The invisible share: who AI recommends

We asked five AI engines the question any Spanish freelancer or small business asks before switching invoicing software. Without putting a single brand in front of them. "I want to stop keeping my invoices in Excel, where do I start?". "I need something to run my workshop in Zaragoza, invoices and clients". Thirty different questions, each repeated five times, across five engines. Seven hundred-odd answers, analyzed one by one.

We were looking for one thing: when the model chooses on its own, with nobody whispering a name, who does it recommend?

Quite a lot of people who aren't in the conversation. Of all the recommendation they handed out, 55% went to brands that weren't even on our list of the ones that matter in this sector. FacturaDirecta, Declarando, Cuéntica, a3factura, STEL Order. The immediate reaction is to think that's noise, filler brands. It isn't. It's the first clue that AI isn't playing the game you think it's playing.

This is the first cut of an Elevam Labs study, measured with Antropus, our GEO measurement platform. Let's take it piece by piece.

What AI hands out is not market share

When it recommends, a model doesn't consult a ranking of who bills the most or who has the most clients. It doesn't have that table. What it has is a map —its knowledge graph— where every brand, every concept and every intent is a node, and the nodes are joined by the strength with which they appear together across all the text the model has read. When someone asks "software to invoice as a freelancer", the model starts at that intent's node and pulls the brands it has most strongly wired to that node. The recommendation is the result of that walk.

And that map isn't built with invoicing. It's built with text. That's why what AI hands out is a different thing from market share, a thing almost nobody is measuring: your presence inside the model's graph for the intents that feed you. Call it graph share. It may look like your market share or it may have nothing to do with it. And when a client asks ChatGPT, the second one is what rules.

This, which sounds abstract, explains with precision everything we found in the data.

Sage exists, but it's plugged into the wrong cable

Sage is the brand with the most muscle in the set. When you ask a model about it directly —"is Sage good for keeping a small company's accounts?"— it answers with judgment, recognizes it as a serious option, describes it well. On reputation, when you ask about it, it scores very high. The model knows exactly what Sage is.

And even so, when the user doesn't name it, Sage barely shows up. This is the blind recommendation share of the seven known brands:

Sage, the one the model knows best, in sixth place and below four smaller natives. It's not a contradiction; it's exactly what the graph predicts. The model has Sage built as a strong entity, yes, but wired to the node "company accounting, multinational, corporate software", not to the node "the Spanish freelancer who wants to stop fighting with their accountant". It's a huge node connected to the wrong intent.

The small ones don't win despite being small. They win by being well connected

That 55% leaving the set isn't better or bigger brands. Declarando, Cuéntica, FacturaDirecta are, several of them, small businesses. They win recommendation for one concrete reason: their trail across Spanish text is densely glued to the freelancer's specific intent. They appear again and again in articles about the 130 tax form, about freelancer income tax, about Verifactu, about how to register as self-employed. The model has them wired, hard, to exactly the node the user is touching.

The graph doesn't reward the size of your brand. It rewards the density of the connection between your entity and the intent. A tiny brand can be, for a very specific intent-node, a stronger node than a multinational. And that completely changes where the game is played: you don't win by being bigger, you win by being, for that intent, more unambiguous.

There is no "the AI". There are five different graphs

Here's the data point that should stop in its tracks anyone building their AI visibility strategy as if it were one block. The same brand is built differently in each model. Look at Holded:

Holded is the outright leader in Claude and in Google's generative block, and just another name in the crowd in ChatGPT. Not because ChatGPT treats it badly, but because each model learned from a different corpus and compresses differently: ChatGPT tends to spill long lists, and in a long list the leader loses weight; Claude gives few options and backs the ones it picks hard. Each one has its own map, with its own cables.

Which translates into an uncomfortable sentence: "I show up in AI" means nothing until you say which one. What puts you on top in one graph can leave you diluted in another. A strategy that doesn't work engine by engine is shooting at a blur that is really five targets, each moving on its own.

