Talk to us

Entity SEO for LLMs: how to make language models know, find and cite your brand

Entity SEO for LLMs is the work of getting large language models to describe your brand correctly, both from what they absorbed in training and from the pages they fetch when they search. In practice: a clear, consistent entity, independent mentions, facts a model can quote and crawlers that can reach your site.

Entity SEO for LLMs: how large language models learn about a brand from training data, retrieval and entity signals

Ask ChatGPT about a well-known company and the answer is already in its memory; any search it runs mostly confirms what it knew. A small company has to be looked up. And a brand it has never met may simply be read as a typo for something else, as we found on October 27, 2026.

That gap is what entity SEO for LLMs is about. Other guides on this site cover being cited by AI search in general. This one looks at the model itself: how it learns about a company, why some brands live in its memory and others only in search results, and what turns a business into an entity a model can recognize and describe.

What entity SEO for LLMs means in practice

Entity SEO for LLMs means optimizing for the reader that now sits between your website and many buyers: a large language model. Its goals overlap with generative engine optimization, and the labels are used loosely, but the emphasis differs. GEO usually starts from the answer engine and its citations. This work starts from the model and asks two questions: what does it already believe about us, and what will it find if it looks? For the wider tactics of getting named and cited by assistants, Slibur.com, a tool from the same owner as Vértigo, keeps a guide to LLM SEO.

Those two questions map to two kinds of memory, and each needs different work.

Training memoryRetrieval (live search)
What it isPatterns learned from text before the model’s cutoff datePages fetched at question time and read before answering
How freshMonths to years oldAs fresh as the index the assistant searches
Shows a sourceNoUsually, as a citation or link
What decides itHow often and how consistently the web described you before the cutoffWhether your pages rank for the searches the assistant writes, and can be crawled
Crawlers involvedGPTBot, ClaudeBot, Google-Extended (for Gemini)OAI-SearchBot, Claude-SearchBot, PerplexityBot, Googlebot
How long a fix takesUntil a new model with a later cutoff shipsDays to weeks, once pages are crawled and ranking

Sources: OpenAI, Anthropic, Google and Perplexity crawler documentation, checked October 27, 2026.

If you want the wider comparison of SEO, AEO and GEO, our side-by-side guide covers it. Here we stay with the model and the brand.

Training data vs retrieval: two ways a model learns about you

A model learns about you once, slowly, in training, and then looks you up again, quickly, whenever a question needs it. Researchers named the second part years ago. A 2020 paper from Facebook AI Research called it retrieval-augmented generation, or RAG: a model’s built-in “parametric” memory paired with a searchable index of documents it can pull in at answer time. Every AI assistant that searches the web works on that principle.

Training: not all text counts the same

Training data is a weighted mix, not a flat copy of the internet. The GPT-3 paper is one of the few that published its recipe. Filtered Common Crawl, a public scrape of the web, supplied 410 billion tokens and 60% of the training mix. English Wikipedia supplied only 3 billion tokens and 3% of the mix, but the model passed over Wikipedia 3.4 times during training, against 0.44 times for the Common Crawl text. Curated, trusted text was deliberately read more often.

Newer models do not publish recipes in that detail, so treat the numbers as an illustration rather than today’s formula. The lesson still holds. A sentence about your company on a heavily cited reference site carries more weight in training than the same sentence on a page nobody links to. And anything published after a model’s cutoff is invisible to its memory; our guide on where ChatGPT gets its information explains cutoffs and live search in ChatGPT specifically.

Retrieval: the part you can change this month

When an assistant searches, ordinary search rules apply. Google’s documentation on AI features says a page only needs to be indexed and eligible for a snippet to appear as a link in AI Overviews or AI Mode, with no special files or schema. Other assistants depend on their own crawlers and search partners, which is why crawler access is the first technical check whenever you look at how models see a brand.

CompanyTraining crawlerSearch or retrievalFetches for a user
OpenAIGPTBotOAI-SearchBotChatGPT-User
AnthropicClaudeBotClaude-SearchBotClaude-User
GoogleGoogle-Extended (a robots.txt token for Gemini training and grounding)Googlebot (AI Overviews and AI Mode use the Search index)n/a
PerplexityNone declared (PerplexityBot is not used to train foundation models)PerplexityBotPerplexity-User (generally ignores robots.txt)

Sources: OpenAI crawler overview; Anthropic help center (updated April 7, 2026); Google common crawlers (updated October 15, 2026); Perplexity crawler docs. Checked October 27, 2026.

