How Google Learns That Your Company Is a Thing, Not a Name

There are at least four companies called Atlas in Brazil. One makes fans, one is a brokerage, one is a language school, one is an auto shop. Crowded names work the same way in every market.

When someone types Atlas, the search engine has to choose. It does not choose by the name, because the name is identical in all four cases.

It does something else, and that something else has a birth date.

To the machine, the name of your company is a string of letters

On May 16, 2012, Amit Singhal published a post on the Google blog with a title that said it all: “Introducing the Knowledge Graph: things, not strings”. Things, not strings of letters. ¹

The announcement came with two numbers. More than 500 million catalogued objects, and more than 3.5 billion facts and relationships between them. ¹

An object, there, is a thing in the world. A company, a person, a place, a product. Each one gets its own machine identifier, inherited from an earlier system and exposed today in the kg: format. ²

That identifier is what separates the four Atlases. They share a name and carry different codes.

This is called an entity.

What Google stores about a thing

The three-part sentence

The record does not store paragraphs. It stores short three-part sentences: subject, predicate, object. ³

Your company, written that way, becomes a stack of lines. Acme LLC, is headquartered in, Denver. Acme LLC, works in, corporate architecture. Acme LLC, was founded in, 2014.

Each of those lines can be confirmed or contradicted by an outside source. That possibility of checking is what makes the structure work.

The original catalog was not born inside Google. It came from the purchase of Metaweb, owner of a base called Freebase, announced in July 2010. Wikipedia, Wikidata, and structured data from the web came later. ⁴

The size of the record, in two snapshots

At the 2012 launch, it was 500 million objects and 3.5 billion facts. ¹

On May 20, 2020, Google published another number: more than 500 billion facts about 5 billion entities. ⁵

Eight years between one snapshot and the other. Both are official Google statements, with a date, and not market estimates.

Where the facts come from, and what Google admits about them

In that same 2020 post, Google describes the sources of the record. The sentence is short: “Wikipedia is a commonly-cited source, but it’s not the only one. We draw from hundreds of sources from across the web”. ⁶

Hundreds of sources. Wikipedia is one of them, it gets cited often, and it is not the only one.

The post admits something else that rarely gets quoted alongside it: “Inaccuracies in the Knowledge Graph can occasionally happen”. Inaccuracies happen, and there is a feedback channel in the panels for correction. ⁷

Put the two sentences together and one practical consequence is left over. If the record feeds on hundreds of sources, what you publish about yourself is one of them. It is the one you control, and for that very reason it weighs least when it stands alone.

The name for this is reconciliation, and Google is the one who says so

Deciding that two records describe the same thing has a technical name, and the name is not a market invention.

In the Google Cloud documentation, there is a service described as “a standalone API that wraps around the Google core entity reconciliation”. Entity reconciliation, in those words. ⁸

One limit, and it is mine, not the document’s. That service is sold to companies for reconciling databases of their own. Its documentation does not describe how Search works. I use the term because it is the term Google uses for the operation. The bridge between the cloud product and the search engine is my reading. ⁹

Reconcile is the right verb for what happens to your company. Mentions of it sit scattered around, with different spellings, different addresses, sometimes different facts. Someone has to decide whether all of that is a single thing.

Three patents, and what they actually prove

The three, with number and date

Google patented mechanisms tied to this, and the three patents below are granted, not merely filed.

US 9,477,759 B2, “Question answering using entity references in unstructured data”, was filed on March 15, 2013 and granted on October 25, 2016. It describes entity recognition in free text, meaning text with no markup at all. ¹⁰

US 10,235,423 B2, “Ranking search results based on entity metrics”, was filed on December 12, 2012 and granted on March 19, 2019. ¹¹

US 11,328,218 B1, “Identifying subjective attributes by analysis of curation signals”, was filed in 2017 and granted on May 10, 2022. ¹²

What a granted patent proves

Here is the part the industry tends to skip.

A granted patent went through examination. That proves investment and engineering direction. It does not prove the system is in production, and it does not prove how it works today. ¹³

Anyone citing a Google patent number as if it were an operating manual is making a promise the document never makes.

There is also a bridge that keeps coming up in this subject, between entity recognition and the Hummingbird update of 2013. The chronological relation exists and the conceptual one makes sense. No primary source links the two documents, and that is why this bridge enters here as my reading. ¹⁴

Two things Google tried and shut down

Google researchers published a paper in 2014 on a system called Knowledge Vault. It gathered 1.6 billion triples, of which 324 million cleared a confidence threshold. ¹⁵

Knowledge Vault never became a product. Google itself said so publicly, in August 2014, after the press treated the paper as an announcement. ¹⁶

Freebase, the catalog bought in 2010, also came to an end. The shutdown was announced on December 16, 2014. The base froze for editing on March 31, 2015. Site and APIs went offline on June 30, 2015, and the data moved to Wikidata. ¹⁷

I keep those two stories here for a reason. Maintaining a record of facts about the world is expensive, and Google shuts down what does not hold at scale. Anyone waiting to find a signup button is waiting for the opposite of what this history shows.