The story AI tells about you is written by someone else

What does the model build your brand's profile with? With what you say about yourself, or with what others say?

With what others say, almost always.

The node the model has of you is made, mostly, of someone else's material. If a comparator scores you badly, if an article puts you on the wrong list, if a two-year-old video describes you with a product you no longer sell, that's what goes into your profile. You don't control your entity in the graph; third parties are writing it, and you're the last to know. Showing up in AI and governing what it says about you are two different things, and the second is what really protects a business. That, deep down, is what correcting what ChatGPT says about your company is about.

From ten options to four

In a Google search you see ten results and, if curiosity strikes, you scroll down for more. AI doesn't let you: it gives you a short answer with a handful of names and that's it.

Let's be honest about the nuance, because it matters so we don't sell smoke: it's not brutal compression where three brands take everything. The choice spreads across about four, and the three leaders only grab the bulk of the recommendation in 7.6% of answers. There is no monopoly. But going from ten to four already leaves out more than half of those who were visible. On Google, being sixth still gets you a line. In AI, being sixth is not existing for that user.

What this changes for whoever splits the budget

Five things, for whoever decides where the money goes and not for whoever fills in a visibility report.

1. 01

   You don't know your share in AI until you measure it

   And it's not the one you assume from your Google ranking or your awareness. The first thing is to know where you really stand, and engine by engine, because there is no single snapshot.

2. 02

   Your SEO and brand budgets aren't buying you this seat

   They're for other things. This is played on how the model has you built as an entity and which intents it has you connected to, which is a separate board and worth treating as a separate game.

3. 03

   The underlying work isn't only your website

   It's your entity —the model knowing clearly what you are and for whom— and your connections —being densely associated with the specific intents that make you money, in the sources the model reads—. Four out of five times, those sources are third parties.

4. 04

   You have competitors you aren't tracking

   Inside your client's buying decision there are brands AI recommends that you aren't watching. Watching only the usual suspects leaves you blind to more than half the market AI hands out.

5. 05

   The goal stops being getting recommended

   The goal is to exist, well defined and well connected, as an entity in the model's graph for what matters to you. Your market share and your graph share can diverge to a frightening degree, and it's the second that decides the recommendation. You can lead your market and be, to AI, a weak node.

If you want to see your own snapshot —where you stand in each engine, which intents it has you connected to, and who's eating your recommendation without you knowing—, that's exactly what we do in Elevam's AI visibility audit.

How we measured it

Thirty real commercial questions in Spanish, split across six types —reputation, generic discovery, comparison, purchase case, price and integration—, most without naming any brand, which is where the model truly chooses. Each one five times, in ChatGPT, Claude, Perplexity and Google's two generative modes.

The whole analysis runs on Antropus: it's what scores the strength of each recommendation with a fixed rubric, what resolves the entity so it doesn't confuse a brand with a homonym, and what classifies each source as owned or third-party. Share is calculated only on the blind questions, weighting the strength of each recommendation and counting any brand the engine names —not only the tracked ones—, which is what reveals the invisible 55%.

The limits, no makeup. This is a pilot of a single niche —management and invoicing for freelancers and small businesses in Spain—, a snapshot, not a time series or a census. Gemini was left out of the share because its API quota ran out halfway through the study; it enters the next round, which will go from five engines to six. And one thing for intellectual honesty: we went in expecting to find a source oligopoly, a few domains controlling almost everything. It didn't pan out: concentration exists but is moderate, the five most-cited domains add up to 21%. We report it because we measure to say what's there, not what we set out to find.

Methodological and verification annex

This section exists for one thing only: so anyone can check that the numbers come from real measurements and not from nowhere. It's what makes the study truthful. Everything here is traceable down to the raw data.