Google-Extended deserves a note because it confuses people: Google says it does not affect inclusion or ranking in Google Search. Blocking it is about Gemini training and grounding; AI Overviews and AI Mode draw on the regular Search index. A typical robots.txt for a business that wants to be found, while keeping its options open on training, looks like this:

# Search and retrieval: allow, so assistants can cite you
User-agent: OAI-SearchBot
Allow: /

User-agent: Claude-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

# Training: a business decision, made separately
User-agent: GPTBot
Allow: /

User-agent: ClaudeBot
Allow: /

User-agent: Google-Extended
Allow: /

Example only. Change Allow to Disallow for any training crawler you want to opt out of; it will not remove you from that company’s search answers.

Opting out of training has a cost: the next model will know less about you from its own memory, and its first search about your category may not include your name. The free AI crawler checker from Slibur.com, a tool from the same owner as Vértigo, shows what your robots.txt currently allows.

Does the model know you? 8 brand questions checked

To see memory and retrieval at work, we asked eight “what is” and “who makes” questions about brands of very different sizes: household names, two AI startups, and Slibur.com, the young AI visibility tool mentioned above. Each ran once on chatgpt.com, logged out, from the US, and once on Google with the AI Overview captured. The Slibur.com question ran on November 6, after the tool took that name; the rest ran on October 27.

Our own check, run on October 27, 2026. AI answers change over time, so the same question may get a different answer today.

QuestionChatGPT searched?What ChatGPT citedWhat the AI Overview cited
what is notionNoA link to notion.com, from memoryMerriam-Webster and Dictionary.com: it defined the word
what is sliburNoNothing: “Did you mean Sliver?”Sibur’s own sites, on the Russian petrochemical company
what is hubspotYesThree HubSpot pagesA Reddit thread, HubSpot, a software company’s blog
what is profound aiYesProfound’s site plus two investors’ portfolio pagesA Reddit thread, Profound, its LinkedIn page
what is peec aiYesPeec AI’s fact sheet for AI assistants, G2, a glossary siteThe same fact sheet, a review blog
who makes instant potYesA press release, Wikipedia, the brand site, two explainer sitesWikipedia and Amazon
who makes kirkland signatureYesCostco pages, GOBankingRates, WTOP, Simply RecipesTwo Facebook posts from a consumer news segment
who owns patagoniaYesPatagonia’s ownership page and its 2022 announcementPatagonia and CNBC

Vértigo Marketing, October 27, 2026 (the Slibur.com question, November 6, 2026). One run per question; answers vary between runs.

What the answers showed

  • Memory wrote the searches. For Patagonia, ChatGPT’s search already named the Holdfast Collective and the Patagonia Purpose Trust before a single result came back. For Instant Pot it searched for the current owner of Instant Brands. The model knew the entity and used the web to confirm and update it.
  • An unknown name got autocorrected. With no search at all, ChatGPT asked whether Slibur.com meant Sliver, and Google’s AI Overview answered about Sibur, a Russian petrochemical company. A brand one letter away from a better-known word or name has to work harder to become its own entity.
  • The same name meant different things. ChatGPT took “what is notion” to mean the app and answered from memory. Google’s AI Overview defined the English word. Which reading wins depends on how strongly each system ties the name to the company.
  • Brand sites were the main source. Every one of the seven answers with citations cited the company’s own website. Patagonia’s ownership page and Costco’s Kirkland Signature page did the job Wikipedia was once expected to do.
  • Investors, directories and press releases filled the gaps. Two investors’ portfolio pages vouched for Profound, G2 did the same for Peec AI, and a press release about new leadership backed up the Instant Pot answer.
  • One brand wrote a page for the machines. Both ChatGPT and the AI Overview cited a page on Peec AI’s site that describes the company and says it is intended for AI assistants. One case proves little, but it shows a model will use a clear, dated fact sheet when it finds one.

Wikipedia was a minor player. ChatGPT cited it once in these eight answers, and once in the 149 answers we saved that day across all our English guides. Semrush measured a similar slide in 2025: in its data, Wikipedia went from roughly 55% of ChatGPT responses to fewer than 20% by mid-September. Wikipedia’s weight in training is a separate matter, as the GPT-3 recipe shows.

Your brand as an entity: Wikidata, Wikipedia and the Knowledge Graph

An entity is a thing a system can tell apart from everything else with the same name. Search engines have worked this way since Google launched its Knowledge Graph, a story told in our guide to what AEO is, and language models absorb the same structured facts. For entity SEO, the work is making sure the web describes your company as one clearly defined thing.

