Generative Engine Optimization (GEO): The Complete Guide

What is generative engine optimization (GEO)?

Generative engine optimization is the practice of making your content and your brand easier for AI systems to retrieve, understand and cite when they generate an answer. The term was formalized in a 2024 research paper, which defined it as optimizing content to be more visible inside generative engine responses.

The word gets used two ways. The academic meaning is about content: which edits to a page make an AI system more likely to quote it. The practitioner meaning is wider: how AI models come to know who you are and what you are credible for, including in conversations where no live search happens. This guide covers both and says which one it means.

How does GEO relate to AEO and SEO? We traced where each term comes from in a separate piece, so we will not repeat it. The short version: the foundations overlap heavily, Google's own position is that optimizing for its AI features is still SEO, and the outcome you chase differs (being named in an answer, not ranked in a list). Our AEO and GEO documentation shows how we split the terms in our own work.

How do generative engines retrieve content and build an answer?

A generative engine answers in stages. It decides whether to search, rewrites your question into several narrower ones, retrieves pages for each, and writes one synthesized answer with citations. Your content has to win at retrieval first, then at synthesis. A page that is never retrieved cannot be quoted, however good it is.

The platforms document this themselves. OpenAI says ChatGPT can search automatically when a question benefits from current information, and that when it uses search partners it typically rewrites the question into one or more targeted queries, then may send more specific follow-ups after reading the first results. Google describes the same pattern for AI Overviews and AI Mode as query fan-out: concurrent related queries generated by the model, with answers grounded in pages from its core search index. Its example turns a question about a weedy lawn into searches on herbicides and chemical-free removal.

Two consequences follow.

Eligibility comes first. For Google's AI features, a page must be indexed and eligible to appear in Search with a snippet. For ChatGPT, Claude and Perplexity, their search crawlers should be allowed to fetch you. More on that below.

Sub-questions are separate chances. A business can be absent from the obvious answer and still appear on a narrower one. In our live audit of 29 B2B companies, comparison-style questions named a business 69% of the time, while problem-first questions, phrased the way someone searches before they know what to buy, named one only 21% of the time (145 live searches, five per company).

There is a second mode too. When a model answers without searching, it relies on what it learned in training, so what the wider web has said about you shapes the answer, and nobody can edit that directly. That is an inference from how these systems work, not a measured finding, but it is why the off-site work later in this guide matters.

What does the research actually show works?

The foundational study is "GEO: Generative Engine Optimization" by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, published at the ACM KDD 2024 conference (arXiv, DOI). It is the first large-scale attempt to test, under controlled conditions, which content changes make a source more visible in generated answers. We read the full paper, and its headline claim of "up to 40%" needs unpacking.

How was the study set up?

The authors built GEO-bench, a benchmark of 10,000 queries from nine sources across 25 domains, about 80% of them informational, each paired with the cleaned text of the top five Google results. A generative engine (GPT-3.5-turbo) answered each query from those five sources. For each query, one source was picked at random, rewritten by a language model using one of nine methods, and scored on how much more visible it became in the answer, averaged over five runs.

Visibility was measured two ways: a position-adjusted word count, which weighs how much of the answer cites the source and how early, and a subjective score judged by a language model. The nine methods included Cite Sources, Quotation Addition, Statistics Addition, Fluency Optimization, Authoritative tone and Keyword Stuffing.

What were the results?

The paper reports that its three best methods improved the position-adjusted metric by roughly 30 to 40 percent over baseline and the subjective metric by 15 to 30 percent. Making text more fluent and easier to understand also helped, by 15 to 30 percent in the authors' words. Two findings cut against common habits:

• Keyword stuffing did little or nothing. It scored below baseline on the main metric (17.7 against 19.3).
• A more persuasive, authoritative tone gave a smaller gain. It scored 21.3, and the authors describe it as no significant improvement overall, though it did better on debate-style queries.

The results also varied by domain, as the authors say plainly. Statistics worked best for Law and Government and opinion-style questions, citations for factual statements, quotations for People and Society and explanatory questions. Combining methods helped: Fluency Optimization plus Statistics Addition beat any single method by more than 5.5% in a 200-query subset.

One more finding deserves its own table. Broken down by where a source ranked in the underlying search results, in a setup where every source was optimized at once, the biggest gains went to the lowest-ranked pages.

The authors read this as a leveling effect: engines condition on page text rather than backlinks, so smaller sites can gain ground. It is encouraging, but it is a hypothesis from one benchmark, not a promise.

What are the limits of the study?

This is the part most summaries skip. The limits are in the paper or follow from its design.

