Entity optimization: the complete guide

An entity is a thing in the world with a stable identity: a business, a person, a place, a product, a credential. Search engines and AI assistants think in entities, not strings. If your business is recognized as a clean, well-connected entity, you get recommended, cited, and disambiguated correctly across every retrieval surface that matters in 2026. If it is not, you compete keyword-by-keyword on a shrinking SERP. This is the complete playbook for entity optimization as it actually works in May 2026, across both traditional search and AI environments.

What an entity actually is

In information retrieval, an entity is a disambiguated, structured identity for a thing in the world. Each entity has:

• An identifier: a Knowledge Graph machine-readable ID, a Wikidata Q-number, a Companies House number, an SRA roll number, an ORCID. Stable; survives renames.
• A type: LocalBusiness, Person, Organization, Product, MedicalClinic. Constrains which properties apply.
• Properties: name, address, opening hours, credentials, area served, founded date, sameAs links to other identifiers for the same entity.
• Relationships: edges to other entities. A solicitor works for a firm; a firm operates in a jurisdiction; a clinic is contained in a city; a product is offered by a business.

Google formalised this view of search in 2012 when it launched the Knowledge Graph. The corporate-blog explainer states the system holds information about roughly 5 billion entities and over 500 billion facts, with local-business records populated through Google Business Profile claims rather than Wikipedia-style notability. Wikidata, the open sibling, holds well over 100 million items with a much lower notability bar — most established businesses with independent press coverage qualify, even small independents.

Entity SEO is not better keyword SEO. The unit of work is the entity record, not the page. Once your entity is correctly resolved, disambiguated, and connected, every page on your domain inherits that identity in retrieval. Without it, every page competes alone.

How retrieval has changed: 2024 advice no longer survives

Most of the entity-SEO advice still circulating online was written before three structural shifts. Each one changes what you optimize for.

The entity stack: where your identity actually lives

For a UK local business, "your entity" is not one record but a graph of records that need to converge on the same canonical identity. Search and AI extractors use this convergence as a confidence signal. The more authoritative sources agree on the same entity, the higher the retrieval confidence.

When all six layers reference the same canonical entity, retrieval systems converge with high confidence. When they disagree (different names, addresses, missing nodes, broken sameAs), the system either picks one and gets it partially right, or fails to surface you at all. Entity convergence is the goal.

Entity salience: the public proxy for "how strongly does Google associate this entity with this page?"

Google's Cloud Natural Language API still publicly exposes a 0-1 salience score for every entity it extracts from a document. The same NLP family underpins much of Google's organic ranking and AI surface logic. Salience is not "the ranking signal" — but it is the most direct public proxy practitioners have for how Google's NLP weights an entity within a page.

Practical use:

• Run your service pages through the NL API. The intended primary entity for the page should be the highest-salience entity returned. If it is not, your page is semantically about something else.
• Look at adjacent high-salience entities. They reveal the topic neighborhood Google is reading the page in. Useful for spotting drift (a "boiler service" page where "kitchen renovation" is the top adjacent entity is in trouble).
• Watch for entities that appear in the page text but never extract. Almost always a content-structure or schema problem. Add internal links, headings, or a structured-data block that names the entity explicitly.

Schema.org as the structured handoff to LLM extractors

Schema.org released v30.0 in March 2026, adding new types (Credential, ConferenceEvent, OnlineMarketplace, InstantaneousEvent) and equivalence annotations to GS1, Dublin Core, and Open Graph. Two facts shape how to use it in 2026:

1. Schema is now the primary structured handoff to LLM extractors, not just a rich-results lever. ChatGPT, Claude, Gemini, and Perplexity read your JSON-LD when they ground answers, even when no SERP rich result is on offer.
2. Google deprecated rich-result eligibility for several types from January 2026 (Q&A, Practice Problem, Sitelinks Search Box, SpecialAnnouncement). The schema types themselves are not deprecated — and AI extractors continue to read them. Decide on a per-type basis whether the maintenance cost is justified by AI-extraction value.