Google’s Knowledge Graph
Google generates knowledge panels automatically from information across the web; its help pages explain how to claim one, not how to create one. If your company has a panel, an official representative can get verified and suggest changes. If it has none, the signals below are what eventually create one.
Wikidata
Wikidata accepts an item if it meets one of three criteria: it has a link to a Wikimedia project page, it describes a clearly identifiable entity that can be documented with serious, publicly available references, or it meets a structural need. A company with real press coverage can usually qualify under the second. Keep the item factual: name, website, founding date, headquarters, industry, official social accounts.
Wikipedia
A company article needs “significant coverage in multiple reliable secondary sources that are independent of the subject”. Press releases, sponsored posts and coverage based on press releases do not count. If you work for the company, the conflict of interest guideline says you are strongly discouraged from editing the article directly; propose changes on its talk page instead, and paid editors must disclose who pays them.
Your own structured data
Organization markup on your home page states your official name, logo, URL and contact details. Its sameAs property lists the pages on other websites that describe the same organization, such as your Wikidata item and LinkedIn page, which tells machines that all of them are you.
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "name": "Example Analytics",
  "alternateName": "Example Analytics Inc.",
  "url": "https://www.example.com/",
  "logo": "https://www.example.com/logo.png",
  "description": "Example Analytics makes inventory software for independent bike shops.",
  "foundingDate": "2019",
  "sameAs": [
    "https://www.wikidata.org/wiki/Q00000000",
    "https://www.linkedin.com/company/example-analytics",
    "https://www.crunchbase.com/organization/example-analytics"
  ]
}

Organization JSON-LD with sameAs. Replace every value with your real ones; list only profiles you control or that describe you accurately.

Fixing name collisions

If your name is also a common word, a person or another company, consistency is your main tool. Use the same full name and the same one-line description everywhere: your site, social profiles, directory listings, press kit and author bios. “Slibur.com, an AI visibility tracker” is harder to mistake for Sliver or Sibur than “Slibur.com” on its own. Over time, the pairing of name and category is what a model learns.

Third-party mentions: what models read about you

Mentions on other people’s sites tell a model what you are and whether you matter, and they show up in both training and retrieval. The best large dataset on this so far comes from Ahrefs, which compared 75,000 brands against how often Google’s AI Overviews mentioned them. Published May 26, 2025, the study found the strongest correlation with branded web mentions (0.664), well ahead of Domain Rating (0.326) and the number of backlinks (0.218).

Ahrefs is careful to say correlation is not causation, and its sample only included sites with a Domain Rating above 40. Still, the gap is large: brands in the top quartile for web mentions averaged 169 AI Overview mentions, against 14 for the next quartile down. Being talked about seems to matter more than being linked to.

Our brand check pointed the same way, with a twist: the mentions that got cited were rarely glamorous. Investors’ portfolio pages, a software directory, a press release on a newswire, explainer articles from personal finance and recipe sites. These are the places that describe what a company is in plain words, which is exactly what a “what is” question needs.

Source typeExample from our check or the studiesWhy it helps your entity
Directories and review platformsG2 cited for Peec AIA neutral one-line description plus category
Investor and partner pagesSequoia and Comcast LIFT Labs pages cited for ProfoundThird-party confirmation that the company exists and what it does
Newswires and pressA press release cited for Instant Pot; CNBC in the Patagonia overviewDated facts about ownership, leadership and launches
Explainer articlesGOBankingRates and Simply Recipes cited for Kirkland SignatureAnswers to “who makes” and “who owns” questions
ForumsReddit threads in the HubSpot and Profound overviewsReal users describing what the product is like
Reference worksWikipedia cited for Instant PotHeavily weighted in training; occasional live citation

Vértigo Marketing check of October 27, 2026, and Ahrefs (May 26, 2025).

Forums deserve their own strategy. Google’s AI Overviews cite Reddit far more often than ChatGPT does, which our guide to Reddit SEO covers with data and the rules for taking part.

Content a language model can quote

A model lifts sentences that answer a question on their own. Since your own site was cited in every answer with sources in our check, the pages that describe your company are entity assets, not housekeeping. Write them for a reader that will quote one paragraph and leave.

