Entity optimisation: a practical guide for search and AI assistants

Entity optimisation is the work of making a business, a person or a product easy to identify. A search engine or an AI assistant that meets your name has to decide which real thing it refers to and which sources to trust for the facts about it. When your records agree, that decision is easy. When they disagree, the result is a vague answer, an old address, or a competitor named where you should have been.

This post is the practical version. It covers what an entity is to the systems that matter, the jobs that make yours unambiguous, how to check how you are understood, and what a paid service can and cannot deliver. The theory, the full schema.org patterns and the rules on Wikipedia are in the entities and structured data guide.

An entity is one distinct thing that facts can be attached to: a company, a person, a product, a place. Three systems shape how the word is used in search.

• Google's Knowledge Graph holds entities with identifiers. Its public Search API returns each match in JSON-LD using schema.org types, with an ID such as kg:/m/0dl567, a name, a list of types and a short description.¹
• Wikidata stores each entity as an item with a Q number, a label, a description and statements. Several items can share a label, so the description is there to tell them apart.² Wikidata also records Google Knowledge Graph IDs as a property of its items,³ which puts it close to the centre of this work.
• Schema.org is the vocabulary websites use to describe entities in structured data. Its sameAs property points to a page that unambiguously indicates the item's identity, such as its Wikipedia page, Wikidata entry or official website.⁴

ChatGPT, Claude and Perplexity offer no public lookup for their own picture of an entity. What they say about you comes from what the model learned in training and what it retrieves while answering. For most businesses the second matters far more.

Language models are weakest on the kind of entity most businesses are. In a study published at ACL 2023, Mallen and colleagues tested 10 models on PopQA, a set of 14,000 questions about entities of varying popularity. The models struggled with less popular facts, scaling them up did little for the long tail, and models given retrieved documents beat unassisted models orders of magnitude larger. Unassisted models stayed competitive on questions about highly popular entities.⁵ Kandpal and colleagues, at ICML the same year, found that a model's ability to answer a factual question tracks how many relevant documents it saw in training, and that retrieval reduces that dependence.⁶

A local firm, a niche manufacturer or an independent consultant is likely to appear in very few training documents. When an assistant answers a question about one, the answer rests mostly on what it retrieves at that moment. Perplexity sources content from the web in real time and cites it,⁷ Claude searches the web for topics that benefit from current information and cites its sources,⁸ and Google's AI features are rooted in its core ranking systems and highlight content from its Search index.⁹ Retrieval only works in your favour if the system can match the pages it finds to the right entity. Consistent names, descriptions and identifiers are what make that match possible.

What other sites say about you counts as well. Ahrefs compared 75,000 brands, all with an Ahrefs Domain Rating above 40, with their mentions in AI Overviews in a study published in May 2025. Branded web mentions had the strongest correlation of the factors it tested, 0.664, against 0.218 for the number of backlinks, and the author stressed that correlation is not causation.¹⁰ Bing's revised webmaster guidelines, as Search Engine Journal reported in February 2026, ask for entity names that are clear and consistent, with no ambiguous references.¹¹ Bing's own advice for AI answers is to align text, images and video so they consistently represent the same entities.¹² Brand signals in AI search covers the mentions side in more depth.

The jobs are the same for a business, a person and a product. Only the records differ. Do them roughly in this order, since each one is read against the ones before it.

One name and one description

Choose the exact name you want to be known by and use it everywhere, spelt and punctuated the same way each time. Then write one factual description of a sentence or two: what you are, what you do and where. Wikidata's descriptions are a good model, because their whole job is to tell one item apart from others with the same label.² For an invented firm, “Northfield Joinery, a joinery firm in Harrogate that makes fitted furniture” does that job. “Passionate craftspeople delivering quality” tells a machine nothing.

The legal name can differ from the trading name, and both can be stated honestly. Schema.org's legalName is for the official name, such as the registered company name, and alternateName holds an alias.¹³ For a person, settle on one form of the name, with or without a middle initial, and use it on the byline, the profile page, LinkedIn and any professional register.