The load-bearing types for a local business

The sameAs property is doing more work than you realize

sameAs declares "treat all of these URLs as references to the same entity". Modern AI extractors weight authoritative identifiers (Wikidata, Wikipedia, regulator pages, Companies House) far more heavily than another link to your own LinkedIn. The 2024 advice to "stuff every social profile in" is now a missed-opportunity pattern.

A high-quality sameAs array for a UK local business looks like:

• Wikidata Q-number URL
• Your Google Business Profile URL
• Corporations Canada or provincial registry page
• Regulator or professional-body record: provincial law society, College of Physicians and Surgeons, CPA Canada, OSC, IIROC
• One or two top-tier press mentions
• Two to three primary social profiles

Three patterns to avoid: paid sameAs link networks designed to inflate entity confidence (extractors weight by destination authority, so most contribute nothing); broken or 404-returning sameAs targets (read as active disconfirmation); inconsistent name or address strings between the targets (forces the extractor to pick a winner and may pick the wrong one).

Wikidata: the open Knowledge Graph

Wikidata sits below Wikipedia on the notability ladder. Per the official policy, an item must satisfy at least one of three criteria: a valid Wikimedia sitelink; a clearly identifiable conceptual or material entity with serious public references; or a structural need to support other items. A business registry listing alone is not enough — you need independent reliable-source coverage to pass the second criterion. Most established businesses with a few years of trading and any press clear that bar.

1. 1

   Search before you create

   Wikidata may already have an item for your business — created by a Wikipedia editor, an industry catalog, or a previous merger record. Search by name, by domain, and by any prior trading name. If an item exists, the work is to enrich and correct it, not to create a duplicate.

2. 2

   Establish notability before you create

   Independent reliable-source coverage from at least two distinct sources is the practical threshold. Local press, sector trade press, regulator publications, and academic citations all count. Press releases, your own marketing, and client testimonials do not. If you do not pass the bar, work on earning the coverage first.

3. 3

   Create or commission the item

   Editing your own entry is technically allowed but discouraged. Better is to provide a Wikidata-active editor with sourced facts and let them create the entry. The Wikidata community has an active 2025-26 RfC on notability reform; expect minor policy shifts but not radical ones.

4. 4

   Populate the load-bearing properties

   P31 instance-of, P17 country, P159 headquarters location, P856 official website, P1448 official name, P1454 legal form. For UK businesses, P1320 OpenCorporates ID and identifiers for Companies House, the FCA register, or relevant trade bodies close the entity-disambiguation loop.

5. 5

   Close the loop on your domain

   Add the resulting Q-number URL to your Organization JSON-LD's sameAs array. This is what tells extractors that your domain and the Wikidata entry describe the same entity. Without that loop closed, the Wikidata entry is half-useful at most.

Authoritative co-occurrence: not dead, but reframed

The classic "co-citation" idea was that if your business name appears alongside a competitor's name on enough authoritative pages, search engines learn to associate you with the same category. That is still true, but the threshold for "authoritative" has tightened and the threshold for "manipulative" has loosened in enforcement.

Two things have changed materially:

• Site reputation abuse. Google introduced this as a spam category in March 2024 (manual actions only) and moved to algorithmic enforcement in the August 2025 spam update. Hosting third-party commercial content under a strong site's authority — paid placements, subfolder rentals, "loan content" arrangements — is now actively dangerous, not just risky. The European Commission opened a DMA investigation into the policy's application to news publishers in November 2025.
• Trust filtering at retrieval. AI assistants weight destination authority during retrieval, not just at training. Manufactured co-occurrence on low-quality directories contributes nothing to AI citation rates and correlates with manual-action triggers.

What still works: earning genuine mentions in recognized press, sector publications, regulator and professional-body listings, and curated "best of" round-ups. Each authoritative co-mention is both a citation-source for AI answers and a co-occurrence signal for ranking. See the AI search visibility guide for the earned-media playbook in detail.

Author and Person entities: where YMYL gets hardest

Your Money or Your Life queries — medical, legal, financial, safety, and now (per the September 2025 Search Quality Rater Guidelines update) government and civic information — are handled more conservatively by every major retrieval surface. Models default to authoritative sources, surface caveats more readily, and decline to recommend specific providers more often than in lower-risk sectors.

