Ask an AI assistant a question and a retrieval system quietly reads a handful of web pages, a model stitches the useful parts into one answer, and a few of those pages earn a named citation. GEO is the work of being one of those pages. The term was coined in a 2023 academic paper, "GEO: Generative Engine Optimization" by Pranjal Aggarwal and colleagues (later published at KDD 2024), which tested ways of changing a page so that a generative engine was more likely to surface it. Nearly three years on, the discipline has settled into a clearer shape: a thin layer of genuinely new practice on top of sound SEO, surrounded by a lot of folklore.
What does GEO actually mean?
GEO is optimisation for the point where a model assembles an answer from several sources. The unit of success changes from a ranked link to a mention or citation inside generated text. Closely related terms, answer engine optimisation (AEO) and "AI SEO", describe largely the same job, and in practice the labels are used interchangeably. What distinguishes GEO from older optimisation is the reader: a retrieval system and a language model that must find your page, extract a passage with confidence, and attribute it correctly, often for a user who never clicks.
What does the research say works?
The foundational paper reported that adding citations, quotations and statistics to content were the most effective content-level changes it tested, with visibility gains of up to 40 per cent in generative engine responses on its benchmark, GEO-bench. It also found that the effect varied by domain, so a tactic that lifted one subject area did little for another. Two cautions apply. The results come from a research setup with a particular engine and query set, not from a live commercial assistant, and the "up to 40 per cent" is a ceiling, not an average. A second body of research points to the structural reality that AI answers lean heavily on a narrow set of authoritative sources such as Wikipedia, major news outlets and government sites, which is why smaller publishers often find citations harder to win than the tactics alone would suggest.
What does Google say about GEO?
Google has now addressed the question directly. Its Search Central guide, "Optimizing your website for generative AI features on Google Search" (last updated July 2026), says optimising for generative AI search is optimising for the search experience, and so is still SEO. The same guide lists several things you do not need to do to appear in AI Overviews and AI Mode: you do not need new machine-readable files, AI text files, special markup or Markdown, structured data is not required for generative AI search (though it remains useful for rich results), and there is no need to break content into tiny pieces. It recommends distinctive, non-commodity content, clear crawlability and indexing eligibility, and monitoring through the Generative AI performance report in Search Console.
This matters for how you read the rest of the GEO advice market. Much of it is aimed at all assistants, not only Google, and Google's guidance covers only Google's features. But it is a useful brake on the tactics that promise the most for the least effort. If you are weighing a specific tactic, our explainers on whether llms.txt is worth using and whether schema markup helps AI citations go into the detail.
What makes content more likely to be cited?
Across the research and the platforms' own guidance, a short list of properties recurs.
- A direct, extractable answer. A page that states the answer to its own question near the top, in self-contained sentences, gives a retrieval system something to lift without rewriting.
- Specific, sourced facts. Named statistics and quotations with clear provenance are what the original GEO paper found most useful, and they give a model something verifiable to attribute.
- Authority and clear authorship. Systems favour sources they can recognise, so a consistent entity (who you are, what you cover, who wrote the piece) helps.
- Freshness where facts change. For topics with moving figures, a visibly maintained page beats a stale one.
- Crawlability. If the relevant crawler cannot fetch the page, none of the above matters. That means checking robots.txt and bot-management rules for the search and user-initiated agents of each assistant, not only the training crawlers.
Notice what is not on the list: keyword density, hidden prompts to the model, or bulk-generated pages. Google's guide explicitly warns against inauthentic mentions and spam, and assistants that rely on web search inherit those systems' defences.
How do you measure GEO?
The honest answer is that measurement is still probabilistic. A late-2025 study by Rand Fishkin of SparkToro and Patrick O'Donnell of Gumshoe, covering 2,961 prompts run 60 to 100 times each across ChatGPT, Claude and Google's AI tools, found that the same list of brands came back in fewer than one response in a hundred, and the same list in the same order in fewer than one in a thousand. The more stable signal was how often a brand appeared across many runs, not where it sat in any one answer. Fishkin's conclusion was blunt: any tool that reports a ranking position in AI is not to be trusted.
The practical metric stack therefore centres on frequency over a large, fixed prompt set. Track mention rate (how often your brand appears for category questions), citation rate (how often a specific claim is attributed to your page), and share of voice against named competitors, measured per assistant and repeated over time. Pair that with first-party evidence: AI referral visits in your analytics, and server-side logs of which agents are actually fetching your pages. For tool choices and their limits, see our comparison of AI visibility tracking tools; for turning the numbers into a business case, see how to measure GEO return on investment.
Where does GEO stop for publishers?
GEO answers one question: will an AI system find, trust and cite your content? It does not answer the second one: do you earn anything when it reads it? A Live Search Agent can fetch and read a publisher page to answer a user, and because it typically does not execute JavaScript, that read leaves no impression for a traditional ad stack to monetise and often nothing in client-side analytics. Citation is a credit line, not a payment.
That is the split blankspace works on. GEO is the comprehension side, increasing the odds your pages are the ones agents retrieve. blankspace sits on the compensation side, at the CDN edge, where those retrievals can be detected and, where a publisher chooses, monetised. The two are complementary: strong GEO raises the number of reads, and monetisation determines what each read is worth.
Frequently asked questions
Is GEO the same as AEO or AI SEO?
Largely, yes. GEO, answer engine optimisation and AI SEO all describe making content easier for AI systems to retrieve and cite. Some practitioners separate them by surface (chatbots versus AI summaries inside search), but the underlying tactics overlap almost entirely.
Does GEO replace SEO?
No. Google's own guidance says optimising for its generative AI features is still SEO, and AI answers built on web search draw on the same crawled and indexed pages. GEO adds extraction, attribution and cross-assistant presence to a sound SEO base.
Does llms.txt or schema markup improve GEO results?
Google says neither is required to appear in its generative AI features, and it describes no special markup for them. Structured data remains worthwhile for rich results and entity clarity, but it is not a shortcut to citations.
How do I know if AI assistants are citing my content?
Run a fixed set of category questions across the assistants your audience uses, repeat them many times, and record how often you are named. Because answers vary run to run, judge frequency over many samples rather than a single result, and check your server logs and referral data for agent fetches.
How does GEO relate to earning revenue from AI traffic?
GEO affects whether assistants read and cite your pages. It does not pay you when they do. Monetising the read is a separate step, handled at the infrastructure layer by detecting agent traffic and attaching value to it.