Generative engine optimization (GEO): what the evidence says

Short answer: Generative engine optimization (GEO) is making your pages easy for AI answer engines to find, understand and cite. Google says that for its AI features this is still SEO. The vendor docs start in the same place: the engine's own crawler or index has to be able to reach your page (Google and OpenAI state it as a condition; Perplexity recommends it and Anthropic says blocking may cost visibility). After that, the engine runs searches and your page has to be a good answer to them. The evidence for specific tactics is thin. One benchmark found that adding sources, quotes and statistics raised visibility in a test engine. None of the vendors we read publishes ranking factors or weights for citations, so anyone quoting the weights is guessing.

This guide sorts GEO advice by the evidence behind it: documented by vendors, tested, only observed, or just repeated because it sounds right. Then measurement, and a tactic Google says violates its spam policy.

What GEO means

The term comes from a 2023 paper, GEO: Generative Engine Optimization, by Pranjal Aggarwal and co-authors from Princeton, IIT Delhi and elsewhere, published at KDD 2024. They describe generative engines as search systems that retrieve sources and use a language model to write an answer from them. GEO is changing your content so it's more visible in those answers. They built a benchmark of 10,000 queries to test it.

Since then the label has stretched to cover any work aimed at appearing in AI answers:

These are conventions, not standards, and the lines blur. Our glossary draws them the same way for GEO and AEO. Google's guide to generative AI search names both terms and says that, from Google Search's perspective, this work is "still SEO". Microsoft's guidance on AI search answers opens by saying that whether you call it GEO, AIO or SEO, visibility is what counts.

How GEO relates to SEO

The gates are the same. The prize is different.

Every engine in the next section starts with its own crawler or index reaching your page. That's SEO's first job too: be crawled, be indexed, be eligible to show.

The prize changes. In classic search you compete for a position in a list. In an AI answer you compete to be one of the sources the system reads and cites, often for a sub-question you never targeted. The GEO paper makes the point itself: average ranking measures visibility for a list of links, but an answer embeds sources at different lengths and positions, so the authors built new metrics.

The data fits. Ahrefs compared AI Overview citations with the regular results for the same queries (863,000 keywords, 4 million cited URLs, published 2 March 2026). About 38% of cited URLs ranked in the top 10 for the same query, and roughly 31% each ranked 11 to 100 or outside the top 100. Its July 2025 study found 76%. Ahrefs suggests fan-out searches now play a bigger role, and notes it has since improved how it detects citations, so the comparison isn't exact. A study of 15,000 prompts, with data collected in July 2025, found that on average about 12% of links cited by ChatGPT, Gemini, Copilot and Perplexity also ranked in Google's top 10 for the same prompt. Perplexity was highest at 28.6%.

Ranking helps. It isn't the whole story.

How AI answer engines choose sources

Most products here follow a similar loop: decide whether to search, run one or more searches, read what comes back, write an answer, cite some pages. Vendors document the first half of that loop. None publishes how it weighs one page against another.

Google lists no requirement beyond that eligibility. OpenAI says ChatGPT may also follow up with more specific queries, and that it ranks results on multiple factors with no guaranteed placement. Anthropic says blocking Claude-SearchBot or Claude-User may reduce your visibility. Gemini Apps can also ground answers in Google Search, and Google-Extended controls that and training; see our robots.txt and AI crawlers guide.

Two things stand out.

Several searches per question. Google calls it query fan-out. In OpenAI's example, a question about the latest drugs that target CCR8 for cancer might first be searched as "CCR8 immunotherapy drug development 2025", with more specific follow-ups after the first results. The query that matters is often not the one the user typed. We go deeper on ChatGPT in how to rank in ChatGPT.

No published weights. OpenAI names "multiple factors" and lists none. Microsoft says there's "no secret strategy" for being selected. For Google, see AI Overviews and AI Mode: what founders should change.

What moves citations: the evidence, sorted

Each common claim, with the best evidence we could open.

The best-known benchmark study

The GEO paper rewrote web pages with a language model using nine methods and measured how much of a generated answer was credited to the rewritten page. Its test engine passed the top five Google results for each query to GPT-3.5 Turbo. Adding citations, quotations from credible sources or statistics raised a position-weighted visibility score by 30 to 40% and a model-rated impression score by 15 to 30%. Improving fluency gave 15 to 30% too. Keyword stuffing offered "little to no improvement". Which method helped most varied by topic. In a test where every source in an answer was optimized at once, citing sources raised the fifth-ranked page's visibility by 115.1% and cut the top-ranked page's by 30.3% on average. A 200-query repeat on Perplexity.ai again put quotations and statistics on top, while citing sources scored below no optimization on the model-rated measure.

The authors read that as good news for small sites, but the limits matter. It was a 2023 to 2024 lab on an older model, with a simplified engine and scores partly rated by a model. The authors say methods "may need to adapt over time" and didn't test effects on search rankings. We read it as a reason to put real sources, quotes and numbers into pages you'd write anyway, not to pad pages with invented statistics.

