A KGMID is the machine ID Google assigns to an entity in its Knowledge Graph. It is a short string that starts with /m/ or /g/, like /m/045c7b for Google itself, and it is how Google’s systems keep your company separate from every other thing that shares its name. With an ID, Google treats you as a known thing. Without one, you are a string of words it has to guess about.

Google spelled out the difference on May 16, 2012, when Amit Singhal launched the Knowledge Graph under the line “things, not strings.” His example was [taj mahal], a query that could mean a monument in Agra, a Grammy-winning musician, a casino in Atlantic City or the closest Indian restaurant. The words are identical. The entities are not. Google settles that fight by attaching facts to IDs and deciding which ID a query points at.

This guide covers what the ID looks like, where it came from, why AI assistants lean on it, and a four-part model I call the 4M Entity Spine for judging whether your entity record is healthy. If you want the step-by-step lookup instead, jump to the three-method lookup walkthrough.

What does a KGMID look like?

Glowing HTML code on a monitor, the kind of markup that carries entity IDs

There are two formats, and the prefix tells you the age of the record, not its quality.

The first format is a slash, the letter m, another slash and a short code. Google’s own entity is /m/045c7b. IBM is /m/03sc8. These are legacy Freebase machine IDs, and Wikidata’s description of its Freebase ID property states the rule in one line: new Freebase MIDs are no longer issued, but existing /m/ identifiers remain in use, including by Google Knowledge Graph.

The second format starts with /g/ and runs longer, in most cases /g/11 followed by a string of letters and digits. OpenAI is /g/11bxc656v6. Wikidata describes this property as the “identifier for Google Knowledge Graph API, starting with ‘/g/’.” Any company founded in the last decade will almost always carry a /g/ ID, since the /m/ supply stopped when Freebase did.

I pulled each of those IDs from the public Wikidata records on November 18, 2026. They are real, and you can check them yourself.

You will run into these strings in three places. The Knowledge Graph Search API returns them in the @id field with a kg: prefix, as in kg:/m/045c7b. Google search URLs carry them in a kgmid parameter, which is what you see when you share a knowledge panel and expand the link. And Google Trends uses the same IDs to define a topic, which is why a Trends topic URL for a company contains a /m/ or /g/ string in place of plain words.

The format matters less than the count. One ID that points at the right entity is the goal. Zero is a problem, and two is often a bigger one.

Freebase built the ID system Google still runs on

The /m/ prefix is a fossil from a startup. Metaweb Technologies launched Freebase in March 2007 as an open, community-edited database of structured facts, and every topic in it got a machine ID. Google bought Metaweb in July 2010. When the Knowledge Graph launched two years later, Singhal named Freebase as one of its public roots, next to Wikipedia and the CIA World Factbook, and said the graph held more than 500 million objects and 3.5 billion facts.

Google announced on December 16, 2014 that it would close Freebase and help move its data to Wikidata. The site went dark on May 2, 2016. The IDs did not. Google kept the /m/ codes inside its own graph and started minting /g/ codes for everything new. By May 2020, Danny Sullivan wrote on Google’s blog that the Knowledge Graph had “amassed over 500 billion facts about five billion entities.”

Here is the part most explainers skip. Because Wikidata inherited the Freebase mappings and its editors kept adding Google IDs, Wikidata is now the largest public lookup table of Google entity IDs anywhere. I ran a count against the Wikidata query service on November 18, 2026. It returned 4,470,136 items carrying a Freebase ID (P646) and 8,181,926 items carrying a Google Knowledge Graph ID (P2671). Against five billion entities, that is a sliver, but it is the sliver with public, sourced, editable records. If your company has a Wikidata item, it may already list your ID. If it lists the wrong one, that error is visible to anyone who looks, including the crawlers that feed AI products. Our guide to Wikidata for AI search covers how those records get used.

Why AI assistants care about your entity ID

Man in sunglasses using voice features on his smartphone to ask an assistant

No chatbot reads an entity ID aloud. The ID works one layer down, in the systems that decide which facts belong to which name before an answer gets written.

Google has said this out loud. When it introduced AI Mode on March 5, 2025, Robby Stein, VP of Product for Google Search, wrote that the feature lets you “tap into fresh, real-time sources like the Knowledge Graph, info about the real world, and shopping data for billions of products.” AI Mode fans a question out into many related searches and stitches the results together. When one of those searches touches your company, the Knowledge Graph is one of the places it checks, and the graph is organized by ID.

The practical effect is disambiguation. Ask an assistant who founded a company called Summit, and it has to pick a Summit before it can answer. The entity with a clean, well-connected record wins that pick. The entity with a thin record loses it to a namesake, and the answer comes back with someone else’s founder, someone else’s headquarters, someone else’s funding round.

