Search Architecture
Entity Resolution: What Google Knows About You That Keywords Never Captured
Four operations behind entity resolution, why prior probability makes a generic name a permanent tax, why rebrands fragment identity, and why PR outperforms content here.
Quick answer
Entity resolution is the process of deciding that a string of characters refers to a specific real-world thing, and that it is the same thing referred to elsewhere under a different name. Search systems stopped matching keywords and started resolving entities more than a decade ago, and AI retrieval inherits that layer wholesale. The practical consequence is that consistency beats volume: one canonical, unambiguous, externally corroborated identity is worth more than a hundred pages of keyword coverage. And if your name collides with something more famous, you are fighting a disambiguation problem that no amount of content solves.
Entity SEO has been written about for years, almost always as a vocabulary lesson — here is what an entity is, here is a knowledge panel, add some schema. What is rarely explained is the actual computational problem, which is called entity resolution, and which determines outcomes that look inexplicable from a keyword perspective.
Understanding it explains why a small consistent brand can outperform a large inconsistent one, why rebrands are so destructive, and why some businesses simply cannot get a knowledge panel no matter what they publish.
The problem being solved
A system encounters the string “Apollo” across the web. It must decide, for each occurrence, which Apollo — the Greek god, the space programme, the sales intelligence platform, the theatre, or one of several hundred companies. Then it must decide that “Apollo.io,” “Apollo Sales,” and “Apollo (the B2B data company)” are the same entity while “Apollo Theatre” is not.
That is two distinct operations, and conflating them is where most entity SEO advice goes wrong.
| Operation | Question | What helps |
|---|---|---|
| Named entity recognition | Is this string referring to a thing at all? | Grammatical context, capitalisation, surrounding language |
| Disambiguation | Which of the candidate things is it? | Co-occurring entities, topical context, prior probability |
| Coreference | Are these two mentions the same thing? | Consistent naming, sameAs links, shared attributes |
| Canonicalisation | What is the preferred label for this thing? | Frequency of usage, authoritative sources |
Prior probability, and why it is brutal
Disambiguation systems use a prior: absent other evidence, an ambiguous mention resolves to the most commonly referenced candidate. This is statistically sensible and commercially merciless.
Mention: "Oracle"
prior P(database company) very high
prior P(Greek oracle) moderate
prior P(your consultancy) ~0
To override the prior you need CONTEXT that shifts it:
"Oracle" + "Larry Ellison" + "database" → the company
"Oracle" + "Delphi" + "Pythia" → the classical one
"Oracle" + [your city] + [your niche] → maybe you, sometimes
Without strong co-occurring signals, an ambiguous
name resolves to someone else and there is nothing
on your site that changes that.The uncomfortable part
If you chose a generic or colliding name, you have a permanent structural tax and no content strategy removes it. I have watched businesses spend years and considerable budget trying to out-publish a naming decision, and it does not work — you are competing against prior probability accumulated across the entire web. Naming is an SEO decision made by founders before anyone consults an SEO, and it is more consequential than most technical work that follows. Occasionally the correct advice is to change the name, and almost nobody wants to hear it.
What actually builds an entity
Entities are not built on your website. They are built in the corroboration between your website and everywhere else. Your own claims about yourself are the weakest evidence in the system, which is the inverse of how most SEO effort is allocated.
- A stable canonical name, used identically everywhere. Not “Acme,” “Acme Ltd,” “Acme Solutions” and “ACME” across four properties. Every variant dilutes coreference and forces the system to guess.
- Corroboration from sources you do not control. Third-party databases, professional registries, industry publications, academic citation, credible directories. This is the load-bearing evidence.
- Consistent co-occurring attributes. The same founder names, the same location, the same category descriptors, the same founding year, across every surface. Inconsistency reads as two different entities.
- Explicit sameAs relationships. Schema linking your site to your authoritative profiles elsewhere. This is where structured data genuinely earns its place.
- Association with other well-resolved entities. Being mentioned alongside things the graph already understands — named clients, recognised technologies, established institutions — pulls you into a resolved neighbourhood.
Point five is the mechanism behind something people observe and misexplain. When a small brand appears in a credible publication alongside recognised names, the benefit is not only referral traffic or a link. It is that a resolution system now has evidence placing that entity in a specific, well-mapped context.
My position
This is the strongest remaining argument for old-fashioned public relations, and I think the SEO industry undervalues it because PR is hard to attribute. A single mention in a source that resolution systems already trust, placing you next to entities they already understand, does more for entity establishment than a quarter of content production. The mechanism is disambiguation rather than link equity, which is why it does not show up properly in the metrics most teams watch.
Why rebrands are so destructive
Rebrand damage gets attributed to redirects and lost links. Those matter. The larger and slower problem is entity fragmentation.
When you change name, the world does not update at once. Older sources keep the old name, newer ones use the new one, some use both, some use one with the other in parentheses. From a resolution system’s perspective there are now plausibly two entities, with attributes split between them and neither fully corroborated.
- State the relationship explicitly and permanently. “Formerly known as X” in prose on your about page, not only in a press release that scrolls away.
- Update third-party profiles yourself. Registries, directories, professional profiles, conference bios. The ones you can edit are the corroboration you control.
