When Google says it understands "things, not strings", the knowledge graph is the thing it means. It is a database, enormous and private, of entities and the relationships between them: people, companies, places, books, films, events, each with an identifier, a set of facts, and links to the other entities it is connected to. The panel beside a search result is a window onto one entry. An AI Overview about a person is written with that entry as its anchor, when one exists. This post explains what the database is, how a person gets an entry, and how to read the little you can see of your own.
A database of things, not a list of pages
The web index is a list of pages with the words on them. The knowledge graph is different in kind: it holds a record for each thing Google is confident exists, independent of any single page. The record for a person might contain a name, alternative names, a description, a date and place of birth, education, employer, works, a photo, and a set of links to the pages that describe them. Each of those facts was extracted from one or more sources and given a confidence. Each of the relationships (works for, wrote, educated at) points at another entity's record, which is what makes it a graph rather than a table.
Google announced it in 2012 with around 500 million entities. It has grown by orders of magnitude since. Much of it was seeded from Wikipedia, Wikidata and other structured databases, and it is continuously extended from the open web, which is why your own site can contribute to it.
How a person gets in
There is no form. An entity is created when Google's systems become confident that a particular string, your name, refers to one real thing with a consistent set of facts attached. In practice that confidence comes from a handful of signals that agree:
- A page that presents itself as the canonical description of the person, and states facts plainly.
- Structured data on that page declaring a Person with a stable identifier and typed relationships.
- Profiles on major platforms that use the same name, photo and facts, connected to that page.
- Third party pages that describe the person consistently: interviews, articles, directories.
- An entry in Wikidata or Wikipedia, which Google treats as a direct feed.
None is required and none is sufficient. Enough of them, agreeing, over enough time, and the record forms. How to get a Google knowledge panel for yourself is the practical version of that sentence.
The identifier, and how to see yours
Every entity has an identifier. For older entries it looks like /m/0abc12 and for newer ones /g/11abc1234. You will never be told yours, but you can find it: when a panel appears for you, the address bar of that search contains kgmid= followed by the code. Copy it. From then on you can look up your entity directly by putting google.com/search?kgmid=/g/XXXX into the address bar on the All tab (the Web tab, which adds udm=14, strips panels by design).
That lookup answers two questions a plain name search cannot. If the identifier resolves and shows a panel, the entity exists and any problem you have is about triggering, which is confidence. If it resolves to nothing, the entity has been merged into another or dropped. Checking it monthly tells you whether your record is stable, and once you have claimed the panel the same identifier is what you add to your schema and to Wikidata as the Google Knowledge Graph ID, closing the loop between your record and Google's.
Why there can be two of you
The graph forms entities from evidence, and evidence can be split. A stale author bio on a trade site, describing you under an old title, can be enough on its own for a second record to form, with its own identifier, disconnected from the one your current site feeds. Then two entities compete for one name, and Google, unsure which is canonical, often shows neither on the plain query.
You find the duplicate the same way you find your own: it will have its own kgmid, usually discovered by searching the old title plus your name. The fix has two parts. Report the duplicate through the feedback link under the panel, category "Incomplete", because the missing thing is a merge, citing your canonical page and Wikidata as evidence. And fix the stale source, by asking the site to update the bio and link your home page, because a merge that leaves the cause in place tends to split again. Claiming a knowledge panel covers the feedback route in detail.
What the graph does with your facts
Each fact in your record carries a provenance. Birthplace from Wikidata. Employer from your site and three interviews. Photo from the panel owner's suggestion. When sources disagree, the fact with the stronger provenance wins and the other is dropped or held at low confidence. This is why the same fact stated in several places, in the same words, matters more than a new fact stated once: agreement raises confidence, and confidence is what makes a fact show.
It is also why first party claims do not appear. A line on your site saying you have advised two hundred companies has no second source and no way for the machine to verify it, so it is not a fact in the graph, however true it is. The graph carries what can be corroborated. Everything else stays in prose, which people read and the machine passes over.
Relationships are the point
A person's record connected to nothing is a weak record. The same record with typed edges to an Organization entity (works for), a Book entity (author), a University entity (educated at) and a set of Podcast entities (guest on) is strong, because each edge is a second entity vouching for the first. The company's own record says it was founded by you; the book's record says it was written by you. These cross references are what the hidden layer on your own site is for: it declares the edges, and when the other entity's sources agree, the graph adopts them.
This is also the quiet argument for collecting appearances. Each episode page can become an entity edge, a dated third party relationship between you and a show, and ten of them weigh more than one.
Where the AI assistants fit
It is tempting to think of the knowledge graph as a Google thing and ChatGPT as something separate. In practice the assistants that search the web live lean on the same evidence the graph is built from, and Google's own AI features read the graph directly. Perplexity and ChatGPT with browsing retrieve the top pages for your name, extract facts, and cite sources; when a Wikidata entry and a canonical page with structured data exist, those are the pages they retrieve and cite, because they are the most extractable. Gemini and Google's AI Mode anchor on the graph entity when there is one. The models that answer from training data learned from a snapshot of the same web, which is why an improved record reaches them last.
So building for the graph and building for the assistants are the same job. One canonical page, one identifier, consistent sources, Wikidata where it applies. The graph is where that evidence is stored. The assistants are where it is read back to the person asking.
What you can and cannot change
You cannot edit the graph. You can claim the panel for your entity, which gives you suggestions on a few fields and a feedback channel. You can change the sources the graph reads: your site, your profiles, Wikidata, and the third party pages you can persuade to update. You can make everything agree. That is the entire toolkit, and it is enough, because the graph is built to follow consistent evidence.