AI search visibility in the EU AI Act market: July 2026 benchmark

This July 2026 benchmark covers 61 distinct EU AI Act questions, each tested once on perplexity/sonar-pro and openai/gpt-4o-search-preview. Across all 122 observations, a Responsible AI Platform, Embed AI or LearnWize domain was cited as a source 46 times, while at least one of these brands was named in the answer text 17 times. This is a time-bound snapshot, not a general search ranking.

The research design and results were published by Zahed Ashkara, legal professional and AI governance specialist. Last substantive review: 21 July 2026.

• Download the complete safe dataset as JSON
• Download all 122 observations as CSV

Key figures

The measurement separates two events:

• Named: at least one portfolio brand appears in the answer text.
• Cited: at least one portfolio domain appears in the source references returned by the model.

A row counts as named or cited as soon as at least one of the three brands or domains is found. Brand counts below can therefore add up to more than the number of positive rows.

Which brand was named

These figures show the number of answers in which a brand appeared. One answer can name several brands.

Which portfolio domain was cited

These are answer counts in which the domain appeared at least once as a source. They differ from total source URL counts because one answer can cite several pages from the same domain.

Sources that appeared most often

The next table shows the five highest source-reference counts in each run. One count is one extracted source URL. It is not a search-result position, and a domain can occur more than once in a single answer.

Within this sample, aiactblog.nl had the most source references in the purchase-oriented Perplexity run, with 15. It ranked second in the Perplexity knowledge run with 44, behind artificialintelligenceact.eu with 51. In the GPT-4o knowledge run, the domain shared the highest count, 7, with the European Commission's AI Act Service Desk.

What the results show

1. Being cited is not the same as being recommended

On Perplexity's knowledge questions, a portfolio domain was cited in 57.8% of answers, while a portfolio brand was named in 4.4%. The material was therefore used as a source much more often than the organisation or provider appeared explicitly in the answer text.

2. Results vary sharply by model

For the same 45 knowledge questions, Perplexity produced 26 answers with a portfolio source and GPT-4o Search Preview produced 7. Brand mentions showed the opposite pattern: 2 on Perplexity and 7 on GPT-4o. A single combined visibility score would hide that model dependence.

3. Purchase-oriented questions produce more brand names

In the purchase-oriented set, the share of answers containing a portfolio brand rose to 31.3% on Perplexity and 18.8% on GPT-4o Search Preview. The set is small and weighted toward Dutch prompts, but the difference from the knowledge set supports measuring both types separately.

4. Portfolio source authority is visible, but not exclusive

Responsible AI Platform appeared frequently as a source alongside official EU sources, independent AI Act publications, consultancies and training providers. The measurement supports an authority claim within this specific question set, but not a claim that one source is number one everywhere.

Methodology

Sample and collection dates

• The knowledge set contains 45 distinct prompts. The prompt set is dated 28 June 2026 and both model runs were collected on 30 June 2026.
• The purchase-oriented set contains 16 distinct prompts. Both model runs were collected on 6 July 2026.
• Each prompt was run once per model. The full dataset therefore contains 61 distinct prompts and 122 prompt-model observations.
• The same prompts were submitted to both models in the same order for each question set.
• Each request allowed a maximum of 600 output tokens.

Technical measurement rules

Requests were made through the OpenRouter Chat Completions API. The measurement tool:

1. sent each prompt as one user message;
2. scanned answer text for predefined portfolio brand patterns;
3. extracted source URLs from the model's citations field or URL annotations;
4. normalised source domains by removing www.;
5. marked an observation as cited when at least one source URL referred to aiactblog.nl, embedai.nl or learnwize.ai;
6. counted competitor names only when a predefined text pattern appeared in the answer.

A legacy technical brand label in the purchase configuration was normalised to Responsible AI Platform in the public dataset. Three Dutch prompts were incorrectly marked EN in the source configuration. Their public language value is corrected to nl-NL, while the original label remains available in a separate field for auditability.

Data minimisation

The public dataset contains prompts, model identifiers, derived true-or-false fields, detected brand and domain names, source counts and safe source fields. Full model answers and answer excerpts are not published. Query strings and fragments were removed from cited portfolio URLs.

Limitations

• This is a snapshot of two model configurations, not a representative market study.
• Each prompt was run once per model. There are no repeat runs or confidence intervals.
• Model indexes, retrieval systems and answer wording can change without notice.
• After language correction, the purchase-oriented set contains 13 Dutch and 3 English prompts, so it gives extra weight to the Dutch market.
• Answers without source URLs remain in the denominator. This occurred especially in the GPT-4o Search Preview run.
• Brand matches measure presence in text, not sentiment, order or recommendation strength.
• Source counts measure URL references. They are not Google rankings, market share or unique visitor counts.

A future edition can repeat the same fixed prompts. Only then will the dataset form a time series showing change by model, question type, brand and domain.