AI engines don’t read your brand name — they resolve it to a node in a graph, and if a same-named competitor, book, or public figure has a stronger node, your name gets attached to their facts. “Entity resolution” is the unglamorous plumbing behind every AI answer about a company, and it runs on three separate systems — Google’s Knowledge Graph, Wikidata, and whatever an LLM absorbed during pretraining — that don’t always agree on who you are.
What “entity resolution” actually means
When someone asks an AI assistant about a company, the model (or the search layer feeding it) doesn’t search for a string. It tries to map that string to a specific real-world entity — a node with attributes: founding date, industry, people, official site. This is the same problem search engines have solved for over a decade, and Google’s own documentation describes the mechanism plainly: Organization structured data properties like iso6523 and naics exist specifically “to disambiguate your organization from other organizations,” while properties like logo and url only affect how the entity is displayed once it’s already been matched (Google Search Central, Organization structured data, last updated April 15, 2026).
The separation matters. A brand can have a beautiful, complete Organization schema block and still get resolved to the wrong entity, because disambiguation and display are handled by different fields, and only a handful of them do the identity work.
Three systems, three different identity anchors
None of these systems share a single source of truth. Each one anchors identity differently, and each is queried differently by AI products layered on top of it.
| System | Identity anchor | What it returns | Primary source |
|---|---|---|---|
| Google Knowledge Graph Search API | Internal machine ID (e.g. kg:/m/0dl567) |
JSON-LD entity with a resultScore confidence value, capped at 500 results per query |
Google Knowledge Graph Search API reference |
| Organization structured data (on-page) | sameAs, iso6523Code, naics, vatID |
No required properties — Google explicitly recommends adding “as many properties that apply,” with legal/tax IDs doing the actual disambiguation | Google Search Central |
| Wikidata | Q-number (e.g. Q42) |
A structured item accepted only if it clears one of three notability criteria | Wikidata:Notability policy |
| LLM pretraining / retrieval layer | None formally — statistical association from training text and retrieved documents | Whichever entity has the densest, most consistent signal in the corpus or retrieved set | No official spec; behavior inferred from vendor documentation and independent testing |
The schema.org property doing the actual work
sameAs is the closest thing entity resolution has to a standard. Schema.org defines it narrowly: “URL of a reference Web page that unambiguously indicates the item’s identity. E.g. the URL of the item’s Wikipedia page, Wikidata entry, or official website” (schema.org/sameAs). It’s also one of the more heavily used properties on the web — schema.org’s own usage-statistics dataset, built jointly with Google, puts sameAs in the 10-million-plus-domain bucket as of July 2026. Wide adoption doesn’t mean correct adoption, though: the property is a single URL or list of URLs, and Google’s guidance is that it should point only to pages that “genuinely represent the same person, organization, or thing” — not every social profile a brand happens to hold.
For the properties to matter, the target pages have to be crawlable in the first place. If your sameAs targets or your own Organization markup live behind client-side rendering that a crawler never executes, none of this reaches the graph — see our notes on why JavaScript-rendered content is invisible to AI, and on which bot rules in robots.txt actually govern AI crawlers in the first place.
Where it breaks: entity collisions
Entity resolution fails loudest when two things share a name. Independent GEO research aggregating AI-accuracy studies through mid-2026 found that small businesses have their brand name confused with another company in AI answers roughly five times more often than large companies — 4% of answers about SMEs versus 0.7% for large companies, per a Searchable accuracy study covering more than 13,000 queries across ChatGPT, Perplexity, and Gemini (cited in Stanislav Peev’s 2026 AI-accuracy research roundup, sourcing Searchable’s July 2026 study via SME Magazine). The mechanism is straightforward and not really “hallucination” in the classic sense: the model or retrieval layer correctly retrieves facts, just about the wrong entity, because the competing entity has a denser signal footprint. That’s a distinct failure mode from the ones catalogued in our Limitations Ledger — it’s a resolution error, not a generation error.
The same study found the gap widens with scale: large companies had a fabricated fact in 32% of answers versus 50% for small businesses, and the researcher’s own caveat applies here too — the underlying sample is UK-based, and no comparable large-scale US study exists yet, so treat the exact percentages as directional rather than definitive.
Wikidata’s gate, and why B2B brands stall at it
Wikidata matters because several major consumer AI products and search knowledge panels lean on it as a cross-referenced identity layer, but getting an entity into Wikidata requires clearing one of three explicit criteria. Per the current policy text, an item is accepted if it has “at least one valid sitelink” to Wikipedia or a sister project, or if it “refers to an instance of a clearly identifiable conceptual or material entity that can be described using serious and publicly available references,” or if it “fulfills a structural need” for other items (Wikidata:Notability). For most B2B software companies, criterion two is the realistic path — a funding announcement, a regulatory filing, or trade-press coverage substitutes for a Wikipedia page — but “serious and publicly available references” is a judgment call enforced by volunteer editors, and as of this writing the policy itself is under an open request for comment on reform. That’s worth flagging as unresolved rather than settled: the gatekeeping criteria brands are optimizing against are themselves in flux.
None of this replaces the more basic step covered in our guide to how AI assistants decide which brands to recommend — entity resolution determines whether the model knows who you are at all; recommendation is a separate, later decision about whether it likes what it finds. And structured markup alone doesn’t solve resolution if the crawler can’t reach it or a stronger competing signal exists: pairing entity work with the retrieval mechanics described in llms.txt explained and schema markup that AI engines actually read covers more of the pipeline than any single fix.
FAQ
Does adding sameAs links guarantee an AI engine resolves my brand correctly?
No. sameAs is a signal, not a guarantee. Google’s own documentation frames properties like sameAs as inputs that help disambiguation “if applicable,” alongside legal identifiers like iso6523Code and vatID — it’s the combination and consistency across sources, not any single property, that determines resolution confidence.
Why does a small company get confused with a bigger one that has a similar name more often than the reverse?
Because resolution defaults to the entity with the denser, more consistent signal footprint across the sources a model or retrieval layer draws on. A 2026 accuracy study found this happens roughly five times more often to small businesses than large ones (4% vs. 0.7% of answers), which tracks with the general pattern that AI systems resolve toward whichever version of a name has more corroborating data.
Is a Wikidata entry required to be recognized as an entity by AI systems?
No single source is required — the Google Knowledge Graph, on-page Organization schema, and LLM training or retrieval signals are all separate inputs. But Wikidata is one of the few structured, cross-referenced identity layers multiple systems can point to, so lacking one removes an anchor other signals could otherwise reinforce.
Last updated August 12, 2026. This page is refreshed as benchmarks and scores move.