Real coverage (what went into the numbers and what didn't)

Totals across five engines: 739 valid answers · 469 blind (the share's base) · 99 reputation · 171 comparisons. Gemini stays out of the share analysis because it's not a finding about Gemini, it's an API quota limit; its 5 valid answers do count in the source analysis.

How we validated that the measurement is reliable

The 900 answers were scored by a model following a fixed five-value rubric. To not take it for granted, a sample was re-scored with a larger model (the expensive one) and agreement was measured:

They agree in 96% of cases within one rubric step, with zero large divergences except a single case. Share rests on a measured score, not an assumed one.

Before closing numbers, the pipeline went through an adversarial audit with verification by independent refuters: 22 problems detected, 21 confirmed and fixed. Three were serious enough to move the headline figure:

• Comparison questions were slipping into the share calculation, inflating the known brands to a false 57.1%. Fixed, the real figure dropped to 44.9%.
• Gemini's redirect-link resolution wasn't working (it recovered 0). Fixed, it recovered 63 of 63.
• A check that the batch was complete before computing was missing. Fixed: the analysis now warns if the data is half-baked.

Others fixed: homonyms (Sage Publishing, the Inca quipu) and the "not mentioned" cases stopped counting as presence; variants of the same name were unified (STEL Order = Stelorder); non-brands were vetoed (Verifactu, which is the regulation; the Tax Agency); owned subdomains count as owned; and percentages close to 100.

Independent verification of the dataset

To not rely on the pipeline's word, the raw dataset was recomputed from scratch with a separate analysis —without using the engine's code, only the 900 raw rows—:

Reproducing each brand to the last decimal requires the Antropus engine, which is proprietary; the raw data, which is what sustains credibility to third parties, is what is documented here. Total cost of the pilot: between 10 and 11 euros (preparation and tests, full batch, and the validation of the scoring model).

Frequently asked questions

What is graph share?

It is your presence inside an AI model's knowledge graph for the intents that bring you clients: how strongly your brand is connected to the question the buyer asks. It is not your market share (how much you bill) or your awareness (how well known you are), it is which intent-node the model has you wired to. When someone asks without naming brands, AI recommends according to that connection, not according to your size.

Why does Sage, the brand AI knows best, barely get recommended?

Because recognition builds a big node, but it doesn't guarantee it's plugged in where you sell. In the study, Sage scores very high on reputation when you ask about it, but blind it lands sixth (3.5%), below four smaller natives. The model has it wired to "corporate, multinational software", not to "the Spanish freelancer who wants to leave their accountant". It's a huge node connected to the wrong intent.

Is appearing in ChatGPT the same as in Claude or Perplexity?

No. There is no "the AI" as a single place: each engine learned from a different corpus and compresses differently. Holded takes 26.1% of the recommendation in Claude and only 3.2% in ChatGPT. "I show up in AI" means nothing until you say which one; a strategy that doesn't work engine by engine is shooting at five different targets thinking it's one.

How is recommendation share in AI measured?

You ask the engine real commercial questions without naming brands, repeated several times to average out the model's natural variation, and you weight the strength of each recommendation with a fixed rubric (from main recommendation to mention with reservations). In this study it was 30 questions, 5 engines and 5 repetitions, analyzed with Antropus, counting any brand the engine names —not only the tracked ones—, which is what reveals the invisible 55%.

The entity and knowledge-graph framework on these pages is the mechanic that best explains what we measured. It's the first cut of an Elevam Labs series: the next installment takes it to the data —how AI has each brand's profile built, attribute by attribute, and which intents it has it connected to—. That X-ray, not the recommendation share that is only its symptom, is the real terrain of GEO. The first chapter of the series, on how recommendation shifts with the language of the question, is here.

If you want to replicate it on your brand or see your graph share engine by engine, let's talk.

Asier Lopez Ruiz is CEO of Elevam and Antropus, a SEO and GEO pioneer in Spain. 2x Forbes Best SEO Agencies Spain. Data generated with Antropus.