Does AI name you? The Dutch AI Visibility Benchmark

Five AI assistants, ChatGPT, Gemini, Perplexity, Claude and Grok, readily name names when you ask them for a domain expert. In 95 to 96 percent of recommendation questions, at least one concrete name appeared, and that held evenly across the three professions I measured: architects, therapists and photographers. There is no void. The question is not whether AI names anyone, but which names, and what decides whether you are one of them.

From 1,125 measured answers, two findings hold up across all three professions.

The first is a law: AI does not recommend professionals, it recommends entities. Whether you get named does not depend on how good you are at your craft, but on how strongly and how unambiguously you stand on the web as a recognizable party, plus the platform the user happens to open. The sharpest evidence sits in a pattern that recurred in each of the three professions: the one expert who was structurally missed always shared a name with a better known namesake. An architect who loses to an academic, a psychologist who loses to an M&A partner, a photographer who loses to a British director. Three professions, one mechanism.

The second finding corrects a conclusion that is quickly drawn too broadly. Whether AI names you as a person or as a company depends on how your profession is structured. For architects, a human appears in only a third of the answers; the rest are firms. For therapists, about half. For photographers, nine out of ten. AI does not hide the individual expert on principle; it mirrors whether your market revolves around people or around brands.

Beneath that lie three stable patterns. Who gets named is largely chance: the overlap between identical, repeated questions ranges from 18 to 46 percent, depending on the platform. Being known is also something other than being recommended: asked directly, AI recognizes almost everyone with their own website (82 to 89 percent the right person), but that recognition does not automatically translate into a spontaneous recommendation. And the source AI draws on is, in all three professions, mostly the expert's own domain (32 to 41 percent of the cited links), not review sites and not the press.

This is a baseline measurement: the first reference line, at one moment, across three professional groups. The next edition will measure the shift.

In a hurry and want to know what this means for you? Jump to What works for you.

No one had systematically measured how visible Dutch domain experts are in the answers of AI assistants. That is the reason for this report: it closes a factual gap. More and more clients begin their search not at Google but in a conversation with ChatGPT, Gemini or Perplexity. Whoever is looking for an architect for a renovation, a trauma therapist for a loved one, a photographer for a corporate shoot: what those assistants answer helps decide who comes into view and who does not. Until now that was a black box.

The main question of this study: how often and how consistently do AI assistants recommend Dutch domain experts by name, and which sources feed those recommendations? Five sub-questions sit beneath it. Do expert searches yield concrete names at all? How consistent are those recommendations on repetition? Which source types are cited? Do the platforms differ structurally? And what sets apart the experts who do get named from those who do not?

What this study is not: proof that any particular agency or service works. It measures the current state of the market, a baseline against which future editions can be set. All raw answers are kept as verifiable evidence.

Data collection and analysis were carried out with AI, through automated, repeated querying of five AI assistants.

The measurement is deliberately reproducible. Anyone who wants to check it or redo it should be able to.

Platforms and models. Five assistants through their APIs, with web search or grounding switched on, as the average user has them in the consumer app: ChatGPT, Google Gemini, Perplexity, Claude and Grok. The exact reported model version per answer is logged automatically.

Professions and sample. Three professional groups, each deep and specific: architects (circular construction, repurposing, timber), therapists and psychologists (trauma and EMDR, couples and family therapy, anxiety and burnout) and photographers. For each profession, a sample of experts was assembled exclusively through neutral sources: professional registers, trade associations and trade media, including the BNA, the EMDR Association Netherlands, the NVRG, the NIP and DuPho. Emphatically not via AI suggestions and not from a personal network. For the therapists, ZorgkaartNederland was deliberately not used as a selection source; it is a ratings guide, not a professional register. Every list was reviewed by hand beforehand.

Questions. Three prompt types, built on real buying questions in natural Dutch, the way a client or patient actually phrases them.

• Type A, category questions: "Who is a good [specialist] for [need] in [region]?"
• Type B, problem questions: the question as someone asks it before knowing which specialist they need, for example "Our 1950s building needs to be made more sustainable, what kind of expert do I need and who does that well?"
• Type C, name verification: "Who is [name]?" for a fixed sample of five existing experts per profession.

The recommendation questions (A and B) were asked at three regional levels, national, metropolitan and regional, to test whether scarcity strikes regionally.

Repetition is the measurement. Every question was repeated five times, with a clean context per run: no memory, no system instructions, no session spillover. That is not a detail but the heart of the method. AI assistants almost never give the exact same list twice, so a single run is worthless. The core metric is frequency across repetitions: in how many of the five runs does expert X appear? Five out of five points to a strong entity position, one out of five to chance.

This yields 375 API answers per profession: 10 recommendation questions times 5 platforms times 5 runs (250), plus 5 recognition questions times 5 platforms times 5 runs (125). Across three professions, that is 1,125 measured answers.

Extraction and validation. From each answer, the named people and organizations, the cited URLs, the source type per URL and any refusals were extracted automatically. The first dozens of extractions per profession were checked by hand before the rest were processed.

Disambiguation at recognition. A crucial extra step. Whether a model recognizes "someone with that name" is different from whether it recognizes our expert. With common names that difference disappears if you do not check it explicitly. So every recognition answer was assessed by hand on one question: was the right person described, or a namesake? That step proves decisive. Without it, a fabricated or confused person would wrongly count as "recognized".

Manual check in the apps. Alongside the API measurement, the core questions per profession were replayed by hand in all five real consumer apps, in a clean environment. This tests whether the API findings also hold in the interface people actually use.

