What GEO actually means, and how we measure it

Where the term comes from

The term was introduced in a paper by researchers from Princeton University and IIT Delhi, with co-authors then at Georgia Tech and the Allen Institute for AI, first posted in November 2023 and published at the KDD conference in 2024. It describes GEO as a way for content creators to improve their visibility in AI-written answers.

What the paper measured

The authors built a benchmark of 10,000 queries and tested nine ways of rewriting a web page. They measured visibility in two ways: how many words of the AI answer were attributed to a source, weighted by where in the answer they appeared, and a subjective score given by a language model. Adding quotations, adding statistics and citing sources did best, with gains of up to about 40% on their own benchmark. Keyword stuffing did not help.

Other names for the same work

You will also see answer engine optimization (AEO), AI SEO and LLM optimization. The names differ in emphasis, but in practice they describe the same work: being used by an AI system when it answers a question. Whether that work differs from ordinary SEO is a separate question, covered in Is GEO just SEO?.

Why a definition needs a measurement rule

"Visible in an AI answer" hides at least two different outcomes, and they rarely happen together. A company can be named in the text without its website being a source, and its website can be a source without its name appearing. Any GEO report that does not say which one it counts is not comparable with any other.

Two outcomes, counted apart

• Mentioned: the company's name appears in the text of the answer.
• Cited: a page of the company's website appears in the list of sources behind the answer.

In our own data, across 17,168 answers from five AI systems between 10 August and 11 September 2026, a company's website was cited 2,173 times, and in 74% of those citations its name did not appear in the text. The Dutch analysis is in Als AI je site gebruikt, valt je naam meestal niet.

When a brand counts as cited

A company counts as cited in an answer when at least one link in the answer's source list points to its website. "Its website" means the registrable domain, such as example.nl, so every subdomain and every page on that domain counts. The position we record is the place of the first matching link in the source list.

The rule in detail

• Domain, not page. We compare the registrable domain of each source link, lowercased and without www, with the company's domains. shop.example.nl/prices counts for example.nl; example.com counts only if the company has told us it owns it.
• Redirect links. Gemini returns its sources as redirect links on a Google domain. The real domain is in the title of the link, so for those we match on the title.
• Pages elsewhere do not count. A company profile on LinkedIn, a directory or a review site is not the company's website. If that page is the source, the company is not cited, even if the answer is about it.
• Read but not cited is separate. Some systems return pages they consulted but did not cite. Where they do, we record a match there in a separate field, and it never counts as a citation.

When a brand counts as mentioned

A company counts as mentioned when its name, or a name it is known by, appears as a whole word in the text of the answer after source labels and bare URLs are removed. Matching ignores capitals and accents. A name inside another word does not count; a name directly followed by a hyphen or a capital letter, as at the start of a compound, does.

The rule in detail

• Source labels removed first. Before we search, we delete source chips, link labels that consist only of a domain name, and bare URLs. A domain in a source label is not the answer telling the reader about the company.
• Whole words. A match needs a non-letter, non-digit character, or the start or end of the text, on both sides. A brand called "Safe" is not found in "safety".
• Compounds. German and Dutch glue words together. "SafeHome-Alarmanlage" and "SafeHomePakket" count as mentions of "SafeHome", because the name is followed by a hyphen or a capital letter. A name at the end of a compound, as in "SuperSafeHome", does not count.
• Short names. A name shorter than three characters is not searched for at all, and a three-letter name counts only as a separate word, never as part of a compound.
• Only listed names. We search for the company's name and the aliases on its profile. A misspelling that is not on that list does not count.

Where the line falls

Situation in the answer	Mentioned	Cited
Name in the text, own site not among the sources	Yes	No
Own site among the sources, name not in the text	No	Yes
Domain appears only as a source label, such as [example.nl]	No	Yes, if the link is in the source list
A subdomain, such as shop.example.nl, among the sources	—	Yes
The company's LinkedIn page among the sources, name not in the text	No	No
Brand "Safe" and the word "safety", in lower case, in the text	No	—
Brand "SafeHome" and the word "SafeHome-Alarmanlage"	Yes	—
Brand name that is also an ordinary word, used as that word	Yes, a known false positive	—

How often and where we apply it

For each company we put 100 fixed buying questions to ChatGPT, Gemini, Perplexity and Google AI Overviews three times a week, and to Claude three times a month. We report each outcome per system as a share of answers, with a 95% margin per buying question, never as a single answer, because the same question gets a different answer from one run to the next.

Where the rule falls short

• Not the logged-in app. We query ChatGPT, Gemini, Perplexity and Claude through their interfaces for developers with web search switched on, and retrieve Google's results pages through a SERP data provider for AI Overviews. A logged-in user with a history and personal settings can see a different answer.
• Literal names. Name matching finds only the names we list. It misses misspellings and can count an ordinary word that happens to be a brand name.
• Presence, not quality. Neither rule says whether the answer is correct, whether it recommends the company, or whether anyone clicked.

Frequently asked questions

Is GEO the same as SEO?

They share most of their foundations: a page that a search engine cannot read, an AI system with web search cannot read either. They differ in what is measured. SEO measures a position in a list of results; GEO measures whether you appear in a single written answer.

Is GEO the same as AEO?

In practice, yes. Answer engine optimization is an older term, used before AI-written answers for featured snippets and voice assistants; GEO comes from the 2023 paper about AI-written answers. Both are now used for the same work.

Why not combine mentions and citations into one score?

Because they answer different questions. A mention tells you what the reader sees; a citation tells you which page the AI system relied on. Adding them up hides which of the two is missing, and that decides what you would change.

Does a mention without a link have any value?

It can. A reader who sees your name may search for you later. But it does not show that the AI system used your website, and it is harder to influence through your own pages.

Where we stand

AIGeschikt applies these rules every week to the buying questions of Dutch companies and reports mentions and citations separately, per AI system, with a margin per buying question. That measurement decides what we write and change on a client's website. How the service works is described on AI-zichtbaarheid laten meten (in Dutch).