What is generative engine optimization (GEO)?

What is generative engine optimization?

Generative engine optimization, also called LLM optimization or LLM SEO, is the set of practices that make a page, a brand or a fact more likely to show up inside an answer written by an AI system. A "generative engine" is any product that retrieves information and then has a large language model write a response, usually with links to the sources it used.

SEO competes for a slot in a ranked list of links. GEO competes for two other things: a place among the handful of sources an answer is built from, and your brand's name in the answer text. Those are separate outcomes. A review site can be the cited source for an answer that names you, and your page can supply a fact in an answer that recommends a competitor, so you track them separately.

The audience is big enough to justify the work. Alphabet reported in July 2025 that AI Overviews had more than 2 billion monthly users17 (Alphabet Q2 2025 remarks), and OpenAI said in October 2025 that ChatGPT had 800 million weekly users18 (TechCrunch). More numbers are on our AI search statistics page.

Where did the term GEO come from?

It came from the paper "GEO: Generative Engine Optimization", first posted to arXiv on November 16, 2023 and accepted at KDD 20241. The six authors are Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. The author block lists Princeton University and IIT Delhi, with two authors listed as independent researchers.

The authors treated GEO as a black-box problem. Publishers can't see how an engine picks sources, so they need edits that raise visibility without access to the engine's internals. To test edits, the team built GEO-bench, about 10,000 queries drawn from nine datasets, each paired with real web sources.

What did the GEO study actually find?

Adding evidence beat rewriting for style, and classic keyword tactics backfired. The team applied nine rewriting methods to one source at a time and measured how much of the generated answer drew on that source1.

They used two metrics. Position-Adjusted Word Count counts the words in the answer attributed to a source, weighted toward sources cited earlier. Subjective Impression is a model-graded score for relevance, influence, uniqueness and how likely a reader is to click. Both baselines were 19.3 on GEO-bench.

The gains were largest for sources that ranked lower in the underlying search results: citing sources lifted visibility for fifth-ranked pages by 115%. The best method also varied by domain. Statistics helped most on law and government queries, quotations on history and society queries.

Combining methods helped a little more. On a 200-query subset, fluency optimization paired with statistics addition was the best pair, beating any single method by more than 5.5% on the word count metric1. The effects also held on a live engine: on Perplexity, quotation addition raised the word count metric from 24.1 to 29.1, and statistics addition raised the impression metric from 24.7 to 33.9.

Keep the limits in view. The study changed one source at a time on a fixed benchmark, engines have changed a lot since 2023, and the authors didn't measure effects on traditional rankings. Use the percentages to rank tactics against each other. They won't predict the lift on your own site.

What does a GEO rewrite look like in practice?

Here is one paragraph rewritten using the methods that scored best in the study. The "before" is a typical draft we wrote for this example; the "after" uses real, linked data from Pew Research Center14.

Before

AI Overviews are changing everything about search. Users now get their answers directly from AI and click on websites far less often, which is a huge challenge for marketers. To stay visible in AI search, brands need an AI search strategy, AI SEO content and AI-optimized pages that can appear in AI Overviews.

After

Google searches that show an AI summary send far fewer clicks to websites. In Pew Research Center's tracking of 900 US adults across 68,879 searches in March 2025, people clicked a traditional result in 8% of visits to a results page with an AI summary, against 15% when there was none. Clicks on the sources inside the summary were rarer: 1% of visits. In Pew's words, "users very rarely clicked on the sources cited." For a brand, being named inside the summary is an outcome worth tracking on its own, separate from the click.

What changed, and which finding each edit maps to:

The "after" is also the more quotable version. An engine assembling an answer to "do AI Overviews reduce clicks?" can lift its second sentence whole, with the source attached. The "before" offers nothing an engine can't write itself.

How do generative engines retrieve and cite sources?

For fresh or factual questions, most engines use retrieval-augmented generation: search an index, pull candidate passages, and have the model write an answer grounded in them, citing some. A page that is never retrieved can't be cited, so crawlability and search visibility still come first.

The search step rarely uses your exact words. Google says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources2 (Google). OpenAI's help center says ChatGPT search typically rewrites a question into one or more targeted queries sent to its search partners4 (OpenAI).