An about page that states the facts

Give the entity one page on your own site that states its facts in visible text. For a business that means the name, the legal name and company number, the address or service area, the founding year, the people behind it, what it does, and links to its profiles elsewhere. For a person it is the profile or author page, and for a product it is the product page. Jason Barnard of Kalicube calls this the Entity Home: the single authoritative page an algorithm uses as its primary source of truth about an entity, usually the About page.¹⁴

For a UK limited company, start from the public record. Companies House publishes the registered address, incorporation date, current and resigned officers and previous company names free of charge,¹⁵ so the facts on your page should agree with that first. Put every fact in the page text as well as in any markup. The post on schema markup and AI citations explains why assistants that fetch a page live often never see JSON-LD.

Organization or Person markup with sameAs

Markup states your side of the record in a form a search engine can read without guessing. Put Organization markup, or the most specific LocalBusiness subtype for a single-location business, on the home page or the page that describes the organisation, which is where Google recommends it.¹⁶ Put Person markup on each profile page. Give each entity a stable @id and point sameAs at profiles that are unmistakably the same entity. Organization markup helps Google understand an organisation's administrative details and disambiguate it in results, and Google describes sameAs as a page on another site with more information about the organisation, such as a social media or review profile.¹⁶

Identifiers do the disambiguation that names cannot, because two firms can share a name but never a company number. Schema.org has vatID, leiCode for a Legal Entity Identifier, and a general identifier property for the rest.¹³ Show any identifier you mark up on the page as well. The JSON-LD patterns, with several entities linked by @id, are worked through in the entities guide.

A product's identity rests on its GTIN. GS1's rules make the brand owner responsible for allocating it, with narrow exceptions such as items made for a single customer,¹⁷ and schema.org's gtin property takes an 8, 12, 13 or 14 digit number with a valid GS1 check digit.¹⁸ A retailer should use the manufacturer's number and the manufacturer's model name, because one product sold under five slightly different names is five weak records. Feeds and markup for shopping are covered in the post on product data for AI shopping.

Wikidata, where notability allows

A Wikidata item is the strongest sameAs target most organisations and people can realistically get, and the bar is lower than Wikipedia's. An item is acceptable if it has a sitelink to a Wikimedia project, if it refers to a clearly identifiable entity that can be described using serious and publicly available references, or if it fulfils a structural need. When an item's notability is disputed it can go to a request for deletion, and the community decides.¹⁹

In my view an item is worth creating when there is a register entry plus independent coverage, such as a trade journal profile or a local news story, that supports each statement. If the only sources are your own website and directory listings, wait. Wikipedia is a different matter. A company is presumed notable if it has had significant coverage in multiple reliable secondary sources independent of it, and press releases, routine announcements and paid articles do not count.²⁰ People with a conflict of interest are strongly discouraged from editing articles directly, and paid editors must disclose who is paying them.²¹

Consistent third-party profiles

Then make the records you do not fully control agree with the ones you do: Google Business Profile, Bing Places and Apple Business for a local business, LinkedIn for the company and its people, industry directories, trade bodies, review sites and marketplaces. Search separately for your old names, old addresses and old phone numbers, because those searches find listings that a search for your current name misses. Fix the most visible records first. The guide to local SEO covers listings for businesses with a location or service area.

The Knowledge Graph Search API

Google's Knowledge Graph Search API lets anyone look up entities in the graph. It is read-only and returns individual matching entities without the graph around them. Google says it is not suitable as a production-critical service, is migrating it to Cloud Enterprise Knowledge Graph, and points new users to that product.¹ For a one-off check it is still the quickest way to see whether Google holds an entity for you. Requests carry an API key in the key parameter, as in Google's own example.¹

https://kgsearch.googleapis.com/v1/entities:search?query=Northfield+Joinery&types=Organization&languages=en&limit=5&indent=True&key=YOUR_API_KEY

{
  "result": {
    "@id": "kg:/g/11example00",
    "name": "Northfield Joinery",
    "@type": ["Organization", "Thing"],
    "description": "Joinery company in Harrogate",
    "url": "https://www.example.com/"
  },
  "resultScore": 112.4
}

The types parameter restricts results to entities of the schema.org types you name, prefix matches the start of names and aliases, and limit defaults to 20.²² Read the result field by field.

