brand-profile agent resolves the WRONG Wikidata entity for same-named brands (not fixed by #359)

[nsdere]

Summary

The brand-profile agent resolves the wrong Wikidata entity for brands whose name collides with an unrelated same-named entity, because findWikidataId searches on the brand name alone and ignores the strongest disambiguator we already have: the site domain (e.g. the TLD) and the scraped site content.

PR #359 (#359) anchors the Wikipedia article fetch to the resolved QID and guards on wikibase_item === QID. That fixes the case where the QID is correct but the article lookup went astray. It does not fix the case where the resolved QID itself is wrong — the agent then faithfully anchors to the wrong entity and still emits contaminated products / sub_brands / competitors.

This is the sibling of the data-quality report adobe-rnd/llmo-data-retrieval-service#3200.

Two distinct failure classes

Class	Cause	Example	Fixed by #359?
A — QID correct, article wrong	old step grabbed a same-named article despite a correct QID	Lovesac → Q6690181 (LoveSac, furniture) ✅	Yes — regen now returns furniture
B — QID wrong	entity resolution picked a same-named unrelated entity	Capella → Q12970 (the star Capella) ❌	No — regen still returns star data

Reproduction (Class B — verified on prod after #359)

capella.edu is Capella University (an online university; .edu). Its resolved QID is Q12970, which is the star Capella (constellation Auriga), not the university.

Before #359 (name-search article): products were a children's TV series ("Cappelli & Company", songs, cassettes).
After #359 regen: products became stars — Capella Aa, Capella Ab, Capella H (category "Star (primary giant in Capella system)"), with products_metadata.brand_wikidata_id: "Q12970" and wikipedia_verified: true. Still wrong — the guard passed because the article does match the (wrong) QID.

Root cause

src/agents/brand-profile/services/wikipedia.js → findWikidataId(brandName):

• calls wbsearchentities with search: brandName,
• returns the first result whose description contains a company term, else the first result.

There is no use of the site domain or scraped content to disambiguate. capella.edu (a .edu) is an unambiguous "this is a university, not a star" signal that is discarded.

Prevalence (fleet scan)

Of 851 brands that have a stored products QID, ~247 have a stored QID whose Wikidata entity does not match the brand (Class B; upper bound — includes some acronym false-positives such as AIB→"Allied Irish Banks", which is actually correct). Clear examples:

• AIA (insurance) → Q25228 = Anguilla (a country)
• American Airlines → Q838581 = "comitative case" (grammar)
• AAA Membership → Q1823 = Aceh
• Capella University → Q12970 = the star Capella
• OECD → polycrystalline silicon; Borussia Dortmund → BYD; Logista → Lovisa jewellery

Blast radius

Same as #3200: products.items[*].category is the priority‑1 category source for synthetic-persona prompt generation, and the profile also feeds Semrush, Citation Attempt / Strategy Chat, and GSC. A wrong entity contaminates all of them.

Suggested fixes

1. Disambiguate with the site domain + scraped content. Pass the base URL (esp. the registrable domain / TLD) and the industry/vertical into findWikidataId; prefer the candidate whose Wikidata sitelink/official-website (P856) or description matches the domain/industry. A .edu should never resolve to a star.
2. Post-resolution guard. After picking a QID, verify its Wikidata description/instance of (P31) is consistent with the brand's known vertical; reject on clear mismatch.
3. Fail safe on low confidence. When no candidate matches the domain/industry, return no data rather than a wrong entity (consistent with fix(brand-profile): anchor Wikipedia fallback to the resolved Wikidata QID #359's "no data rather than wrong data" principle).

Links

• Original data-quality report: adobe-rnd/llmo-data-retrieval-service#3200
• Partial fix (Class A): fix(brand-profile): anchor Wikipedia fallback to the resolved Wikidata QID #359

[nsdere]

Fix proposed in #361 — adds an entity-validation gate (official-website P856 domain match + LLM content check against site domain/industry) after QID resolution, discarding wrong same-named entities (no data rather than wrong data). Verified with tests incl. the Capella Q12970 (star) case.