GEO-092·Issue 08·Technical Foundation·92 / 132
Knowledge Graph Basics: Wikidata, sameAs, and Entity Disambiguation
An assistant picks the wrong company with your name before it reads your copy. Eight sourced fields, a backlink test for sameAs, and a split-first order.
Issue focusOff-Site Authority: Digital PR, Communities, and the Knowledge Graph
A US brand selling under-sink water filters into Germany gets a screenshot from Munich. She asked an assistant whether the filters ship to Germany; the answer describes a Stuttgart water-treatment contractor, with its registered office, founding year, and service hotline. Only the photo is the brand’s. The copy is fine. A volunteer editor had registered a knowledge-base item; another merged it into the contractor’s item, both names matching and both saying water. The markup still points at the identifier from before the merge.
A knowledge graph entry is not a privilege for large companies; it is a disambiguation file — eight sourced fields, sameAs values that point back, a split-first order — and leaving one part out files you under another company.
Identity resolves before attributes
Asked who a brand is, a model decides which entity you are, then pulls attributes. Failure at the first step is not a gap but a confident error, sharpened by every field you add.
Three cues carry the resolution. The name string is weakest: same-name firms, a brand name equal to a store name, language variants on separate items. Co-occurrence is stronger: category words, domain suffix, address. Graph edges are strongest: whether the site, marketplace store, trade coverage, and reviews point at one identifier.
A knowledge base does not rank you; it gives scattered pages one place to hang, so ten documents become ten facts about one company, not ten candidates. Small brands invert the big-brand approach: fewer fields, fewest but hardest links.
Eight fields, each with a source
Eight fields decide whether you are a different legal subject; the rest is description.
| Field | Write | Source | If missing |
|---|---|---|---|
| Entity type | Brand, manufacturer, or retail brand — one | About page plus independent article | Filed with resellers |
| Official name | One primary; legal name where legal | Filing or terms page | Split into two items |
| Official domain | One primary; regional stores state the link | The site | Citations on a reseller site |
| Founding year | Four digits, “founded” or “launched” | Filing or coverage | No ruling on a collision |
| HQ | City and country, never “global” | Filing or site | Absorbed into a namesake |
| Industry | Category words, not a slogan | Marketplace category, trade press | Wrong consideration set |
| Product lines | Series spelled as on product pages | Site category pages | A model number read as a brand |
| Anchors | Registration number, store ID, city plus year | Page anyone can open | Nothing to arbitrate |
A source is a page a stranger can open, not a fact you hold. Knowledge bases are community-edited with inclusion thresholds (as of 2026-08), so attach the site and one independent article: the About page proves what you claim, the article what others recorded.
sameAs is a claim until the target points back
sameAs sits in Organization structured data, linking the site to LinkedIn, a knowledge-base item, a marketplace store, a review page. The model treats each value as a pending link: it merges only when the target is fetchable, the subject matches, and ideally the target links back. A one-way declaration is ignored on conflict.
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Example Brand",
"url": "https://example.com/",
"sameAs": [
"https://www.wikidata.org/entity/Q000000",
"https://www.linkedin.com/company/example"
]
}A validator reads syntax; the model reads whether the same subject sits behind those addresses. Three rules follow. Include only endpoints that can point back; a press release usually cannot. Canonicalize: https, no tracking parameters, an entity page rather than a search result. Keep five or fewer, each explainable in one sentence; fifteen is noise.
Five steps: split the collision first
Fix what makes you another company before descriptive fields; reversed, you add detail to the wrong node.
| Step | Do | Owner | Passes when |
|---|---|---|---|
| 1 Dedupe | Brand, English name, category: knowledge base, search engine, two assistants | Growth | Each namesake’s country, industry, domain logged |
| 2 Define | One value per field, with source and as-of date | Brand lead | Three surfaces copy it without asking |
| 3 Split | Anchors into About page, markup, store profile; a merge reversed by split request | Brand, front end | Asked thrice, the question stops returning the other HQ |
| 4 Prune | Keep endpoints that point back; delete redirects, search pages | Front end | Each endpoint states the same subject |
| 5 Retest | Re-ask; log “wrong” and “thin” separately | Growth | Errors fall, no field contradicts another |
Step 3 can sit in community review for weeks; fix the site and store meanwhile to give the model a second citable source. Step 5 needs sampling, not a screenshot — use the cadence of an AI visibility baseline, since answers move by region, account, and day.
Example: one merged item, three edits
Example — the brand, domains, and identifiers are invented; no real company is described.
Brand Halden sells under-sink filter housings into Germany and collides with Halden Wassertechnik, an industrial water-treatment firm in Stuttgart. Ask whether Halden is a German company, and the answer merges them: Stuttgart HQ, industrial treatment, filters that will not ship. The About page says Delaware; the markup points at a QID that now redirects; the item’s website field still shows the old domain.
All three need editing. Open the About page with one self-contained disambiguating sentence — what the brand makes, for whom, where it is based, unrelated to the same-name firm — and add legalName, address, registration number, and disambiguatingDescription to the markup. Drop the redirecting QID from sameAs, and file an edit or split request with the site and one trade article.
Three weeks later the headquarters answer is correct, and “do the filters ship to Germany” quotes the shipping page. Disambiguation changes who you are merged with, not every attribute; product relations are another graph, covered in the Product Graph.
Where this breaks
White-label catalogs. The platform owns the brand fields; do listings, reviews, and fulfillment first, or the entry gets proposed for deletion.
No independent source yet. The entry is removed and leaves a record. Build the site, the store profile, and one asset only you could publish, like citable assets.
An unresolved corporate structure. Tangled parent brand, regional entities, distributors copy the contradiction four times.
The entry as a ranking lever. It only makes the machine describe you accurately; recommendation rests on price, fulfillment, and outside evidence, and padding a wrong node only describes the other company.
This issue’s lead, AI’s off-site trust network, shows why a recognizable entity decides whether that network has anywhere to land; the rest sits on the technical foundation hub.
This month: search the brand name, English name, and category in a knowledge base and two assistants, and log the nearest namesake’s country, industry, and domain. Then add one sentence to the About page that separates the two, and cut sameAs to endpoints that point back.