• No result for your trading name may mean Google files you under another name. Try the legal name, the name with the town, and prefix=true before concluding there is no entity.
• The wrong type, a Person where you expected an Organization or a Place where you expected a business, is worth investigating. It can mean your records are being mixed up with a person or place of the same name.
• description is a short description of the entity, and url is its official website, if Google has one.²² A missing or wrong url suggests Google has not connected the entity to your site.
• resultScore is an indicator of how well the entity matched the request constraints.²² It describes the match to your query, so do not compare it across queries or read it as a measure of authority.
• The ID is worth keeping. Wikidata stores IDs that begin /g/ under its Google Knowledge Graph ID property, files IDs that begin /m/ under Freebase ID, and links each one to a Google search using the kgmid parameter.³ If you have a Wikidata item, check whether it carries the ID.

Knowledge panels and Wikidata

Search for your exact name in a private window. Knowledge panels are created automatically once there is enough information about an entity on the open web,²³ so whether one appears, what it says and which sources it links to is a direct read of what Google holds. If one exists you can claim it by signing in to an official profile such as YouTube, Search Console or Facebook. Not every panel can be claimed yet, and a business serving customers at a location or in a service area should use a Google Business Profile instead.²⁴ Verified users can suggest edits with supporting URLs, and Google checks each suggestion against other public information on the web,²⁵ so fix the underlying records before you submit.

If you have a Wikidata item, open it and check the label, description, official website and identifiers. Then search Wikidata for your name and look for other items with the same label that a system could confuse with yours.²

Asking the assistants

With no lookup tool, the only way to check an assistant is to ask it. Answers vary from one run to the next. SparkToro and Gumshoe.ai had 600 volunteers run 12 prompts a combined 2,961 times in ChatGPT, Claude and Google's AI results in November and December 2025, and found less than a 1 in 100 chance that ChatGPT or Google's AI would return the same list of brands in any two responses.²⁶ That study measured recommendation lists, which are a different test from factual questions about one named business. The lesson carries over all the same: ask each question several times before reading anything into the answer.

If you use Bing Webmaster Tools, its AI Performance report shows how often your pages are cited as sources in Microsoft Copilot and AI-generated summaries in Bing, and the grounding queries the AI used when it retrieved them.¹² Grounding queries that contain your name show how Microsoft's systems look you up. How to measure visibility in AI search sets out a fuller method, and when an AI assistant gets your business wrong covers tracing a wrong fact back to the record it came from.

Entity optimisation is sold under several labels, including entity SEO, brand entity optimisation and AI trust signals. The useful work underneath is the list above, done carefully, and most of it is auditing and editing records. A reasonable engagement looks something like this.

What nobody can promise

• A knowledge panel. Panels are created automatically by Google's systems,²³ so there is nothing to order.
• Panel edits on demand. Google compares each suggested change with other public information before it acts on it.²⁵
• A Wikipedia article. Notability depends on significant independent, unpaid coverage,²⁰ and undisclosed paid editing breaks Wikipedia's rules.²¹
• A Wikidata item that stays. An item whose notability is disputed can be put up for deletion, and the final decision is the community's.¹⁹
• Citations or mentions in AI answers. In a study published in May 2026, Ahrefs tracked 1,885 already-cited pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched pages that did not. Citations in ChatGPT and AI Mode moved by amounts statistically indistinguishable from zero while AI Overviews citations fell slightly.²⁷ Google says there is no special schema.org markup to add for its AI features,⁹ and Bing's guidelines say this kind of optimisation does not guarantee citations, as Search Engine Journal reported.¹¹
• Rankings. Google's advice on hiring an SEO says no one can guarantee a number one ranking.²⁸
• A quick change to what a model learned in training. What a model learned stays until the provider releases a newer one, and no outside service controls that.

Questions to ask before you pay

• Which records will you change, and will every account stay in our name?
• Will you edit Wikipedia or Wikidata for us, and how will you disclose it?
• What will you measure, how and how often? Can we see the baseline before the work starts?
• What will you refuse to do? A good provider names the items in the list above without being asked.
• Can you show previous work, and explain in plain terms what you intend to do? Google's hiring advice treats guaranteed rankings, claims of a special relationship with Google and firms that will not explain what they intend to do as warning signs.²⁸

In my view a service earns its fee when the records are a mess: after a rebrand, a move or a merger, for a founder whose old company still dominates searches for their name, or for a product range sold under inconsistent names. A business with one location and tidy profiles can do this work itself, with the entities and structured data guide open alongside.