Window and evidence. The whole measurement ran within 48 hours, everything time-stamped, because models change under your hands. The raw answers are kept and are not edited; they form both the evidence and, later, content material.

The value of measuring three professions at once lies in what generalizes. What is the same in architecture, healthcare and photography is no accident of one market, it is structural. What differs per profession tells you something about that market. First the factual backbone, then the interpretation.

The manual check in the real consumer apps confirmed the API measurement and brought a sharp extra insight: each platform behaves differently, and that profile is stable across the three professions. It helps decide whether you appear as a human or as a logo, and whether you appear at all.

• ChatGPT (OpenAI) puts firms above people most strongly, has the lowest name ratio everywhere and is the least consistent (18 to 21 percent overlap). With less well known names it first asks for context or poses counter-questions.
• Gemini (Google) names the individual expert most often, neatly lists the namesakes when in doubt, and matches the API almost one to one in the app. In healthcare and for photographers it has the highest recognition of the right person.
• Perplexity is everywhere the most reproducible (43 to 46 percent) and the only platform that reliably retrieves the obscure, local expert; it was the only one that found Cora van Dijk every time. On regional questions it pulls in a local business directory, with addresses and ratings, that the API does not show.
• Claude (Anthropic) is, in all three professions, the weakest at recognizing the right person and, when uncertain, tends to fill in rather than hold back, hence the nonexistent senator. In healthcare it is also the most reluctant to recommend individuals. It also varies by model: the app sometimes could not place an entity that the API did know.
• Grok (xAI) sits in the middle; it names names readily, both firm and person, and backs them up neatly with sources. At the name confusions it chose, like most, the better known namesake. Its X search advantage did not fire: none of the citations came from X, because Dutch experts have barely any footprint there.

Which AI decides your visibility thereby helps decide what information the user receives. At Perplexity you are the specialist from your city; at another platform you are not there, or someone else stands under your name.

One line runs through all of it: AI does not recommend professionals, it recommends entities. Whether you get named does not depend on how good you are, but on how clearly you stand on the web as a recognizable party, and on which platform the user happens to open.

One word captures what it comes down to: entity strength. An entity is a person, firm or brand that the web can place unambiguously and link to other information. Strong entities get found, recognized and recommended; weak entities disappear into the noise or into a namesake. What an entity exactly is, and how you build one, is explained in this companion article.

Three things hold across the three professions.

• Your own domain is the source. In all three professions, the own website is the largest traceable source (32 to 41 percent of cited links). Not reviews, not the press, but what you publish yourself. Without an own domain there is no raw material for AI to draw on.
• A distinctive identity is your entry ticket. Whoever shares a name with a better known namesake disappears, in every profession again. An own, unambiguously placeable entity, with name, domain and profile, is the precondition for being found. That goes for everyone.
• Person or brand depends on your trade. What "becoming visible" concretely means differs per profession.

And then, tailored to your market:

• Are you an architect? The system runs on the firm, not on the person. To come into view as an individual architect, you have to become a recognizable entity alongside the firm, otherwise the logo appears. For an established firm that is less urgent; it is already an entity.
• Are you a therapist or psychologist? There are two entrances, and you need both: your own domain (35 percent of sources) and the official registers (17 percent, a second entrance that is absent in other professions). In a sensitive domain, registration is both a findability signal and a trust signal.
• Are you a photographer? The message is the most direct: the person is the brand, there is no firm to step out from behind. Your visibility is literally your entity. A strong portfolio site under your own name is the foundation, and being findable under your name, not only your Instagram handle, decides whether you appear or a namesake does.

What this comes down to, putting a person on the map as a recognizable entity that AI can place unambiguously, is the work Identity First Media does. Honestly scoped: not equally urgent for everyone. An established firm is already an entity; a photographer with a strong own site already has the foundation in place. The measurement shows precisely where it does take hold: with whoever wants to step out of the shadow of a brand as an individual, and with whoever needs to claim their name unambiguously so as not to be lost to a namesake.

No measurement tells the whole story. Below are the limits of this one, as concretely as possible, so you know what the numbers do and do not say.

• The API is not the app. These numbers come from the APIs. The consumer apps have memory, personalization and sometimes a stronger live search. In the manual check, the apps found the obscure expert more often than the API, so "AI is blind to the small player" is less absolute in the real app. Personalization can go far: the logged-in Grok app visibly tailored care recommendations to the user's profile. For a clean measurement the API is purer, the app is the reality check. Both are reported.
• The professional spectrum is partly structure. That photographers appear as a person 92 percent of the time is partly because photography practices are often named after the person. That is real person visibility, but it is also how the trade is named.
• Frequency is a proxy. How often a name appears is an indication of entity strength, not a proven causal mechanism.
• The largest source category is a remainder group. In each profession, part of the source links fell into "other" (for architects 50.5 percent). The statement about the own domain is robust; the fine breakdown within the remainder group is less so.
• This is a baseline, not a verdict. Fifteen questions per profession, one short measurement window, five experts per profession at the recognition questions. Enough for large, consistent patterns across three professions, not meant as the last word. The thinnest spot is the number of tested names per profession; that is where the first scale-up lies, should there be one.

This is the first edition of the AI Visibility Benchmark by Identity First Media. It sets a reference line: the state of things at this moment. Every quarter I repeat the measurement with the same questions, so the next edition shows what is moving. The raw answers are kept and available for inspection on request.

Identity First Media. Research and writing: Paul Veth.