GEO: what makes a page citable by AI

Where the term comes from

It is a different kind of visibility from ranking: you do not occupy a position in a list, you become the material the answer is built from. And it is worth knowing where the word comes from, because it changes how you read the numbers in circulation.

GEO did not start in marketing. It comes from a November 2023 academic paper, GEO: Generative Engine Optimization by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, later accepted at the KDD 2024 conference.

The authors built GEO-Bench, a benchmark of ten thousand queries drawn from nine different collections, each accompanied by Google's top five results. The paper's main results are measured there, on a generative engine reconstructed in the lab: those five results passed to GPT-3.5. The authors then repeated the test on Perplexity, a real commercial engine, and obtained improvements of the same order.

The number that circulates most, 40%, should be read for what it is: a maximum, not an average. It is the gain of the best strategy on one specific metric, and other strategies tested in the same study make visibility worse rather than better. The authors themselves note that effectiveness varies by sector and that domain-specific methods are needed.

GEO, SEO and AEO: what actually differs

The three acronyms get presented as separate disciplines. They are better understood as three different objectives on the same work.

There is only one useful distinction: in SEO the objective is that a person clicks, in GEO that a system uses your text. But the means overlap almost entirely, and Google explains why.

What Google actually says

This is the part most guides leave out, and the most useful one because it saves work.

On its page about AI features and your website, Google writes that there are no additional requirements for appearing in AI Overviews or AI Mode, and no special optimisations needed. It adds that you do not need to create new machine-readable files, AI text files or AI markup to appear in these features, and that no special schema.org structured data has to be added.

The eligibility conditions for being shown as a supporting link are three, and they are the usual ones:

• the page has to be indexed;
• it has to be eligible to appear in Search with a snippet;
• it has to meet Search's technical requirements.

Meeting them makes the page eligible; it does not guarantee the citation.

In its guide to optimising for generative AI, Google is equally explicit about what is not needed: an llms.txt file is not used by Search, there is no need to chunk content so the AI understands it, writing for machines does not help, and structured data is not a requirement for generative answers, though it remains useful for rich results.

What Google points to as decisive is something else: creating content people find unique, interesting and useful will likely influence your site's presence in generative AI search, over the long run, more than any other suggestion in that guide.

Translated: a good share of the “GEO techniques” sold as new are required by nobody. The real work is upstream, and it concerns what the page says and whether the page can be read.

The controls you have

They are the usual ones: robots meta tags, an HTML attribute, and a directive in robots.txt. It is worth knowing what each does, because they get confused often.

• nosnippet and max-snippet, robots meta tags, limit the previews Google may show. The robots meta tag documentation is explicit: nosnippet prevents content being used as direct input for AI Overviews and AI Mode.
• data-nosnippet is an HTML attribute and does the same thing to portions of text rather than to the whole page.
• noindex removes the page from the index, and with it any chance of appearing at all.
• Google-Extended is a robots.txt token, not a meta tag. It limits the use of content for training Gemini models and for grounding — retrieval from the Search index to the model in Gemini apps and Vertex AI. Google states that it does not affect a site's inclusion in Search and is not a ranking factor.

Why crawling comes first

None of this counts if the page is not read, and this is where the technical part is decided.

Answer systems largely work by retrieving documents at question time — what the literature calls RAG, retrieval-augmented generation: the system searches an index for relevant documents, reads them and builds the answer citing them. The typical chain is this, and every link depends on the one before it: the page is crawled and enters an index, it is retrieved at question time, and only then can it be cited. Some systems also fetch an address directly when the user asks for it, but the starting point does not change.

Three minimum technical conditions follow:

1. The answer's text is present in the initial document. Crawlers reach content in different ways, and the independent testing available shows that a fair number of AI system crawlers do not run JavaScript: what appears only after the code executes may not exist as far as they are concerned. It holds for Google too, where rendering happens in a separate queue after crawling.
2. The answers are not hidden behind an interaction. A questions-and-answers section that only opens on click is a technical problem before it is an editorial one, if the text is not in the HTML anyway.
3. No rule blocks the answer engines' agents. AI systems use their own crawlers, with names different from the search engines' and often with separate agents for training and for retrieval. They want watching in the logs and managing in robots.txt deliberately, blocking only what genuinely should not be read.

Where to start

The order matters, because the first two steps make the others possible and cost less than people expect.

1. Check that the pages can be read. Content in the initial HTML, no accidental block in robots.txt, no key answer hidden behind a click. Without this, everything else is pointless.
2. Check the preview controls. That nosnippet or max-snippet are not limiting the text you want cited, and that nobody has blocked the answer engines' agents by mistake.
3. Rewrite the openings. A definition or a direct answer in the first lines of the pages that matter most, in a block that stands on its own.
4. Build the basket of questions and take the first reading, before changing anything else. Without a starting point you will not know whether what you do next works.
5. Then the rest, which is the usual editorial work: covering the questions completely, with verifiable data and a structure where every section closes something.

None of these steps is a new technique. That is why Google says no special optimisations are needed.

What makes a page usable

A generated answer is assembled by taking verifiable statements from several sources. The pages that end up cited, in our experience, have one thing in common: they contain sentences that stand on their own, without the surrounding paragraph acting as context.

This changes how you write more than it sounds like it would. A definition at the opening, a number with its unit beside it and its source, a sentence that closes a question: those are extractable blocks. An argument that only closes at the end of the page is not.

What cited pages have in common

• An explicit definition in the page's first lines, written to hold up outside its context.
• Statements that stay true outside their paragraph, with no references like “as mentioned above” or “in this case”.
• Numbers carrying their source and the period they cover, because a number with no provenance is the first to be discarded.
• A structure where each section closes a single question, rather than touching five and closing none.

It is not only field experience: it is also what the study that introduced the term GEO measured. The three writing strategies with the largest gains turned out to be adding citations from authoritative sources, adding statistics in place of generic claims, and stating sources explicitly. Strategies of the opposite kind, such as stuffing keywords into the text, made visibility worse.

It is worth noticing that these are the same things that make a page readable to a person in a hurry. That is not a coincidence, and it is consistent with what Google says about helpful content: there is no writing for machines separate from writing for people.

How to measure it

Citations do not show up in classic ranking reports. For Search's AI features, Google states that traffic from AI Overviews and AI Mode is included in the Search Console performance report, within the “Web” search type: there is no separate report. For every other engine there is no report at all: citations have to be collected by querying them on the questions that matter and recording what happens.

The method matters more than the tool, because without a method the data is not comparable over time.

How to build the measurement

1. Define a fixed basket of questions, the ones your customers actually ask. Fixed is the important word: change the questions and you lose the comparison.
2. Repeat each question several times, not once. The same engine asked twice about the same question can cite different sources, and a single reading cannot tell a real improvement from random variation.
3. Record the same things every time: the question, the engine, the date, whether your page appears as a source, which other sites appear, and how the claim attributed to your page is worded.
4. Compare over time and against competitors. What tells you something is not the absolute number, it is how it changes after an intervention and how it sits against whoever is cited in your place.

The last thing on that list, the wording of the attributed claim, is the most underrated: reading how the engine rephrased your sentence tells you whether the page was understood, and often explains why it was not picked the next time.

Do citations bring traffic?

Less than a first position, and differently. There is no reliable public data on this, so what follows is an observation from people who do the work, not a number: someone arriving from a citation has already read an answer and is after verification or depth. Low volume, highly qualified.

Be wary of anyone promising traffic volumes from citations, because right now nobody can support them with data. The value is in appearing at the moment a person is deciding, and in doing so with a sentence attributed to your site rather than to a competitor.

Common questions