There is a second effect that hurts more because it is quiet. If Google holds two IDs for the same company, the facts about that company get split across them. One record has your founding date and your CEO. The other has your product line and your press coverage. Neither looks complete, so neither gets trusted. An AI answer built on half a record reads as vague at best and wrong at worst.

Other assistants run their own pipelines and do not publish them. But the inputs overlap. Wikidata, Wikipedia, schema.org markup and major news sources feed most of them, and those are the same sources where Google IDs get recorded and cross-referenced. A clean entity record is one of the few pieces of AEO work that pays off in more than one engine at once. If the broader idea is new to you, start with our knowledge graph explainer.

Meet the 4M Entity Spine

I use a simple four-layer model to grade an entity record. I call it the 4M Entity Spine because each layer has to hold before the one above it carries any weight. Score one point per layer you pass. Four is healthy. Two or below means you have work to do before any PR or content push will land the way you want in AI answers.

LayerThe questionWhere you checkWhat failure looks like
MintedDoes an ID exist for you at all?Knowledge Graph Search API, panel share linkNo result for your exact name
MatchedIs there exactly one ID, and is it you?API with a type filter, WikidataTwo IDs for one company, or your name tied to a namesake
MirroredDo your own properties point at that ID?Wikidata item, your Organization schema, social profilesWikidata missing the ID, schema sameAs empty or stale
MentionedDo independent sources describe you the same way?News coverage, directories, interviewsConflicting founding dates, titles or locations

Minted is binary. Google either has a record or it does not. If it does not, nothing else on this list matters yet, and your job is to create the conditions for one: consistent facts on your own site, a notable presence in sources Google reads, and third-party coverage that names you without ambiguity.

Matched is where most mid-sized brands fail. They have an ID, sometimes two, and one of them is stuck to an old company name or a founder’s prior venture. Our breakdown of duplicate knowledge panels walks through how those splits happen.

Mirrored is the layer you control. Your Wikidata item can list your ID. Your Organization markup can list your official profiles in sameAs. Some SEOs also add a Google search URL with the kgmid parameter to sameAs. Google has not documented that as a ranking signal, so treat it as a tidy cross-reference, not a lever.

Mentioned is the slowest layer and the one that compounds. Every article that states your founding year, your founder’s name and your category in the same words gives the graph one more vote for the same facts. Conflicting coverage does the opposite.

Where entity IDs go wrong

Even the biggest names in tech carry messy records. When I pulled the Wikidata item for Anthropic (Q116758847) on November 18, 2026, it listed two Google Knowledge Graph IDs: /g/11nn47s4yv and /g/11csrrqr2z. Apple Inc. (Q312) listed two Freebase IDs, /m/0k8z and /m/04st9hr. From the outside, you cannot tell which of each pair Google treats as primary, or whether the second record still holds facts of its own.

For Apple and Anthropic, a split costs little, because the volume of coverage overwhelms any confusion. For a 40-person company with six press mentions, the same split can mean an assistant knows half of what it should.

The other common failures are easier to spot once you know the shape. There is the company with no ID at all, which shows up in AI answers as a guess pieced together from its own homepage. There is the founder merged with a namesake, which looks like a knowledge panel with someone else’s photo or birth year. There is the ID that resolves but never produces a panel, which tends to mean Google has the record but lacks confidence in it. And there is the ID that vanishes after a Wikipedia deletion or a Wikidata merge, the scenario we cover in why a knowledge panel disappears.

Each of these maps to a layer in the 4M Entity Spine. No ID is a Minted failure. Two IDs or a namesake merge is a Matched failure. A missing Wikidata link is a Mirrored failure. Conflicting coverage is a Mentioned failure. Name the layer and you know which fix comes first.

Put your KGMID to work this week

Block 30 minutes and run four moves in order.

First, find the ID. Use the three methods in our lookup walkthrough and check your company, your founder and your flagship product. Write each result in a sheet with today’s date.

Second, count. If any entity returns two IDs, flag it. That is your highest-priority fix, because every other improvement gets diluted across the split.

Third, score each entity on the 4M Entity Spine and write the number next to the ID. A founder at 2 and a company at 4 tells you where the next quarter of effort should go.

Fourth, set a calendar reminder for 90 days out and run the same lookup again. IDs change, records merge, and the only way to catch a drift before it shows up in an AI answer is to check on a schedule. If you want a faster read on where your brand stands across the major assistants first, run the free AEO rating and bring your entity IDs to the review.