Before a search engine ranks a business, it has to identify one. That sounds like a formality, and for large, well-documented brands it is. For everyone else it is a live question with commercial consequences. Google decides who you are before it decides where you belong, and every ranking signal, review, mention and backlink is only worth what the engine can confidently attribute to the right entity. When the identification is shaky, the signals scatter. The brand then underperforms in ways no keyword audit will surface.
Entity SEO is the work of making that identification unmistakable. It gets less attention than content or links, and it underpins both. In AI search it is no longer optional.
An entity is a thing, not a string
The distinction comes from Google itself. When it introduced the Knowledge Graph, Google described the shift as moving from strings to “things, not strings”, from matching the letters in a query to recognising the real-world objects behind them. The launch example was “Taj Mahal”: as a string it is two words; as things it is a monument, a Grammy-winning musician, a casino and a restaurant, each a separate entry with its own facts. At launch in 2012 the Knowledge Graph already held more than 500 million objects and more than 3.5 billion facts and relationships between them, and it has been the basis of Google’s understanding ever since.
An entity is a distinct thing, such as an organisation, a person, a place or a product, that exists independently of any particular words used to name it. It has facts (what it is, where it operates, who runs it, what it sells) and relationships (to its founders, its locations, its industry, its parent company). A keyword is a string. An entity is the thing the string refers to, and search engines rank entities.
The practical consequence is that the engines hold an entry for your brand, explicit or inferred, assembled from everything they have read. Entity SEO is the work of making sure that entry exists, is unambiguous and is correct.
Recognition precedes ranking
Consider a firm called Apex Legal. There are several firms with that name in Australia. When a signal attaches to the string, whether a review, a citation in an article, a directory listing or a link, the engine has to decide which Apex Legal earned it. If your entity is well defined, the attribution is clean and the signal compounds. If it is not, three things can happen: the signal is credited to a competitor with the same name, it is discarded as unresolvable, or it seeds a second, partial version of you that dilutes the first.
This is why entity recognition comes before ranking. Relevance and authority are judgements about an entity, and they cannot be made about a string the engine has not resolved. A business with strong content and genuine authority but a poorly defined entity loses the benefit of both.
For AI engines the dependency is sharper. A ranking system that misattributes a signal loses some accuracy, while a generative system that misattributes an identity produces a wrong answer with your name in it: the wrong services, the wrong city, or a merged description of you and your namesake. A model that cannot work out which “you” the retrieved passages describe usually takes the cautious route and leaves you out of the answer. In generated answers, the cost of ambiguity is being absent.
How an entity gets assembled
Engines build their picture of an organisation from two layers, and both are workable.
The declaration: what you state about yourself. This layer is fully in your control. It starts with Organization schema markup, which exists for exactly this purpose. Google’s structured data documentation describes it as helping Google “better understand your organization’s administrative details and disambiguate your organization in search results”. Name, legal name, address, telephone, logo, founder and the services offered all get stated in machine-readable terms that leave nothing to inference. The about page belongs to the same layer. Most businesses treat it as a branding exercise, and engines treat it as a primary source. It should read as a fact sheet in good prose: legal name, category, locations, specialisations, key people and history, written as plain declarative sentences. Write it for the machine as much as the buyer, and keep it to facts. The broader machine-readable layer of schema, llms.txt and structured entity facts is covered in its own article.
The corroboration: what everyone else says about you. Engines do not take a brand’s word for its own identity. They cross-check the declaration against independent sources. For an Australian SME the realistic corroboration set is concrete: the company register and ABN lookup, Google Business Profile, LinkedIn, the relevant industry association, established directories, supplier and partner pages, press coverage with a named author, and conference and event listings. Wikipedia and Wikidata are generally treated as the strongest identity anchors on the web, but most small and mid-sized businesses do not qualify for a Wikipedia article and should not manufacture one, because a thin promotional page does more harm than good. The tier below is achievable by almost anyone, and it is enough. What matters is that multiple sources you do not control describe the same organisation in the same terms.
The two layers verify each other. A declaration nothing corroborates reads as a claim, and corroboration with no clear declaration is hard for an engine to attribute. The entity forms when the two agree.
The sameAs web: consistency beats volume
The property that connects the two layers is small and underused. schema.org defines sameAs as the “URL of a reference Web page that unambiguously indicates the item’s identity”, such as the item’s Wikipedia page, Wikidata entry or official profiles. In practice it is your organisation schema saying that the entity on this website is the same entity as this LinkedIn page, this Google Business Profile, this directory listing and this registry record. It hands the machines a map of your identity across the web.
The map only works if the sources agree with it. Every profile you point to with the same name, description and facts strengthens the entity. A profile with a stale address, a superseded description or an old trading name introduces a contradiction at a location you flagged as authoritative. This is why consistency beats volume in entity building: ten sources that agree outweigh a hundred that almost agree. Aim to be described identically everywhere you are listed, and correct or remove the places you cannot keep true.
Entity gaps
Most entity problems are inherited from earlier decisions, and they share one trait: nothing visibly breaks. No error appears in any console. Rankings soften, citations fail to arrive, and the knowledge panel shows someone else or nothing at all, while every conventional audit comes back clean.
The recurring gaps are familiar to anyone who has looked for them. A rebrand can leave the old name alive in directories, press and profiles, so the engines hold two half-entities instead of one. A merger may never consolidate its identities. An alias problem develops when legal name, trading name and colloquial abbreviation are used interchangeably, each seeding its own partial record. A name collision with a larger business absorbs every ambiguous mention. The most common gap is a thin about page with no schema, which leaves the engines to infer identity from scattered fragments. They make that inference cautiously, incompletely, or in favour of a namesake.
Because nothing breaks, these gaps persist for years. Finding one is often the most useful discovery in an AI search assessment.
The sequence for establishing an entity
The work has a natural order, because each step feeds the next.
Decide the canonical facts first. Settle on one name written one way, one category noun the business will commit to, one two-sentence description, and a definitive list of locations, services and named people. This is a leadership decision. Skip it and the later steps will encode the ambiguity.
Publish the declaration. Rewrite the about page as declarative fact. Implement Organization schema with the canonical facts and sameAs links to every profile you intend to stand behind. Make the visible page and the markup say the same thing, because engines treat any difference between them as a discrepancy.
Reconcile the corroboration. List every place the organisation is described, including registries, profiles, directories, partner pages and old coverage, then bring each into line with the canonical facts or retire it. If a rebrand or merger is in the history, this step is the bulk of the work, and it costs coordination more than budget.
Earn the independent layer. This covers association memberships, press with named authors, event listings and published profiles for key people. The layer compounds slowly, so start it early.
Verify from the outside. Search the brand name. Check what the knowledge panel shows, if one exists. Ask the AI engines directly who the business is, what it does and where it operates, then compare their answers to the canonical facts. A well-established entity returns correct and consistent answers to all of it. The method for doing this systematically is set out in the self-audit guide.
Where entities fit in AI citation
Entity work is one factor in a larger frame. The four-factor citation framework lists content structure, entity clarity, authority signals and consistency of facts, and the other three all depend on the entity. Structure decides whether your passages can be quoted, and the entity decides who gets the credit. Authority accrues to an entity and lands nowhere when the entity is ambiguous. Consistency is largely entity consistency: the same facts about the same thing, everywhere the engines look.
Entity gaps often turn out to be the reason a brand publishes well and still is not cited. In those cases the content was fine and the engines were not sure who was speaking. Establishing the entity, closing the gaps and checking how the engines describe the brand is foundational work inside an AI search optimisation engagement. The first step costs nothing: ask the engines who you are, and see whether the answer is the one you would have given.