Two 2025-26 shifts shape the response:

• On 1 February 2026, Google added a dedicated "Authors" section to Search Central documentation, codifying author-entity transparency as an explicit quality signal.
• The December 2025 core update was the most significant YMYL-focused update since the 2018 Medic update; legal verticals were among the hardest hit. The March 2026 core update further amplified the "Experience" leg of E-E-A-T, rewarding first-hand specifics and verifiable author credentials.

How each major surface treats entities, in May 2026

Different retrieval stacks are wired differently. Knowing which stack each one uses tells you which signals you need to invest in for each surface.

Crawler access: the audit nobody runs

AI crawlers honor robots.txt. Many UK sites accidentally block them via wildcard rules, agency-installed defaults, or templates inherited from years ago when AI crawlers did not exist. The list is longer than most operators expect:

llms.txt: a candid status report

llms.txt was proposed by Jeremy Howard in September 2024 as a Markdown file at /llms.txt describing a site's content for LLM consumers. As of May 2026: no major LLM operator (Google, OpenAI, Anthropic, Microsoft, Apple) has officially endorsed it for ranking or retrieval. Adoption hovers around 10% on large-domain samples; independent statistical analyzes to date have found no measurable effect on LLM citation frequency. Implement it if your stack makes it trivial — a clean structured summary of your content costs little — but do not make it a focal investment.

Knowledge Graph eligibility for local businesses

Most local-business operators conflate the Knowledge Graph (the underlying data store of around 5 billion entities) with the Knowledge Panel (one of many surfaces that render data from it). A panel is not proof the graph trusts you; absence of a panel is not proof of exclusion. The reliable path:

Sector heuristics for entity work

Patterns observed across UK sectors for what entity signals dominate. Useful for prioritisation, not as a substitute for sector-specific research.

What to avoid: 2026 penalty patterns and outdated tactics

The 90-day entity-optimization program

A practical phased plan for a UK local business. Each phase covers both the entity-record work and the schema/page work that anchors it on your domain.

1. 1

   Days 1 to 30: foundations and audit

   Audit robots.txt explicitly for each AI crawler (GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, Claude-User, Claude-SearchBot, PerplexityBot, Perplexity-User, Google-Extended, Applebot-Extended, CCBot). Validate every JSON-LD block on the site through the Rich Results Test and Schema.org's validator. Confirm GBP, Bing Places, and Apple Business Connect all match. Inventory existing sameAs targets and replace any social-profile-heavy chain with one anchored in regulator records, Companies House, and authoritative identifiers.

2. 2

   Days 31 to 60: entity convergence

   Create or claim a Wikidata Q-number where notability allows. Add the Q-number URL to your Organization JSON-LD's sameAs. Implement the load-bearing schema types: LocalBusiness with the most specific subtype, Organization with sameAs, Service per offered service, Person for any named author or expert, Review and AggregateRating from first-party data. For YMYL businesses, add hasCredential references to regulator records. Aim for one or two earned mentions in authoritative third-party sources by end of phase.

3. 3

   Days 61 to 90: measure and iterate

   Run a structured query test set across Search, AI Overviews, AI Mode, ChatGPT Search, Perplexity, Claude, Microsoft Copilot, and (if relevant) Apple Intelligence. Measure entity inclusion rate, page citation rate, earned-media citation rate, and competitive set per surface. Use the Bing Webmaster AI Performance preview for Copilot citation visibility. Identify gaps (where competitors appear and you do not, on which surface, for which query type) and target the underlying signal. Re-test monthly.

Reference numbers

Where this is going

Entity optimization is becoming the connective tissue of every modern retrieval surface. Search engines, AI assistants, and the graph-RAG architectures that increasingly power them all converge on the same question: which entity is this, and how confident are we? Businesses that close the loop between GBP, Wikidata, regulator records, schema on their domain, and authoritative third-party coverage become first-class citizens of that retrieval substrate. Businesses that do not get partial-credit answers when the system guesses, or no answer at all when it cannot.

The good news: the work is concrete and finite. A clean entity record is not infinite content investment. It is a one-time foundational build with quarterly maintenance. The teams that ship it now compound the advantage every time a new AI surface launches and reads from the same structured layer.

Where to go next