What the observational data adds

The Ahrefs figures describe overlap, not cause. A University of Toronto preprint, Generative Engine Optimization: How to Dominate AI Search (September 2025), compared Google with ChatGPT, Perplexity, Gemini and Claude on ranking-style prompts, such as asking for the top 10 brands in a category. For software products in the US, the domains cited by OpenAI's search-enabled GPT-4o model were 72.7% earned media (independent media, review and comparison sites), against 45.4% in Google's top ten. A model classified the domains, and the tests ran in 2025.

Google says its AI features can reflect what blogs, videos and forums say about a product, and that chasing inauthentic "mentions" doesn't help. Honest coverage survives both.

Be wary of GEO services

Google's guidance on third-party SEO tools and advice names tools that promise "AEO" or "GEO" improvements, and says outside tools can't see Google's internal ranking data or guarantee performance. That includes ours.

A GEO plan for a small site

1. Let the search crawlers in. Check OAI-SearchBot, PerplexityBot, Claude-SearchBot and Bingbot with the AI crawler checker, then your CDN or firewall.
2. Check Google eligibility. Important pages indexed, no accidental nosnippet, Search generative AI set to Include. Steps are in the AI Overviews guide.
3. Write 20 to 30 buyer prompts using the method in measuring AI visibility without paid tools, and run them before you change anything, for a baseline.
4. Cover the sub-questions as sections of one page. Open with a direct answer, then the evidence.
5. Add what only you have: numbers, screenshots, mistakes and sourced facts. That's Google's non-commodity content. Sourced facts also overlap with the GEO paper's best-performing methods: citations, quotations and statistics.
6. Get listed honestly on third-party pages: reviewed directories, comparison sites, communities where you take part for real. A free launch on LaunchRanked gives you a permanent product page.
7. Re-run the prompts on the same schedule, such as every two weeks, and compare like with like.

Don't build a page for every fan-out query

Once people learn that AI search splits a question into sub-queries, the tempting plan is one page per sub-query. Google's guide addresses that. Creating "separate content for every possible variation of how people might search", fan-out queries included, primarily to manipulate rankings or AI responses "violates Google's scaled content abuse spam policy" (see scaled content abuse). It adds that a high quantity of pages doesn't make a site higher quality or more relevant. Google's documentation update of 15 May 2026 also clarified that its spam policies apply to generative AI responses in Google Search.

We learned the practical version on Versely, our own product: about 2,500 posts in one month, many crawled but left out of the index. Pages Google hasn't indexed can't be supporting links in AI Overviews or AI Mode.

Use the sub-questions as sections of one strong page. Give one its own page only when people search for it alone and you can answer it better than a section can: pricing, a comparison, an alternatives list. The query fan-out generator predicts the sub-queries for a topic so you can check your page against them. It's a prediction. Search Console's AI report doesn't list queries, so you can't see the real ones.

How to measure GEO

You can't see most of the searches an engine runs for you, and answers change between runs, so measure from several angles.

Google. Search Console's Generative AI performance report was rolled out to all sites on 31 August 2026. It reports impressions from AI Overviews and AI Mode by page, country, date and device. If two links to your site appear in one AI response, the chart counts one impression. Clicks from AI features also land in the normal Performance report, under the Web search type. No report? Google says the site usually hasn't had enough impressions yet, or is excluded.

Bing and Copilot. Bing Webmaster Tools' AI Performance report, in public preview since February 2026, shows citations in Copilot and Bing's AI summaries, which pages are cited, and a sample of the grounding queries the AI used. A June 2026 update added citation share.

ChatGPT. OpenAI says ChatGPT search adds utm_source=chatgpt.com to referral URLs. Filter your analytics by that source.

Prompts. A fixed list, run on a schedule in a clean session, more than once per prompt. Count mentions, citations and errors. Our free method takes about an hour a fortnight.

Our free tools. The AI visibility checker writes four questions a buyer might ask (none contain your brand name), asks ChatGPT, Claude, Perplexity and Gemini with web search on, and shows which ones name and link you. It's one free check per account every 30 days, with sign-in. The AI Overview checker shows whether Google gives an AI Overview for one keyword, which pages it lists and whether your domain is one.

Autopilot. Our AI SEO Autopilot tracks 25 prompts across ChatGPT, Perplexity, Gemini and Claude every two weeks, next to the content and Search Console work meant to move them. The free trial covers Perplexity and Gemini. By Google's definition we're a third-party tool: we can't see Google's internal data, we sample what assistants answer, and answers vary. We don't promise citations.

The short version

1. GEO is mostly SEO with a new place to be cited.
2. Let each engine's search crawler in, and be indexed where it searches.
3. Answer the sub-questions on one strong page, not a page per query.
4. Put in sourced facts, real quotes and your own numbers.
5. Ignore guaranteed-citation pitches and anyone selling llms.txt as a ranking lever.
6. Measure with Search Console, Bing's report, ChatGPT referrals and a fixed prompt list. Plain-word definitions: LLM citation and grounding.