Fan-out changes which pages win. Ahrefs found in July 2025 that 76% of pages cited in AI Overviews ranked in Google's top 10. In a March 2026 update covering 4 million cited URLs, that share fell to 37.9%, which Ahrefs attributes to sources found through fan-out searches13 (Ahrefs). So a page on "payroll software for restaurants" can get cited for "best payroll software" because the engine searched the restaurant sub-question on its own.

Each engine keeps its own index and preferences, so a page can win in one and vanish in another. Our breakdown of how ChatGPT, Perplexity, Gemini and Claude choose sources goes engine by engine.

How is GEO different from SEO?

GEO runs on the same crawling, indexing and ranking work as SEO and is judged on a different outcome: whether an AI answer names or cites you, usually with no click. SEO is judged on clicks from a ranked list of links. The Pew numbers in the rewrite above are the reason that matters: with an AI summary on the page, only 1% of visits included a click on a cited source14.

So brand mentions in the answer text become a goal in their own right, and measurement moves from rank tracking to sampling answers. We compare the terms, including AEO, in GEO vs AEO vs SEO.

Which GEO tactics actually work?

Five tactics have published evidence behind them. Each one below ends with a test that tells you the work is finished.

Put a number and a named source behind every claim that matters

This is the strongest lab result in GEO: quotation addition, statistics addition and citing sources were three of the top four methods in the original study1. Open your five most important pages and highlight every adjective doing the work of a number ("fast", "most", "significant"). Replace each with a figure and a link to where it came from, or delete the claim. Done means no key claim on the page rests on an unsourced adjective.

Answer the heading in its first sentence

Engines quote passages, so the first one or two sentences under each heading should answer it completely without the rest of the page. Phrase headings as the questions people ask. Google says you don't need to split pages into small chunks, because its systems can surface the relevant part of a longer page3 (Google). A quick test: read only the first sentence under each H2. If it doesn't answer the H2, rewrite it.

Publish something the other ten results don't have

Google's guide tells site owners to focus on unique, non-commodity content that is helpful, reliable and people-first3. That means original data, first-hand tests, a pricing table you actually checked, or a named expert's view. If your page says what the top ten results say, an engine has no reason to pick yours, because any of them would do.

Update the facts, not the date

Ahrefs analyzed about 17 million citations and found AI assistants cited pages that were on average 25.7% fresher than pages in organic results, with ChatGPT the most inclined toward recent pages12 (Ahrefs). Ahrefs also warns that cosmetic updates to weak pages won't help. Re-check prices, version numbers and statistics on your top pages each quarter and change them when they change.

Describe your company the same way everywhere

A model has to know what you are before it can recommend you. Pick one name, one sentence describing what you do and one category, then use them on your homepage, LinkedIn, Crunchbase, G2 and directory listings, and tie them together with Organization markup. If your G2 profile says "HR platform" and your homepage says "payroll API", expect engines to place you in neither list. Our glossary entry on entities explains why.

Keyword stuffing, hidden text and fake mentions backfire

Keyword stuffing reduced visibility in the GEO study1. Google's guide says you don't need to rewrite content specially for AI, and that chasing inauthentic mentions is less useful than it looks because spam systems filter manipulative tactics3. Hidden text aimed at AI crawlers, self-ranking listicles published at scale and fake reviews all carry platform risk with no tested upside.

What technical foundations does GEO need?

Every engine's crawler has to reach your pages and read the content without running JavaScript. Four checks cover most of it.

Search bots and training bots are separate decisions

AI companies run separate bots for training, search indexing and user-triggered fetches, and each obeys its own robots.txt rule. OpenAI's OAI-SearchBot surfaces sites in ChatGPT search, GPTBot collects training data, and ChatGPT-User fetches pages when a user asks5 (OpenAI). Anthropic splits ClaudeBot (training), Claude-SearchBot (search indexing) and Claude-User (user requests)6 (Anthropic). Perplexity runs PerplexityBot for its index and Perplexity-User for live fetches, and says neither is used to train foundation models7 (Perplexity).

You can block training bots and still allow search bots. Blocking OAI-SearchBot or Claude-SearchBot takes you out of those engines' search answers. Check your file with our AI crawler robots.txt checker and see the full list in every AI crawler user agent. Then check your CDN or firewall too, since a bot-protection rule can block a bot that robots.txt allows.

Most AI crawlers don't run JavaScript

Vercel's analysis of crawler traffic found that none of the major AI crawlers it studied rendered JavaScript, including GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot and PerplexityBot8 (Vercel). Gemini, which uses Googlebot's infrastructure, was the exception. Run curl -s https://yoursite.com/page | grep "a sentence from your page". If the sentence isn't in the raw HTML, those crawlers can't see it, and you need server-side rendering or static generation.

