Client name withheld. Figures are from our measurements/deliverables on the stated dates.
Business: two companies under one ownership — a countryside real-estate developer and a professional cosmetics manufacturer, sold through separate channels but sharing one measurement question: do AI answer engines know these brands exist.
Phase 1: first read, April 26, 2026
Twelve prompts, run in three languages through Perplexity.
- The developer's flagship housing project ranked 2 of 10 in a "top cottage communities near the capital" answer.
- Safety questions (backup power, water, shelters) returned zero mentions of the developer brand, in both the local language and English, across all prompts asking them.
- The cosmetics manufacturer's professional-line brand was named in none of 10 "top hair-color brands" answers in one local language, but in 3 of 4 salon-oriented answers in English.
Phase 2: five models, 90 prompts, April 28, 2026
We extended the same question to five models — Perplexity Sonar 2, GPT-5.4, Gemini 3.1, Claude Sonnet 4.6, Kimi 2.6 — 90 prompts (30 per language) across 11 query types, checked against 18 of the client's brands. Counting runs on a script against the 450 raw responses, not on a model's self-report, so the same numbers come back on a rerun.
- The development's location name is the most visible asset measured: 15.6% to 27.8% of prompts named it, depending on model. The corporate developer brand itself: 1.1% to 5.6%.
- We also measured model discipline: how often a brand is named when it should be (recall) versus named when it shouldn't (false positive). Claude Sonnet 4.6 and Gemini 3.1 held false positives at 5.0%. Perplexity Sonar 2 recalled more (47.6%) but at a 23.3% false-positive rate — close to a quarter of its apparent visibility was noise.
- Visibility fell by language for the top five cosmetics brands, in every model tested: queries in the two local languages averaged 7.3%-14.7% and 5.3%-12.7% mentions, English 3.3%-6.7%.
- By query type, "who makes this" attribution questions averaged 40.0% visibility across the cosmetics brands. B2B/wholesale questions averaged 0.4%. Everyday consumer questions and one nail-care sub-category both came back at 0.0% in all five models.
- For the developer, the location name averaged 20.0% visibility across models against 3.3% for the corporate name, and two sister developments under the same company averaged under 1%.
What we found
Attribution works when asked directly — models correctly name the manufacturer when a prompt asks "who makes X." That signal does not carry over into category or awareness questions, where the same models default to brands already dense in their training data. The gap is not content volume; it is which brands the model already associates with the category.
What we delivered
A prioritized roadmap: near-term fixes (structured data and about-page attribution to convert existing unattributed citations into named-brand credit, category-listing fixes for the near-zero segments) and a longer-term track (Wikidata/Wikipedia entries, category-specific content, a repeat measurement on the same 90 prompts at 60 and 120 days to track any brand's movement above 3 percentage points).
Status
This engagement was baseline measurement and roadmap delivery. We do not have post-implementation remeasurement data to report yet.