There is no form

This is the question that arrives most often, and the answer is short.

There is no form, no registration, and no public submission process for adding an entity to the Knowledge Graph. None. The absence is a verifiable fact: just go looking. ¹⁸

There is a channel for suggesting a correction when a panel is already live, and that is another matter. Correcting a record that exists differs from creating one that does not.

What is left is the long road, and it is the only one. Being mentioned consistently by sources you do not control. With the same facts, written the same way.

Your Monday exercise

Ten minutes, no tools.

Ask ChatGPT and Gemini who your company is. Use the exact name, with no context at all.

Write down what comes back. It may come back right, it may come back wrong, it may come back empty. All three answers are information, and the third is the most common in a company that never worked on this.

Save the screenshot with the date. That is your zero mark for recognition, and without it there is no way to say later whether anything moved.

Frequently asked questions

Can I register my company in the Knowledge Graph?

No. There is no form, no registration, and no public submission process. What exists is a correction channel for panels that are already live. ¹⁸

Does putting JSON-LD on my site guarantee entry into the graph?

It guarantees nothing. Structured data is a statement of yours about you, and the record feeds on hundreds of sources. The statement helps with disambiguation. On its own, it is the weakest source that exists about your company, because it is the only one you control. ⁶

If Google gets my company wrong, can it be fixed?

Google itself acknowledges that inaccuracies happen, and it keeps a feedback channel in the panels. ⁷

Do the three patents prove Google does this today?

No. A granted patent proves investment and engineering direction, and it does not prove a system in production. ¹³

Notes and sources

1. Amit Singhal, “Introducing the Knowledge Graph: things, not strings”, official Google blog, May 16, 2012. Launch numbers: more than 500 million objects and more than 3.5 billion facts and relationships.
2. Google Knowledge Graph Search API, official documentation. MIDs inherited from Freebase are exposed as IDs in the kg: format. Two things with the same name receive distinct identifiers.
3. Triples of subject, predicate, and object: the standard data model of knowledge graphs, reflected in the Knowledge Graph Search API.
4. Acquisition of Metaweb, owner of Freebase, announced by Google in July 2010. Later additions: Wikipedia, Wikidata, and structured data from the web.
5. Google, “A reintroduction to our Knowledge Graph and knowledge panels”, blog.google, May 20, 2020. Official statement of more than 500 billion facts about 5 billion entities.
6. Same post of May 20, 2020. Literal quotation reproduced in the body.
7. Same post of May 20, 2020. Literal quotation reproduced in the body. The text describes automatic prevention systems and a feedback channel in the panels.
8. Google Cloud, documentation of Enterprise Knowledge Graph and of the Entity Reconciliation API. Literal quotation reproduced in the body.
9. Declared limit: the Entity Reconciliation API is a service sold to companies for reconciling their own data, and the documentation does not describe how Search works. The bridge between the two is the author’s inference.
10. Granted patent US 9,477,759 B2, “Question answering using entity references in unstructured data”, Google LLC. Filed on March 15, 2013, granted on October 25, 2016.
11. Granted patent US 10,235,423 B2, “Ranking search results based on entity metrics”, Google LLC. Filed on December 12, 2012, granted on March 19, 2019.
12. Granted patent US 11,328,218 B1, “Identifying subjective attributes by analysis of curation signals”, Google LLC. Filed in 2017, granted on May 10, 2022.
13. Evidence standard applied to the three patents: all granted, meaning all went through examination. A grant proves investment and engineering direction, and it does not prove a system in production.
14. Bridge with Hummingbird (2013): a chronological and conceptual relation, flagged as the author’s inference. No primary source links the two documents.
15. Dong et al. (Google), “Knowledge Vault: A Web-Scale Approach to Probabilistic Knowledge Fusion”, KDD 2014. 1.6 billion triples gathered, 324 million above the confidence threshold cited in the paper.
16. Google’s statement that Knowledge Vault was not a product, published as a postscript after press coverage, August 2014.
17. Outcome for Freebase: shutdown announced on December 16, 2014, base frozen for editing on March 31, 2015, site and APIs retired on June 30, 2015, with the data transferred to Wikidata.
18. There is no form, no registration, and no public submission process for adding an entity to the Knowledge Graph. The absence is a verifiable fact.