Structured data helps understanding, not selection

Google says no special schema.org markup is needed for its AI features, while still recommending structured data as normal SEO practice that must match the visible text2. Microsoft's Fabrice Canel said at SMX Munich in March 2025 that schema markup helps Microsoft's LLMs understand content for Bing and Copilot16 (Search Engine Land). Add Article, Organization, Product and FAQPage markup where it fits, and don't expect it to earn a citation alone. Our schema generator builds valid JSON-LD.

llms.txt is optional

llms.txt is a proposed Markdown file listing a site's key pages for language models. Google's guide states that Google Search doesn't use llms.txt or similar AI-specific files3, and no major AI search engine has publicly confirmed using it to pick sources. It's cheap to publish and can help AI coding tools read developer docs. Our llms.txt explainer covers when it's worth the time.

Why do third-party sites decide so many AI answers?

Because AI engines lean on sources you don't own when they describe or recommend a brand. Researchers at the University of Toronto reported in September 2025 that AI search engines show a "systematic and overwhelming bias towards Earned media" over brand-owned and social content, in contrast to Google's more balanced mix10 (Chen et al.).

Ahrefs studied 75,000 brands and found branded web mentions had the strongest correlation with AI Overview mentions, at 0.664, ahead of branded anchors at 0.527 and branded search volume at 0.39211 (Ahrefs). That's correlation, not proof of cause, but it points the same way as the Toronto study.

Which third parties matter depends on the engine. Profound analyzed 680 million citations collected between August 2024 and June 20259 (Profound):

Reddit led for both Perplexity and AI Overviews, so take part in real threads from real accounts; astroturfed posts get removed and remembered. Wikipedia was ChatGPT's biggest single source, but its conflict-of-interest guideline strongly discourages editing your own company's article, so the way in is independent press coverage that editors can cite. G2 appeared in both ChatGPT's and Perplexity's top 10, which makes a complete profile and a steady flow of genuine reviews worth the effort. Forbes, Business Insider and Reuters were among ChatGPT's most cited sources, and original data is the most reliable way we know to earn that kind of coverage. YouTube and LinkedIn both made the AI Overviews top 10.

How do you measure GEO?

Run a fixed set of buyer prompts through each engine several times and track how often you're mentioned and cited compared with competitors. Single checks mislead: a study by SparkToro and Gumshoe found less than a 1-in-100 chance that ChatGPT or Google's AI would return the same list of brands across repeated runs of a prompt15 (Search Engine Land).

The core metrics are mention rate, citation rate, share of voice, position, sentiment and the domains cited. Add AI referral traffic from analytics and the generative AI impressions report in Google Search Console. For a free first look at whether ChatGPT names your brand, run the Arobis AI visibility checker; the full method, with a worked prompt set and formulas, is in how to measure AI visibility. Once you're tracking more than a few dozen prompts, a paid tracker saves the manual runs; we compare 11 platforms on engines covered and price in the best AI visibility tools for tracking ChatGPT, Gemini and Perplexity.

Six GEO claims the evidence doesn't support

• "GEO replaces SEO." Google says its generative AI features rely on its core ranking and quality systems3. Weak SEO almost always means weak GEO.
• "You need an llms.txt file." Google Search doesn't use it, and no major engine has confirmed using it to choose sources.
• Schema as a citation trigger. Google says no special markup is required for AI features2; markup helps engines read a page and does nothing for a page they never retrieve.
• "We rank #2 in ChatGPT." Answers vary run to run15, so a single position is a snapshot. Report rates over many runs.
• Keyword density. Stuffing scored below the unmodified baseline in the GEO study1.
• "Only your own site counts." Third-party mentions correlated more strongly with AI Overview visibility than anything else Ahrefs measured11.

A 30-day GEO plan with a pass/fail check each week

The order matters: baseline first, access second, pages third, off-site fourth. Each week ends with something you can hand to a colleague and a check that is either true or false.

Weeks 3 and 4 are where in-house teams usually stall, because rewriting pages and earning mentions take writers and outreach time that SEO teams rarely have spare. If you plan to hire that part out, our ranking of GEO agencies lists published prices, who each agency suits and the red flags to check before you sign.

Frequently asked questions

Sources