Why AI doesn't recommend the same to everyone and how to…

Why doesn't AI recommend the same to everyone?

The question seems obvious, but most GEO strategies ignore it entirely. Language models don't work like a classic search engine that returns a static list of results. They work as probabilistic systems that generate responses conditioned by multiple variables: the prompt text, the language, the geographic location, the conversation history and, on platforms like ChatGPT, the user's persistent memory.

A study by SparkToro and Gumshoe.ai (January 2026) quantified it: 600 volunteers ran 2,961 identical prompts on ChatGPT, Claude and Google AI. According to their data, the probability of obtaining the same list of brands in two executions of the same prompt was less than 1%. The lists varied in composition, order and number of items. It's a study with limitations (two-month time window, non-representative volunteers), but the magnitude of variability is hard to ignore.

That variability isn't noise. It's architecture. LLMs are designed to generate diverse, non-deterministic responses. Each execution traverses slightly different probabilistic paths. And when you add user variables — location, language, previous context — divergence amplifies.

We've verified this with clients. When you introduce profile variables (language, travel context, location) in the prompts of a GEO audit, the recommendation lists change completely. It's not theory. It's something any brand can verify in 30 minutes.

LLMs generate probabilistic responses conditioned on the user's profile. The same question produces different recommendations depending on who asks, from what location and with what history. Treating GEO visibility as a single metric is ignoring how these systems work.

What variables change what AI recommends?

Not all variables weigh the same. We've identified three as the most determining in practice:

1. Geographic location and regional context

The user's location directly influences responses, even when the prompt doesn't mention any location. Otterly.ai documented that ChatGPT localizes its responses based on the user's location through two mechanisms: web browsing context (when search is activated) and information inferred from previous conversations. This means that the same question from Madrid and from Munich generates different responses without the user doing anything explicit.

For brands with a local or regional component (tourism, hospitality, professional services), this is critical. Your content may be optimized for one market and be completely invisible to another.

2. User profile and conversational memory

ChatGPT already operates with a two-layer memory system: explicit memories (which the user approves) and implicit insights from previous conversations. A user who has repeatedly talked about traveling with kids will receive different recommendations than someone who has explored solo adventure options. The model doesn't start from scratch: it starts from what it knows about you.

And it's not just ChatGPT. Gemini, Copilot and other models are moving in the same direction: responses increasingly conditioned on the user's accumulated profile.

3. Prompt formulation and implicit segment

SparkToro's research found something revealing: when 142 participants wrote their own prompts about headphones, the average semantic similarity was only 0.081. That is, users with the same intent formulate the question in radically different ways. And each formulation activates different paths in the model.

A CEO who asks "what digital growth consultancy do I need" doesn't write the same as a CMO looking for "tools to optimize visibility in AI Overviews". Same territory, same problem, completely different responses.

The three variables that most condition an LLM's response are the user's geographic location, their accumulated profile on the platform and the specific formulation of the prompt. A GEO strategy that doesn't contemplate these three layers works with an incomplete model.

What data confirms that AI visibility is fragmented?

The intuition that "AI changes what it says depending on who asks" needs solid evidence. And there is.

The analysis of Ahrefs on 730,000 query pairs (September 2025) showed that Google's AI Mode and AI Overviews — two surfaces of the same ecosystem — cite different URLs around 87% of the time for the same search. They only share around 14% of cited sources. And yet, they reach semantically similar conclusions in 86% of cases. Same answer, radically different sources.

On the user side, McKinsey's research (August 2025, n=1,927 consumers in the US) found that approximately half of respondents intentionally use AI-powered search to make purchase decisions. The data suggests use across the funnel: 73% for researching categories, 61% for comparing products, 57% for getting personalized recommendations. Each of these interactions is potentially conditioned on the user's profile.

The fragmentation doesn't stop there. Recent Ahrefs data (February 2026) shows that AI Overviews reduce the CTR of the first organic position by 58%. And that only 38% of URLs cited in AI Overviews also appear in the organic top 10, versus 76% in mid-2025. AI's citation system is actively decoupling from classic ranking.

Asier's opinion: When you see these numbers together, the conclusion is uncomfortable but clear: "GEO visibility" as a single metric is a useful fiction. What you really have is visibility fragmented by platform, by surface, by user profile and by prompt formulation. If you only measure "how many times my brand appears in ChatGPT" without segmenting by persona, you're measuring an average that represents no one.

GEO translation: if your brand appears in 60% of responses to a generic English prompt but in 8% to a Spanish prompt with family context, you have a segmentation problem, not a global visibility one.

