How the AI Visibility Benchmark was measured

The AI Visibility Benchmark is research by a company that also sells services in the field it studies. That is called vendor research, and it is entirely common: Ahrefs, Semrush and HubSpot publish the same way. But it also means you should not have to take my numbers on faith. You should be able to check them.

That is what this page is for. It contains everything needed to verify or redo the measurement: the exact questions, the exact model versions, the definitions behind every metric, the uncertainty margins on every claim, and the complete dataset of all 1,125 answers as a free download under a CC BY 4.0 license. Nothing is summarized, nothing is left out, nothing is cleaned up.

The recommendation questions spread across three region layers per profession: five national, three metropolitan, two regional. Every API call got a clean context: no memory, no system instructions, no session carry-over. A fixed two second delay separated calls. Grok was added as the fifth platform on June 12, after a trial run on the therapist set came back error-free; all three professions were then measured in full on Grok.

The prompt set was fixed up front and not changed during the measurement. That is verifiable in the dataset itself: every raw answer contains the exact prompt text that was sent, timestamped. The questions are phrased in natural buyer language, the way a client actually asks them, not in trade jargon. All prompts are in Dutch, because the buyers are.

Type A is a category question, type B a problem question where the asker does not yet know which specialist they need, type C a name verification. For type C, {expert_name} was filled in per run with one of the five sampled experts per profession.

All five assistants were queried through their APIs, with web search or grounding enabled, the way the average user has them in the consumer app. The table shows both the configured model name and the model version each API reported back. The latter is recorded per individual answer in the model_reported field and is the ground truth; models shift under your hands.

Two honest footnotes. The Claude measurements ran through the official CLI; for 149 of the 225 Claude answers it reports an internal helper model (claude-haiku-4-5) alongside Sonnet, which did not write the answer but drove CLI tooling. That is visible in the data rather than cleaned away. And the Gemini configuration used the gemini-flash-latest alias, which pointed to gemini-3.5-flash during the measurement window; since July we pin the explicit version name, because an alias can jump mid-window.

The expert sample per profession was compiled exclusively through neutral sources: professional registers, trade associations and trade media, among them the BNA, the Dutch EMDR Association, the NVRG, the NIP, DuPho, the BIG register and trade titles such as De Architect and Het Houtblad. Emphatically not through AI suggestions, because then you measure what AI already knows, and not from my own network, because then you measure who I already know. Every list was reviewed by hand before any run.

For the recognition questions (type C), five experts were selected per profession, spread across prominence: one nationally prominent, two mid-field, two with a small public profile. For each expert an entity snapshot was recorded: own domain, schema.org present, Wikipedia and Wikidata, LinkedIn activity, trade media mentions, and name ambiguity. Those snapshots ship in full with the dataset, so you can test the relation between entity position and recognition yourself.

Every metric in the report rests on an explicit definition. Here they are.

• Named answer. A recommendation answer (type A or B) counts as named when at least one concrete person or organization was extracted from it and it is not a refusal. Generic advice without names does not count.
• Mention. One named person or organization in one answer, after name normalization (lowercased, accents and punctuation stripped). The core metric is frequency across the five repeat runs of the same question: in how many of the five runs does expert X appear?
• Consistency. Jaccard overlap of the set of named entities between identical repeat runs of the same question on the same platform.
• Recognized (type C). The answer demonstrably describes our sampled expert, verified by hand against firm, region and profession. An answer describing a namesake, or a convincing sounding but confabulated person, counts as not recognized. All 375 recognition answers were reviewed manually. The original automated recognition field sits next to the manual verdict in the dataset, so you can see exactly where the two differ and why the manual step was necessary.

The extraction of persons, organizations and source links from the answers was done by a language model (claude-sonnet-4-6). The first fifty extractions per profession were validated by hand before the rest was processed.

AI answers vary, and a sample is a sample. Every percentage in the report therefore carries an uncertainty margin, and it belongs in the open. Below are the Wilson 95% confidence intervals for the core numbers. Read them like this: where the interval is wide, the point estimate is an indication, not a precision instrument.

All 1,125 answers are public, unedited and machine readable, under a CC BY 4.0 license. You may reuse, redistribute and build on them, with attribution. Every claim in the report can be checked against these files.

The checksums in manifest.json make every file verifiable: whoever downloads a file can confirm it is the original, byte for byte.

Citation: Identity First Media (2026). NL AI Visibility Benchmark, baseline Q2 2026.

This is vendor research: carried out by Identity First Media, which sells services in the field of AI findability. That double role does not disappear by staying quiet about it, but by compensating it with verifiability. Hence this appendix, the open dataset and the fixed honesty paragraph that ships with every publication.

What this study is: a repeatable baseline of the AI visibility of Dutch domain experts, with every raw observation retained and published.

What it is not: independently validated research. It only becomes that when someone else redoes the measurement and finds comparable patterns. The prompt set, the expert sample and the model configuration in the dataset are sufficient for that. Whoever repeats the measurement against the then current model generation should expect different absolute numbers, because models change continuously. The structural patterns are the claim, not the point estimates. If you redo it and find something else, we will gladly publish that next to ours.

The remaining limits of the measurement, including the difference between API and consumer app, are in the main report.

This is the methodology appendix to the Dutch AI Visibility Benchmark. Background on the core concept: what is an entity?