a single citation in ChatGPT. And the editorial presence that sustains you in ChatGPT is almost irrelevant for Grok. The multimodal layer is essential, but you need to know which engine you are talking to.

A Note on Fan-Out

From the stage it was presented that models consult between 50 and 60 sources in the fan-out phase and end up citing 7 or 8 — a 70-72% reduction. Our measurement of the number of sources each engine exposes in its response fits that order of magnitude for the middle pack (AI Overviews 10.3, Perplexity 9.3, Copilot 3.9) but exceeds it upward in other cases (Grok 45.6, AI Mode 13.8, ChatGPT 13.5).

The practical conclusion does not change: between consulting and citing there is a funnel, and the whole of Layer 2 is about being on the right side of that funnel.

The Five Layers as Actions: What to Do Monday Morning

In order. Skipping a layer invalidates the ones that follow.

Layer 1 · Defensive — first Run a stress prompt battery about your brand across the nine engines and note every invented data point. Publish hard data sheets that contradict them. 29.6% of Real Estate responses arrive with no source. It is the sector where you are most likely to be fabricated.

Layer 2 · Identity — the decisive one Write your canonical sentence and put it in the first paragraph of the homepage. Replicate it in schema, in author profiles and in Wikidata. It is the only layer you control 100% and the one that feeds the other four. Only 5.6% of queries have the same brand at position 1 across all engines.

Layer 3 · Operational — scale Connect the model to your stack via MCP and re-audit every quarter. ChatGPT grew 104% in citation volume in the analyzed period. At this speed, a six-month-old report describes an engine that no longer exists.

Layer 4 · Geographic — local Audit name, address and phone in GBP, website, schema, social networks and directories. Every discrepancy is an invitation for the model to fill the gap.

Layer 5 · Multimodal — by engine If your target engines are Grok, AI Mode or AI Overviews, video is the priority. If it is ChatGPT, your budget returns more in Wikipedia and media. 66.5% of Grok responses cite YouTube. 2.5% of ChatGPT's do.

What We Took Away From Granada

The five layers are not five tactics to choose between. They are an order of precedence: each one only works if the previous one is resolved.

Three operational consequences:

Defense comes before visibility. Amplifying a brand the model describes incorrectly multiplies the error instead of correcting it.

Coherence is worth more than volume. With 10.4% source overlap between engines, there is no content that works for all of them — but there is a definition of yours that can be repeated identically everywhere.

The re-audit cadence is quarterly, not annual. Engines rewrite their citation rules in months.

And an honest caveat about the format of this edition: almost everything presented on stage is field experience from professionals working with real clients, not lab results. We have placed on top of it what we can actually measure. Where the two layers coincide, the signal is strong.

Next stop: Valencia, April 2027. Second edition of Alhambra Search confirmed. Focus on autonomous search agents, consolidated share of model metrics and cases of brands that have moved the needle on their synthetic visibility in the first quarter.

Analysis on a representative sample of the GEO Metrics platform database. Proprietary data, prompts monitored in real production.

Agus and the GEO Metrics Research team

Monitor your share of model across the 9 AI engines → trygeometrics.com