AI visibility is fragmented by platform, surface, location and profile. According to Ahrefs, cited sources diverge around 87% between AI Mode and AI Overviews for the same query. According to SparkToro, recommended brand lists repeat less than 1% of the time in their sample. The single visibility metric is insufficient.

What does this imply for your GEO strategy?

If AI personalizes by profile, your content strategy has to do so too. It's not about creating a page for each type of user (that's unviable and counterproductive). It's something more surgical: understanding which user profiles generate the prompts that matter for your business and ensuring your content answers each of them.

There are three concrete moves:

Map your prompts by persona, not just by keyword

Traditional keyword research assumes intent is static. In GEO, intent is dynamic and conditioned. The same need ("improve my digital visibility") generates completely different prompts depending on who formulates it. A CEO will ask for "digital growth consultancy for a mid-sized company". A CMO will ask "GEO tools to optimize presence in AI Overviews". A digital marketing director will look for "comparison of AI visibility tracking platforms".

Each of those prompts activates different sources, tones and recommendations. If your content only covers one of those angles, you're invisible to the other two.

Segment by location and language

Tools like Scrunch and Gumshoe.ai already allow segmenting GEO tracking by region and persona. And the data shows that regional divergence is significant. An identical prompt executed from Germany and from Spain can generate recommendation lists with no brand in common.

For brands with international markets or with dependence on tourism, this isn't a nuance: it's the variable that explains why content investment doesn't translate into visibility in the markets that matter.

Optimize for "query fan-out", not just for the literal prompt

Google has confirmed that both AI Mode and AI Overviews use a technique called query fan-out: when a user asks a question, the system breaks it down into multiple sub-queries and searches for sources for each one separately. Recent Ahrefs data suggests this mechanism is gaining weight in source selection, especially since the adoption of Gemini 3 as the default model in AI Overviews (January 2026).

This means your content needs not just to respond to the main prompt, but to cover the semantic adjacencies the model will explore. And those adjacencies vary depending on the user's profile.

An effective GEO strategy in 2026 maps prompts by persona (not just by keyword), segments visibility by location and language, and optimizes for the query fan-out that models execute internally when processing each query.

How does this look in practice?

Operational pattern: asymmetric visibility by profile

In recent months we've observed a recurring pattern in at least three clients in the B2B services and experiential tourism sector. We call it internally "asymmetric visibility by profile". It's not an isolated case: it's a pattern that repeats with enough regularity to consider it structural.

How it manifests: the brand has reasonable presence in AI responses for generic prompts in English or neutral Spanish. But when you introduce profile variables — specific language of the target market, family context, mid-high budget, search from a specific location — the brand disappears. It doesn't drop in position: it disappears.

Anonymized mini-baseline (experiential tourism, Canary Islands, Q4 2025): we ran the same prompt in three languages (Spanish, English, German) with and without profile context (couple's trip vs. family with kids) in ChatGPT and Perplexity, 20 executions per combination. The client's brand appeared in 65% of executions for the generic English prompt. In German with family trip context, the frequency of appearance dropped to 5%. In Perplexity, it directly didn't appear in any execution in German. The German family tourism segment was, according to their own sales data, the one that generated the highest average ticket.

Diagnosis: the content was optimized for one user profile (the most generic, in the marketing team's language) but didn't cover the semantic angles that activate responses for other profiles. Coverage was missing in German with family travel vocabulary, authority signals were missing on platforms that LLMs consult for the DACH market, and content was missing that answered the sub-queries (query fan-out) that this type of prompt generates.

Action: three pieces of content specific to the DACH family profile were created, listings on review platforms with presence in the German market were optimized, and the schema was adjusted to reflect the family offering. After 8 weeks, the frequency of appearance in German prompts with family context had risen from 5% to 35%. It's not an absolute success, but it confirms that the problem was content segmentation, not general brand authority.

My opinion: The most common mistake I see is assuming that "if I appear in ChatGPT, I'm already covered". No. You appear for one type of user, with one type of prompt, from one location. For your real ideal client — the one who pays and repeats — you may be completely invisible. And you won't know it until you measure segmented.

This pattern connects directly with what at Elevam we work through the Prompts x Models Matrix (Elevam Labs): a tool that crosses your ICP's real prompts with the responses of the main LLMs, segmented by profile, language and platform. It doesn't measure "average visibility". It measures visibility for who matters.

