Search and AI
Google ranks entities, not keywords.
The shift
The unit of ranking stopped being the word.
For most of its history, Google matched strings: the text on your page against the text in a query.
In 2012 it began moving from strings to things, building a model of the real-world entities behind a search (a brand, a product, a place) rather than the words used to describe them.
The keyword model did not scale. A model of how entities relate to one another did.
One Google patent application describes how that can work at scale: represent a site as a mathematical vector, a fingerprint of what it is and how authoritative it is in its field, then classify it by measuring the distance between that fingerprint and the composite profile of the recognised authorities in the same category.
The method is published. US patent application 20200050707, "Website Representation Vector", Google LLC. A patent application shows a method Google has claimed, not proof of exactly how every search is ranked today.
Google no longer asks "does this page mention golf clubs?" It asks "does this brand look like an authority in golf?"
Part of how it answers is a graph of entities, each with a stable identity. In my experience, if Google cannot resolve your brand to a distinct entity in that graph, you are not so much ranked down as left out of the conversation the systems are actually having.
The proof: run it in five minutes
One identifier follows the entity everywhere.
Every entity Google recognises carries one persistent key: a Machine ID.
Take "golf club," the piece of sporting equipment. The same ID, /m/03gzqq, appears across three independent surfaces, one run by Wikidata and two by Google. This is not a coincidence; it is the identity.
| Where | What it shows | Check it |
|---|---|---|
| Source · Wikidata | Golf club (Q1054671): "A piece of sporting equipment used to hit a golf ball." The Freebase ID property holds the key. | wikidata.org/wiki/Q1054671/m/03gzqq |
| Surface · Google Trends | The entity, not the search term. Trends tracks the thing itself. The MID is sitting in the URL, encoded. | trends.google.com/explore?q=%2Fm%2F03gzqq |
| Surface · Google Search | Query the graph directly. Search by Knowledge Graph ID and Google returns its explicit understanding of the entity. | google.com/search?kgmid=/m/03gzqq |
One identifier · three surfaces · one entity Google is certain about.
Disambiguation
"Golf clubs" is not one thing. That is the whole point.
Type the phrase into Trends and it splits into separate entities, each with its own ID and its own node in the graph.
| Entity | Type | Identity |
|---|---|---|
| Golf Club | Sports equipment | /m/03gzqq |
| Golf course | Topic / place | distinct entity |
| Country club | Topic / establishment | distinct entity |
Ground your content and markup to the wrong one and you are reinforcing an identity that has nothing to do with what you sell.
The prefix matters
Why your brand almost certainly has a /g/, not a /m/.
The prefix tells you the entity's origin, and it changes how you find it.
/m/ Migrated from Freebase
Existed in Freebase, the database Google acquired and later closed, with much of its data moved to Wikidata. /m/ IDs date from the Freebase era, mostly established concepts, places and categories. Lookup property: Freebase ID (P646).
/g/ Created by Google
Minted by Google after Freebase closed. Your brand, your company, most recent people and products live here. Lookup property: Google Knowledge Graph ID (P2671).
The common finding: there is no entity at all.
The trap most people fall into
Not every Knowledge Graph identifier is the same thing.
Google runs several Knowledge Graph verticals (Business Profiles, Books, Podcasts and others), and each can mint its own kgmid.
An ID from one of those may not resolve in the main graph. Drop an unresolvable ID into your sameAs or schema and you have handed Google, and every AI model reading your markup, a contradictory signal.
Rule: verify the ID resolves before you ever cite it.
The AI layer
The same entity layer now shapes whether AI mentions you.
This stopped being only an SEO question.
When ChatGPT, Google's AI Overviews, Perplexity or Copilot answer "who is the best provider of X," they have to work out which real businesses the question is about. Clear, structured, machine-readable identity signals make that easier to get right.
For Google, the kgmid is one of the firmest anchors available because it is unambiguous and language independent.
No entity, the wrong entity, or an unresolvable one, and the risks are predictable: you can be confused with someone of a similar name, represented inconsistently, or left out of the answer entirely.
Entity authority used to be how you ranked. Increasingly, it is the price of being cited at all.
The method
RightView: close the distance to the entity Google expects.
RightView is the diagnostic I run before anything else. It estimates how far a brand sits from the authority profile of the recognised leaders in its category, and what is widening the gap, across six weighted dimensions from entity authority to technical foundation. The weighting is my own model, not a copy of Google's.
The Machine ID work is step one, because there is no point optimising content for an entity that does not yet exist.
- Confirm the entity exists. Find the kgmid, or establish that there is not one. The absence is the finding, and the starting line.
- Confirm it is the right one. Disambiguate. Make sure Google has bound your brand to the correct node, not a similarly named entity or a dormant vertical ID.
- Confirm it resolves. Verify the ID lives in the main Knowledge Graph before it goes anywhere near your structured data.
- Reinforce it. Align schema and sameAs, Wikidata and Wikipedia, and consistent entity associations so every signal points at the same thing.
Work with me
Find out what Google actually thinks you are.
A RightView diagnostic shows you your Machine ID, or its absence, the entities you own versus the ones you should, and the gaps between your brand and the businesses Google already treats as authorities in your category.