  1. Open with a definition. The first sentence of your home and about pages should say what the company is, for whom and where: “Example Analytics makes inventory software for independent bike shops in the US.” That sentence is what you want repeated.
  2. Answer the entity questions on your own site. Who owns the company, who founded it and when, where it is based, who makes the product. Patagonia has a page called Ownership. Most companies leave these answers to strangers.
  3. Date your facts. Prices, plan limits, team size and locations change. Put “updated” dates next to them, as Peec AI’s fact sheet does, so a model can tell current from stale.
  4. Publish numbers nobody else has. Your own survey, usage figures or benchmarks give journalists and bloggers something to cite, and every one of those citations is a third-party mention.
  5. Use tables for comparisons and specs. Rows and columns survive extraction better than a paragraph that buries four facts in one sentence.
  6. Keep every description consistent. If your site says you were founded in 2019 and LinkedIn says 2020, a model has to guess. Fix contradictions before writing anything new.

One file you will hear about is llms.txt, a plain-text summary of a site for language models. Google says its AI features need no special files, so treat it as optional; if you want one, Slibur.com has a free llms.txt generator.

A practical entity SEO plan for the next 90 days

Most companies can cover the basics in a quarter. The order matters: find out what the models believe first, fix what you control second, then work on what others say about you.

WeeksFocusTasks
1 to 2BaselineAsk ChatGPT, Gemini, Claude and Perplexity “what is [brand]”, “who makes [product]” and your top category questions, logged out. Note whether each one searched, what it got wrong and which sources it cited.
3 to 4AccessCheck robots.txt, firewall and CDN rules for the search crawlers. Decide on training crawlers deliberately. Make sure key pages are indexed in Google and Bing.
5 to 6EntityWrite one standard description. Add Organization markup with sameAs. Create or correct the Wikidata item with real references. Align LinkedIn, Crunchbase, directories and social bios.
7 to 9PagesRewrite the about page with a definition sentence and dated facts. Add ownership, founder and location answers. Publish one data-led page worth citing.
10 to 12MentionsUpdate listings on the review platforms your buyers use. Pitch the data page to trade press. Answer questions where your buyers already talk.
13Re-checkRepeat the week 1 questions and compare. Keep a monthly log from here on.

Vértigo Marketing. Training memory only changes with new models, so judge the first 90 days on retrieval: searches, citations and accuracy.

Checking by hand works for a handful of questions. For more, a tracker saves time: Slibur.com offers a 7-day free trial, no card required, that records each AI answer and the domains it cites, and our roundup of LLM tracking tools compares others.

Vértigo Marketing runs GEO and SEO for companies in the US, the UK and Canada. Get a free review

If you would rather hand it over, our GEO services for US companies include the entity and mention work described here, and our SEO services cover the rankings that retrieval depends on.

Frequently asked questions

What is entity SEO for LLMs?

The work of making language models such as GPT, Claude and Gemini recognize your company as one clearly defined thing and describe it accurately, from what they learned in training and from what they find when they search. It overlaps heavily with GEO, AEO and LLM SEO, with more weight on entity signals.

Is entity SEO for LLMs different from GEO?

The goals overlap almost completely. GEO usually focuses on citations in AI search answers; entity SEO puts more weight on what the model itself knows about your brand, which means entity signals and the text it was trained on, as well as live retrieval.

Can I get my brand into ChatGPT’s training data?

There is no submission form. OpenAI’s GPTBot crawls the public web for training, so allowing it and being described widely and consistently on public pages is what makes your brand more likely to appear in future models. Changes only show up after a new model with a later cutoff ships.

Do I need a Wikipedia page for entity SEO?

No. Wikipedia requires significant coverage in several independent, reliable sources, and it discourages people from writing about their own company. A factual Wikidata item, consistent profiles, Organization markup and real press coverage do much of the same work.

How long does it take for an LLM to learn about my brand?

Through search, as soon as your pages are crawled and rank for the questions people ask, often days or weeks. In the model’s own memory, only when a new model trained on later data is released, which can take many months.

Does llms.txt help models understand my brand?

Google says its AI Overviews and AI Mode need no special files or schema beyond normal Search requirements. An llms.txt file is cheap to create and does no harm, but treat it as optional rather than a ranking lever.

How can I see what ChatGPT says about my brand?

Ask it logged out, in a fresh chat, both “what is [brand]” and the category questions your buyers ask. Note whether it searched and which sites it cited, then repeat monthly, by hand or with an AI visibility tracker.

Cookies and preferences

Choose what Vértigo may remember in this browser. You can change your choice at any time.

NecessaryThey store your cookie choice. Always active
Read the cookie policy