The "asymmetric visibility by profile" pattern appears recurrently in brands with international markets or diverse segments: reasonable presence for generic prompts, invisibility for the profiles that actually generate business. The anonymized Canary Islands tourism case illustrates it: from 65% to 5% appearance when switching from generic English to German with family context.

What's the scale of this opportunity (and this risk)?

Context numbers help size what's at stake. ChatGPT surpassed 800 million weekly active users in February 2026. Google AI Overviews already appears in a significant percentage of searches (around 50% according to trend analysis estimates), with a projection to exceed 75% in 2028. McKinsey estimates that up to 750 billion dollars in consumer spending could be channeled through AI search by 2028.

And yet, according to the same McKinsey report, only 16% of brands systematically track their performance in AI search. The remaining 84% don't know if they appear, for whom they appear or what AI says about them.

Now add the personalization layer: of that 16% that does measure, how many segment by user profile? The honest answer is: almost none. Most measure average visibility, which is like measuring the average temperature of a hospital and concluding all patients are fine.

McKinsey research on AI personalization points out that more than 75% of consumers reject content they don't feel relevant to them. If AI recommends your brand to the wrong profile or doesn't recommend you to the right profile, you're not in the game.

McKinsey projects $750B in spending through AI search by 2028, but only 16% of brands measure their AI visibility and practically none segment by user profile. The gap between the scale of the opportunity and the sophistication of measurement is enormous.

How to start segmenting your GEO visibility by profile?

You don't need an enterprise tool to take the first step. But you do need a change of mindset: stop thinking of "my AI visibility" as a single metric and start treating it as a matrix.

An operational approach we apply at Elevam:

01
Define 2-3 real personas from your ICP
Not generic archetypes. People with specific language, location, sophistication level and type of question. The CEO of a €5M company in Barcelona looking to grow digitally doesn't write the same as the Head of Digital of a multinational in Frankfurt.
02
Generate 5-8 realistic prompts for each persona
Not SEO keywords disguised as prompts. Real questions, formulated as each profile would write them. In their language. With their context. If your market includes German tourists, those prompts need to be in German.
03
Run each prompt on at least 3 platforms and 10+ times
ChatGPT, Gemini, Perplexity. Record: does your brand appear? How frequently? In what context? What competitors appear? What sources are cited?
04
Cross results in a Persona × Platform × Visibility matrix
There you'll see where you have coverage and where you have blind spots. Those blind spots are very specific content opportunities.

This process is, in essence, the simplified version of what we execute through Elevam's HSA Protocol when we diagnose a client's AI presence. The difference is that the HSA systematizes it, crosses it with business data and generates a prioritized action plan. But the principle is the same: don't measure average visibility; measure segmented visibility.

To segment GEO visibility, define 2-3 real personas from your ICP, generate realistic prompts for each one, run them 10+ times on 3+ platforms and cross the results in a Persona x Platform x Visibility matrix. The blind spots that emerge are concrete content opportunities.

Conclusion: generic GEO is no longer enough

During the first years of GEO, the question was binary: do you appear or not in AI responses? That question is no longer useful. The relevant question now is: for whom do you appear?

The personalization of AI responses isn't a hypothesis: it's a fact documented by SparkToro, Ahrefs, McKinsey and by the very architecture of the models. Your brand doesn't have "an AI visibility": it has multiple visibilities conditioned by profile, language, location and platform.

Asier's opinion: I think we're at an inflection point. Brands that understand that GEO is inherently segmented — and act accordingly — will gain a disproportionate advantage. Not because the technique is difficult, but because almost no one is executing it. It's the equivalent of the early years of SEO: whoever moves first wins more.

If you want to understand where your brand is on this new map, the starting point is a diagnosis that segments by real profiles. It's what we do through Elevam's CREF© approach: a system that connects visibility, content, conversion and business. Not as separate pieces, but as architecture.

The relevant GEO question is no longer "if you appear in AI", but "for whom you appear". Brands that segment their visibility by user profile, language and platform will capture a disproportionate competitive advantage in a market where 84% still doesn't measure their presence in AI search.

Next reading:

If your company depends on digital visibility to capture customers, you need to know how AI sees you. Explore how our applied GEO and AI service works or check the GEO Glossary to familiarize yourself with key concepts.

Shall we work together?

If you want to apply this in your company with a team that combines technical SEO, GEO and paid acquisition measured against the income statement, request a no-commitment audit. You can also check real case studies or read the public GEO baselines that Elevam Labs publishes every quarter.

Why AI doesn't recommend the same to everyone and how to segment your GEO visibility