Entity Recognition

1. What entity recognition is

Entity recognition is the layer that decides whether an AI engine can tell who or what you are — and join the signals you have earned to the right node. This entry treats it as the GEO-relevant subset of the academic named-entity recognition / entity linking literature, not as the literature itself.

Definition (GEO Wiki working definition): entity recognition, as a GEO signal, is the process by which an AI engine maps a surface form — a name, brand, product, or author string — to a canonical entity it already knows, disambiguating it from collisions, so that any credit attached to that string accrues to the right node.

2. Why an unresolved entity cannot be credited — the load-bearing honesty

This entry’s counterpart to Brand Mentions §5’s “a link with no mention is not a completed authority play” and Citability §2’s “necessary, not sufficient”: a mention, a citation, or a markup assertion only pays out if the engine resolves it to the right node. Resolution is the join, not a nicety. Being named or cited but unresolved is credit that leaks.

   mention / citation / sameAs assertion
            │
            ▼
   [ ENTITY RESOLUTION LAYER ]   ◄── this entry
     surface form → candidate → disambiguate → node
            │
   ┌────────┴─────────┐
   ▼                  ▼
 resolved        UNRESOLVED / MIS-RESOLVED
 → prior &       → credit dropped, or
   credit          attached to the wrong
   accrue to       entity (name collision)
   your node

Where it acts is the whole point. Resolution operates at the pre-retrieval parse and entity layer — the same layer Schema.org for AI §2 places markup at, upstream of the Answer Loop §3 grounding/selection gate that Citability and E-E-A-T govern. Resolution does not pick the passage; it decides whose ledger the credit lands in.

Three orthogonality lines, stated so they cannot be misread:

• Resolution ≠ liftability. Making a passage quotable is Citability’s.
• Resolution ≠ trust. Whether the resolved entity is trusted is E-E-A-T’s.
• Resolution ≠ the node existing. Having a Wikidata node is Knowledge Graph Presence’s; being matched to it is here.

3. The mechanism — the resolution pipeline

Under the hood this is named-entity recognition followed by entity linking — named once as the underlying discipline, then never jargon-walked again. In GEO terms it is one pipeline:

  surface form        "Acme"  (the string on a page or in a query)
       │
       ▼
  candidate generation  which known entities could "Acme" be?
       │                (Acme Corp · Acme Tools · Acme the band …)
       ▼
  disambiguation        context + co-occurrence + prior pick one
       │
       ▼
  canonical entity      the node credit attaches to

Disambiguation draws on three inputs. They are not equal, and which one dominates is what §5 routes by surface.

Input ① is the mechanism Schema.org for AI §4 routes here: sameAs is the resolution join key — Google states it is “used behind the scenes to disambiguate your organization from other organizations” (see Organization structured data). Which markup carries the sameAs key is in Schema.org for AI §4; how the join resolves is below. For the JSON-LD block, see Schema.org for AI or the Schema Implementation playbook. In prose, sameAs simply asserts “the entity at this page is the one at this Wikipedia/Wikidata URL,” and the engine treats that as a candidate-collapsing edge only when the rest of the web does not contradict it.

4. The levers — what makes you resolvable (concept, not runbook)

A concept-level taxonomy of resolution levers, each tagged with the §3 input it feeds. This is not a runbook — the doing (deploy markup, reconcile identity, claim nodes) is the Schema Implementation playbook’s.

The line that ties it back, the mirror of Brand Mentions §6’s “you earn being the named thing”: you do not declare an identity — you make one corroborated identity the only consistent reading across the web. Resolution is downstream of that consistency. Trust framing of the identity is E-E-A-T’s axis.

5. How resolution varies by surface (invariant vs delta)

The §3 pipeline is invariant — it holds everywhere. What varies is which input dominates.

One routed line, not expanded: the same brand resolving across languages (zh ↔ en) is not re-derived here — cross-language entity binding is Multilingual GEO’s.

6. What the evidence says — and what it does not

The mechanism direction is well-attested; the brand-level dose-response is not. Read this table the way the site reads Aggarwal — for direction, not a coefficient.

Honest gap: Generative Engine Optimization’s headline lever — Aggarwal et al. (KDD ‘24, arXiv:2311.09735; paper summary) — measured on-page content rewrites (cite sources, add statistics, quotations), not entity resolution. The up-to-40% figure does not transfer to “improve entity recognition.” Borrowing it here would be the exact over-claim the siblings warn against.

The position, the reciprocal of Schema.org for AI’s “declared, not rewarded” and E-E-A-T §6’s “earned, not annotated”: resolution is corroborated, not asserted — one sameAs does not resolve you; one consistent identity attested across sources does.

7. Anti-patterns — identity ambiguity and false joins

The errors this entry exists to prevent — mirror of Brand Mentions §8 and Schema.org for AI §7.

The load-bearing line: the failure mode is rarely “no identity” — it is the wrong one, or a split one. Consistency is the fix, not volume.

8. Why this matters for GEO + how to act

Resolution is one of the upstream joins that makes any downstream credit reachable at all — pair it with groundability (Citability) and the off-site prior (Brand Mentions). This entry is the concept; the doing is the playbook.

References

Academic:

• Kandpal, N., Deng, H., Roberts, A., Wallace, E. & Raffel, C. (2023). Large Language Models Struggle to Learn Long-Tail Knowledge. ICML 2023 (PMLR v202). arXiv:2211.08411
• Mallen, A., Asai, A., Zhong, V., Das, R., Khashabi, D. & Hajishirzi, H. (2023). When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories. ACL 2023. ACL Anthology · arXiv:2212.10511
• Aggarwal, P. et al. (2024). GEO: Generative Engine Optimization. KDD ‘24. arXiv:2311.09735 · paper summary — boundary reference; entity resolution is not a tested variable

Official:

• Google — Introducing the Knowledge Graph: things, not strings (2012-05-16) — the “entities, not strings” model
• Google Search Central — Organization structured data — sameAs “used behind the scenes to disambiguate your organization from other organizations”
• Schema.org — sameAs — “URL of a reference Web page that unambiguously indicates the item’s identity”

Industry:

• Search Engine Land — Why entity authority is the foundation of AI search visibility (Benu Aggarwal, 2026-03-16)