Entity SEO: making an AI engine understand your company

Type your company name into ChatGPT, cold, in a fresh session, with no other context. Just: "What is [your brand]?"

If it hedges, invents, or confidently describes a completely different company, you don't have an entity. You have a string. And an engine that can't tell what you are will never recommend you for anything, because recommending requires knowing.

That's the whole subject of this post: how a company goes from being a sequence of characters the model has seen a few times to being a thing it understands, with properties, relationships, and enough corroboration behind it that the model will put its name in an answer.

What is an entity, and why does it decide whether you get recommended?

An entity is a thing the engine has a stable internal representation of: what category it belongs to, what attributes it has, what other entities it relates to. A string is just characters.

Google formalized this years ago with the Knowledge Graph. Language models do a fuzzier, more powerful version of the same thing: during training, they build distributed representations where "Stripe" sits near "payments," "API," "Patrick Collison," "Y Combinator," and "Adyen," because those words co-occurred across millions of documents. Ask a model about payment processors and Stripe surfaces because the representation is dense — many documents, many contexts, all agreeing.

Your company probably has a sparse representation. Fifty documents, most of them yours, all saying slightly different things. When the model has to decide whether to name you in a list of five vendors, it's implicitly asking: how confident am I that this thing is what it says it is, and that it belongs in this category? Sparse and inconsistent means low confidence means you don't make the list.

The practical upshot: entity work is confidence work. Everything below is a way of raising the model's confidence that you are what you say you are.

Why does a model need corroboration from independent sources?

Because self-description is unreliable and models have learned that. Every vendor site claims to be the leading platform in its category. That claim carries almost no information.

What carries information is agreement between sources that have no incentive to agree. If your homepage says you're an AI visibility tracker, and Crunchbase says you're an AI visibility tracker, and G2 has you in the AI visibility category, and a TechCrunch piece describes you as an AI visibility tracker, and a Reddit comment says "we use it to track AI visibility," the model has five independent observations converging on one fact. Now it will say it back.

One source is a claim. Four independent sources is a fact. This is the actual mechanism, and it's why the corroboration gap is the number one reason competitors get named and you don't.

How do you actually build the entity footprint?

Five things, in this order. None of them are hard. The hard part is doing all of them and keeping them identical.

1. Write the canonical one-liner, then never deviate

Pick the sentence. Something like:

Spottlo is an AI visibility tracker that shows whether ChatGPT, Perplexity, Gemini and Google AI Overviews mention your brand.

Category noun, what it does, who for. No adjectives that don't carry information. Now paste that exact sentence into: your homepage, your About page meta description, LinkedIn, Crunchbase, G2, Capterra, GitHub, Product Hunt, your press kit, and every guest-post byline. Word for word.

Marketers hate this because it feels lazy. It is the single most effective entity intervention available to you. Every variant you introduce is a document that pulls the model's representation in a slightly different direction, and the entity resolves more slowly.

2. Link your profiles together with sameAs

sameAs is how you tell a machine that all those profiles describe one thing. Put Organization schema on your homepage or About page:

{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://spottlo.com/#organization",
  "name": "Spottlo",
  "url": "https://spottlo.com",
  "logo": "https://spottlo.com/logo.png",
  "description": "Spottlo is an AI visibility tracker that shows whether ChatGPT, Perplexity, Gemini and Google AI Overviews mention your brand.",
  "foundingDate": "2026",
  "sameAs": [
    "https://www.linkedin.com/company/spottlo",
    "https://www.crunchbase.com/organization/spottlo",
    "https://www.g2.com/products/spottlo",
    "https://github.com/spottlo",
    "https://www.wikidata.org/wiki/Q000000"
  ]
}

Two things people get wrong here. First, they put a sameAs array full of profiles they don't actually control or that don't exist. Every URL in there should resolve and should describe you. Second, they only list social profiles. Twitter and Facebook are the least useful entries; Crunchbase, Wikidata, G2 and GitHub are the ones that carry weight. Full schema patterns are in schema markup for AI search.

Add SoftwareApplication schema on your product/pricing pages too, with real offers containing real prices. Models verify pricing claims and a machine-readable price is trivially verifiable.

3. Claim the structured profiles

In rough order of value for a software company:

• Crunchbase. Free to claim. Fill in founding date, category, location, funding, description. It's a top-tier source for "what is this company" questions and it takes an hour.
• G2 and Capterra. Complete every field, get in the right category. See AI visibility for B2B SaaS for why these dominate software answers.
• LinkedIn. Consistent description, correct industry, correct headcount band.
• Wikidata. More accessible than Wikipedia, and it's machine-readable, which is the point. Wikidata's notability bar is lower: you need at least one serious reference (a funding announcement, real press coverage) to justify the item. If you have that, create the item, set instance of: business, add industry, inception, official website, and reference each statement. Then point your sameAs at it.
• Wikipedia. If you qualify, great, and it's worth a lot. Most companies don't, and do not write your own article. It gets flagged, deleted, and creates a paper trail of promotional editing.

4. Make your NAP and facts consistent everywhere

The old local-SEO discipline (Name, Address, Phone consistency) applies here for the same reason it always did: contradictory facts across sources reduce machine confidence.

Extend it beyond NAP. Founding year, founder names, headquarters city, pricing, employee count, and category should be identical wherever they appear. If your site says founded 2025 and Crunchbase says 2024, you've introduced a contradiction the model has to resolve, and the resolution might be "this data is unreliable."

Run the audit: pull up your site, LinkedIn, Crunchbase, G2 and every profile in your sameAs array, side by side, and check the facts match. Ours didn't the first time we did this. Nobody's does.

5. Get independent sources to state the same facts

This is the slow, unavoidable part, and it's the part that actually moves the needle.

You want third-party documents that say, in their own words, what you are. A trade publication piece. A roundup that describes you accurately. An analyst note. A conference talk listing. A podcast description. Each one is another independent observation, and independence is what you're buying.

The tactic that works: when anyone writes about you, give them the canonical one-liner and let them paraphrase it. Don't send a paragraph of positioning. Send the sentence. Most writers will use something close to it, and now you have another document agreeing with the others.

How to know if it's working

The cold test, repeated. Ask three engines "What is [brand]?" every month and grade the answer:

The step from C to B is corroboration volume. The step from B to A is being present on the pages that get retrieved for your category query, which is a different job.

And you can't grade this from memory, because model answers vary between runs. You need the same questions, the same engines, on a schedule. Spottlo runs it weekly and shows you the sources behind each answer, which is how you find out which document is teaching the model the wrong thing about you.

What to do next

• Run the cold test right now, in ChatGPT, Perplexity and Gemini. Write down what each one says. That's your baseline grade.
• Write the canonical one-liner. One sentence: category noun, what it does, who for. Get it approved once and then stop editing it.
• Put an Organization schema block with a real sameAs array on your About page this week. It's an hour of work and it's the cheapest entity signal available.
• Open your site, LinkedIn, Crunchbase and G2 side by side and reconcile every factual contradiction you find. There will be some.
• Claim Crunchbase and, if you have a citable reference, create a Wikidata item. Then re-run the cold test in six weeks and see if the grade moved.