Data study · 18 Oct 2026
64 Claude answers to 32 buyer questions in 2 EU niches: who gets linked
We asked Claude 32 buyer questions about loose-leaf tea and specialty coffee in Belgium and France, twice each (64 answers, 18 Oct 2026). The most-linked shop appears in 31% to 56% of answers, and the top three take 17% to 31% of all shop links.
4 niche-and-country cells · 64 answers · Claude Haiku 4.5 via the Anthropic API, web search on · Download CSV · One-page PDF
The 4 cells
| Cell (full table on its page) | Answers | Most-linked shop | Top-3 share | Sites seen | Run overlap |
|---|---|---|---|---|---|
| Thé en vrac, France | 16 | auparadisduthe.com (5/16) | 17% | 45 | 45% |
| Café de spécialité, France | 16 | lestorrefacteurs.cafe (5/16) | 20% | 39 | 50% |
| Café de spécialité, Belgique (FR) | 16 | cafedujour.be (9/16) | 31% | 30 | 57% |
| Specialty koffie, België (NL) | 16 | cafedujour.be (6/16) | 26% | 33 | 77% |
Top-3 share: the three most-linked shop sites as a share of all shop links in the cell. Run overlap: average share of shop sites that both runs of the same question linked.
What the numbers say
- MeasuredLinking is concentrated. The most-linked shop appears in 31% to 56% of a cell's answers; the top three shops take 17% to 31% of shop links.
- MeasuredA single answer is a poor guide. Two runs of the same question share only 45% to 77% of the shop sites they link, on average per cell.
- MeasuredLinks go to shops' own sites: 279 of 326 links across the 4 cells (86%). Department stores, maps, guides and review sites take the rest.
- MeasuredClaude linked 1 domain with no working page (confirmed by a second check), in 1 of the 4 cells, 1 time in total. It is excluded from the rankings.
- Not tested yetWhy a shop gets linked: published facts (starting price, delivery time, subscription terms), reviews, third-party mentions or technical traits. That needs a linked-versus-unlinked comparison; it is the next measurement round.
Method and limits
- Engine: Claude Haiku 4.5 (claude-haiku-4-5) through the Anthropic API, web search on (at most 2 searches per answer), no history or personalisation. API answers differ from what a signed-in person sees in the Claude app, and other assistants (ChatGPT, Gemini, Perplexity) may link different shops: this study covers one engine only.
- Questions: 32 buyer questions in French and Dutch (8 per cell), each asked twice, 64 calls in total. 64 complete answers kept; failed runs are dropped (none failed).
- “Linked”: the shop's domain is among the answer's links. We do not count names in running text.
- Coverage: 2 niches (loose-leaf tea and specialty coffee), 2 countries (Belgium and France), 2 languages. 64 answers is a small sample: read the percentages as counts, not as market share.
- Limits: one engine, one model, runs on 18 Oct 2026; answers change from day to day. No traffic or sales estimates. Domains are as Claude linked them; we checked which ones load on 18 Oct 2026. Earlier counts from a different method (ChatGPT's app, 13 and 14 Oct) are retired and not mixed in here.
- Raw data: every answer is stored with its timestamp. Pooled CSV: download.
Is your shop in these tables?
We run the same kind of questions for your category and count how often your shop is linked. Free, one page, in your language.