Client: a family-owned group with two businesses: residential property and professional beauty products Work: AI visibility baseline across five models, three languages and 18 brands, plus a GEO roadmap Status: baseline delivered, roadmap handed over. No follow-up measurement yet.
Situation
One owner runs two very different businesses: residential property and a portfolio of professional beauty brands. Their buyers can now ask ChatGPT or Perplexity where to buy a house or which salon hair colour is best. Nobody knew what those assistants answered, and there was no baseline to measure against.
What we did
A quick first pass (12 prompts in three languages on Perplexity) showed the shape of the problem. Then we built a repeatable test: 90 prompts, 30 per language, across 11 query types, from "who makes X" to wholesale and everyday shopper questions. Each went to Perplexity Sonar 2, GPT-5.4, Gemini 3.1, Claude Sonnet 4.6 and Kimi 2.6, giving 450 saved answers.
A deterministic script, not an AI judge, counted mentions of 18 client brands in every answer, so each percentage can be re-run from the raw data. We also hand-labelled which brands should and shouldn't appear for each prompt, to test how disciplined each model is.
What we found
- The place is known, the company is not. The community's place name appeared in 15.6% to 27.8% of answers depending on the model. The developer's brand: 1.1% to 5.6%. Two of the developer's other projects averaged under 1%. In one test the developer's own website was the top cited source, yet the answer never named the developer.
- Not every mention can be trusted. Where a brand had no reason to appear, Claude Sonnet 4.6 and Gemini 3.1 still named it 5.0% of the time, GPT-5.4 15.0%, Perplexity 23.3%. Raw Perplexity numbers overstate real visibility.
- Direct questions work, category questions don't. On "who makes this brand" queries the cosmetics brands appeared in 40.0% of answers. On product-category rankings and everyday shopper questions: 0.0%. On wholesale supplier queries: 0.4%, although wholesale is a core channel for the group.
- Language matters. Visibility dropped from one local language to the other and was lowest in English, in all five models.
- No knowledge-base footprint. 44 encyclopedia checks (11 entities, 4 languages) found nothing.
What's next
Quick wins (two weeks or less): Organization schema, a clear "who builds this" block on the site, category tags on retail catalogues. Then one to three months of knowledge-base entries, English pages and Q&A content. The plan is to re-run the same 90 prompts at 60 days and treat any move above 3 percentage points as real. That re-run hasn't happened, so there's no result to report yet. What exists is a baseline anyone can reproduce.