The Brand Entity Graph: One Story Across Site, Wikidata, LinkedIn, and Marketplaces

A composite case: a DTC espresso-machine brand sells direct in the UK and through a German distributor. A buyer in Manchester asks an assistant whether the brand still services a machine bought in 2023. The answer names a discontinued model as current, calls the company a UK reseller, and links a domain that now redirects to a clearance page. Support gets a screenshot. No product page changed. The distributor’s retired site simply reads more like a company record than the brand’s About page.

An entity graph is not a Wikidata entry. It is one set of fields saying the same thing on your site, LinkedIn, marketplace storefronts, and Wikidata — and when they conflict, a model does not pick the true version; it picks the most record-like one.

Models resolve fields, not pages

Asked who a brand is, a model resolves the entity, then extracts attributes. It compares fields, not prose, and has no prior that yours is the official site — only two tests: can the claim be restated elsewhere, and which source looks like a record? That is why a distributor page can beat an About page: “We love coffee” is not a usable field; “warranty handled by the manufacturer” is.

The second mechanism is a stale clock. The site changes with each release; LinkedIn sits for years; a marketplace brand field waits on review; Wikidata is edited by strangers. Change one record and three keep the old version. Faced with contradictions, the cheapest response is to hedge, or quote whichever third-party page has the most complete fields.

So “create the Wikidata entry first” is wrong. Notability thresholds apply, and you can only file edit requests with sources; a brand without independent coverage is more often deleted than accepted, and the deletion is public. You fully control site and LinkedIn, partly control marketplace fields, and merely influence Wikidata. The rest of this line sits in the authority and off-site hub.

Four records, four levels of control

Equal effort is wasted: the records do not offer equal control.

Record	Who maintains	Model reads	Common defect	Control
Brand site	Brand, front-end	Organization data, About, policies, domain	Market suffix; role unclear	Full; same day
LinkedIn page	Brand operations	Name, industry, size, HQ, website	Retail industry; reseller blurb	Full, if remembered
Marketplace storefront	E-commerce	Brand field, category, model names, returns	Supplier in brand field	Partial; review
Wikidata	Volunteer editors	Entity type, country, website, inception	Forum link; namesake merge	Edit requests only

LinkedIn returns most: structured, editable only by you, and most often abandoned. Wikidata has the highest ceiling, since it drives namesake disambiguation, and the lowest control.

Six fields that change the answer

Most mismatches are cosmetic. Only six change what an answer means.

Field	Consistent looks like	How it fails
Entity and role	One role across legal and LinkedIn copy	Cast as a reseller, behind the channel
Official domain	One primary domain; regions redirect or labeled	Citations land on an old reseller site
Brand and legal name	One primary name; legal only on invoices	The brand splits, and so does its evidence
Model naming	“Aero 2 Pro”, never “Aero2Pro”	One product becomes two
Market boundary	Shipping, warranty, support agree	“Covered in Germany?” comes back wrong
Disambiguation	Two of: founding year, HQ, registration, seller ID	Namesake merge; you inherit its reviews

The first row is not a content task: manufacturing versus reselling, warranty service, and direct versus partner-led markets must be settled internally. The last looks like housekeeping, but a namesake merge is not neutral — the other company’s complaints join your record.

Name and model spelling also have to match platform category language; see building a category tree AI can parse.

A four-step pass with an owner

Step	Do	Owner	Passes when
1 Evidence	Log ten brand-name answers: field, value, source	Growth	A table of the model’s view
2 Standard	Approve one value and variants for six fields	Brand lead	Three teams copy it without asking
3 Align	Site, LinkedIn, marketplace, Wikidata; one field class at a time	Front-end, ops, PR	Same questions re-run same day
4 Cadence	Quarterly check; rename, domain, series, market triggers	Brand lead	Named owner and calendar

Change one field class, re-test, then move on; otherwise you cannot say which edit moved the answer. Triggers matter more than the rhythm: after a rename the records update months apart while the answer cites the old one. Re-test with the prompt set from the AI visibility baseline.

Example: three months after a rename

Example — a hypothetical espresso brand, no real company or figures. It renamed its flagship last year. The site updated on launch day; LinkedIn kept the old positioning for three months; the marketplace brand field sat in review; Wikidata still pointed to a forum link.

A buyer asks in English whether the brand still supports the previous model. The answer reports a rename, then sends the buyer to the forum.

Four moves: keep a legacy comparison page that states “discontinued” in one sentence; replace the LinkedIn blurb with the same positioning and primary domain; unify model spelling across storefronts; file one Wikidata edit request citing the site and one article.

The re-test moves less than expected: the correct domain appears, the model stops splitting one product into two, and the discontinuation question stays unreliable. Alignment repairs what the model says wrong, not what it never knew. Product relationships are a separate graph; see the product graph. Confirm first that the entity is acknowledged: does your brand entity exist?.

Where this breaks

Dropship and white-label sellers. Three records are not yours; fix listings, reviews, fulfillment first.

Brands too new for independent sources. Notability rules will reject or delete the item. Do site and LinkedIn now, Wikidata later.

Unresolved multi-entity structure. Parent, sub-brand, regional agent, local entity — aligning before those relationships are documented copies the contradiction four times.

Sales almost entirely inside a marketplace. Platform AI reads listings, price, and delivery; storefront fields outrank Wikidata.

Consistency read as a recommendation guarantee. Agreement makes the model accurate, not favorable; price and third-party evidence still decide.

A site AI cannot crawl or render. Fix access first; aligning an unreadable page changes nothing.

This month, build the six-row table with three columns — site, LinkedIn, primary marketplace — mark every conflict in red, fix them in order, and file nothing on Wikidata except one edit request with two sources. That table becomes next quarter’s reconciliation sheet.