- Keep the old name discoverable rather than scrubbing it. Erasing the old identity removes the bridge between the two.
- Expect eighteen months. Entity graphs update slowly because they are weighing corroboration across a very large corpus, and there is no request-reindexing button for identity.
Personal entities, which behave differently
For individuals — consultants, authors, founders — the resolution problem is harder in one specific way: name collision is far more common and completely outside your control. There are many people with your name and no naming committee to consult.
What works is a consistent bundle of co-occurring attributes rather than the name alone.
| Signal | Why it disambiguates |
|---|---|
| Name plus consistent professional descriptor | Narrows the candidate set immediately |
| Name plus location, used consistently | Geographic priors are strong |
| Authored work under a stable byline | Creates a corpus attributable to one entity |
| Citation in indexed sources | Academic, legal and journalistic citation carries disproportionate weight |
| Speaking and conference listings | Third-party corroboration with date and context attached |
| Consistent photograph across profiles | Used more than people expect for identity linking |
Why this matters more now, not less
A reasonable objection: if language models read prose and understand context, does an explicit entity layer still matter?
It matters more, for a specific reason. Retrieval systems sit between the user and the model, and those retrieval systems inherit the entity infrastructure that classic search built. When an assistant rewrites a question into search queries, the entities it recognises in that question determine what it searches for. If the system does not resolve your organisation to a distinct entity, you are not in the candidate set before the model sees anything.
There is also a second-order effect. Models generating text about a topic reproduce the entity associations present in their training data. If the corpus consistently associates your name with a specific category, that association appears in generated text without any retrieval at all. That is the one genuine mechanism by which training-data presence helps you — not brand awareness in general, but resolved entity association specifically.
An honest audit
- Search your exact name in an assistant, several times. Does it resolve to you, to someone else, or to a blend of several entities? A blend is the worst outcome and the most common.
- List every name variant in use across your properties. Legal name, trading name, product names, social handles. Count the variants. It will be more than you expect.
- Check attribute consistency. Founded year, location, category, leadership names, across every third-party profile. Contradictions are resolution failures waiting to happen.
- Audit your sameAs coverage. Does your Organization schema link to every authoritative profile you hold?
- Identify your strongest external corroboration and ask whether anything in the last year has added to it. If not, that is the gap, and it is not a content gap.
The conclusion I keep arriving at
Entity work is unglamorous, slow, and mostly happens off your own website. It cannot be sprinted. It produces no chart that moves within a quarter. And it is more durable than almost anything else in this field, because it is infrastructure rather than positioning — an entity that resolves cleanly is legible to every system built on top of the graph, including systems that do not exist yet.
Which is also why I would push back on the reflex to answer every visibility problem with more content. If a system cannot tell who you are, publishing more pages by an entity it cannot resolve does not help. Fix the identity first. It is boring, it takes eighteen months, and it is the foundation everything else sits on.
Where structured data genuinely contributes to this is set out in does schema markup help AI citation, and how the retrieval layer above it behaves is in how retrieval actually selects your page.
Frequently asked questions
What is entity resolution in SEO?
The process of deciding that a string refers to a specific real-world thing, and that it is the same thing referred to elsewhere under different names. It involves four operations: recognising that a string names an entity, disambiguating which candidate it is, establishing that two mentions coreference, and choosing a canonical label.
Why does a generic brand name hurt search visibility?
Because disambiguation uses prior probability — absent strong contextual evidence, an ambiguous mention resolves to the most commonly referenced candidate. If your name collides with something far more famous, you are competing against priors accumulated across the entire web, and no volume of content on your own site overrides that.
How do you actually build an entity?
Mostly off your own website. A stable canonical name used identically everywhere, corroboration from sources you do not control, consistent co-occurring attributes such as founder names and location, explicit sameAs links to authoritative profiles, and association with other entities the graph already resolves cleanly.
Why are rebrands so damaging beyond lost links?
Entity fragmentation. Older sources keep the old name, newer ones use the new one, and from a resolution system’s view there are now plausibly two entities with attributes split between them and neither fully corroborated. State the relationship permanently in prose, update third-party profiles yourself, keep the old name discoverable, and expect about eighteen months.
How do individuals establish themselves as entities?
Through a consistent bundle of co-occurring attributes rather than the name alone: a stable professional descriptor, a consistent location, authored work under one byline, citation in indexed sources such as academic or legal writing, speaking listings with dates, and even a consistent photograph across profiles.
Does entity SEO still matter now that AI reads prose?
More, not less. Retrieval systems sit between the user and the model and inherit the entity infrastructure classic search built. When an assistant rewrites a question into search queries, the entities it recognises determine what it searches for. If your organisation does not resolve to a distinct entity, you are excluded before the model sees anything.
Why is PR undervalued for entity building?
Because its mechanism is disambiguation rather than link equity, so it does not appear properly in the metrics most teams watch. A single mention in a source resolution systems already trust, placing you beside entities they already understand, can do more for entity establishment than a quarter of content production.
How do I audit my entity resolution?
Search your exact name in an assistant several times and check whether it resolves to you, someone else, or a blend — a blend is the worst and most common outcome. Then count every name variant across your properties, check attribute consistency on third-party profiles, audit sameAs coverage, and identify your strongest external corroboration.