• The engine was a research setup. GPT-3.5-turbo read five Google results. That resembles how commercial engines work, but it is not ChatGPT's live pipeline.
• A language model made the edits. Added statistics and quotations were produced by prompting a model to rewrite the source, so the test measures what the engine favors, not whether the additions were true. Publish only real, sourced numbers.
• The Perplexity check was narrower than it sounds. On 200 queries the authors uploaded source text as files, because Perplexity does not let you specify source URLs. That tests synthesis, not retrieval. Quotation Addition improved the main metric by about 22% there, and keyword stuffing came in about 10% below baseline.
• Visibility was the authors' own definition. Word-count share, position and a model-judged score are reasonable proxies, but they are not clicks, sales or recommendations.
• Search rankings were not tested, and the authors note that engines and queries will keep changing.

So the honest summary is "up to roughly 40% on one metric, in one controlled setup, varying by domain." That is real evidence that sourced, quotable, readable content gets favored. It is not a formula.

Why do off-site sources matter as much as your own pages?

Models read far more than your website. Across the studies we could verify, third-party pages account for most of what AI answers cite, and brand mentions across the web track AI visibility more closely than backlinks do. Treat off-site presence as half of GEO, but note that the studies disagree about which sources matter on which platform.

Our own audits. We ran live AI searches for buyer questions in several industries, never naming the business. The business's own site was a tiny share of what got cited.

Instead, answers leaned on directories, review sites, rankings, Wikipedia and other businesses' pages. See the law firm, insurance agency and marketing agency audits for detail, or all of them on our AI visibility research page.

Platform citation mix. Profound's analysis of citations across ChatGPT, Google AI Overviews and Perplexity shows sources differ sharply by platform. One detail is easy to misread: its headline percentages are shares within each platform's top ten sources, not shares of all citations.

The real picture is a long tail: no single source topped 8% of all citations on any platform, which fits our own finding that citations were spread across hundreds of domains.

Correlation and preprint evidence. Ahrefs analyzed 75,000 brands and found YouTube mentions correlated most strongly with AI visibility (about 0.74), branded web mentions next (0.66 to 0.71 depending on platform), and page count only about 0.19. Ahrefs says correlation is not causation, and its sample was limited to domains with a domain rating above 40, so it describes established brands. A preprint by Chen and colleagues (not peer reviewed as far as arXiv shows) ran 1,000 consumer ranking prompts through Google and a web-enabled GPT model. In US software queries, 72.7% of AI sources were earned media (independent reviews and publications) against 45.4% for Google, and brand-owned sites fell from 43.7% to 26.7%.

These studies do not fully agree. Chen's team classed Reddit, Quora and YouTube as social and found almost none in AI results, while Profound found Reddit leading on two platforms. Engines, query sets and classification differ. Do not copy a headline from one study. Find out which sources your category's answers actually cite.

One caution from Google: it says seeking inauthentic mentions is less helpful than it looks, because its systems favor high-quality content and block spam. Earn mentions the slow way: genuine reviews, trade coverage, partner pages and community answers that stand on their own.

How do you build entity and brand consistency?

Entity consistency means every place that describes you uses the same name, category, location and core facts, so a model can resolve all of it to one organization. No controlled study isolates its effect, so treat it as low-cost groundwork backed by how the systems work, not a proven lever.

The logic is simple. A model assembling an answer pulls from many pages. If one calls you an agency, another a studio, and a directory still lists a former name, the picture is muddier and the model has room to guess. That is our reasoning and experience from audits, not a measured result.

• Write a canonical description. One sentence on what you do and for whom, plus the facts that must never vary: legal name, location, services, leadership names.
• State it plainly on your own site first. Home, About and contact pages should say the same thing in the same words.
• Add Organization structured data. Google says it can help it understand and disambiguate your organization, and sameAs links your profiles together. Do not expect a citation lift from it alone (see the schema evidence below), but it is cheap.
• Match the profiles and directories that matter in your category, and fix stale listings.
• Keep author names, roles and profile links consistent wherever your experts publish.

How do you build content for retrieval?

Build pages so a retrieval system can find the right passage and a model can lift it safely: a direct answer first, claims backed by named sources, original information worth citing, and key text in the HTML. Most of this is good SEO. Where evidence exists we say so.

Lead with the answer, then support it

Open each section with a plain answer in a sentence or two, then give the detail. No controlled study isolates answer-first placement, and Google says there is no requirement to break content into tiny pieces. We do it anyway because retrieval picks passages, readers scan first, and the habit forces clarity, which is what the GEO paper rewarded. Our guide to getting cited covers the evidence in more depth.

Add things worth citing

The paper's best performers were statistics, quotations and citations. The takeaway is not to sprinkle numbers for show. Give a model something concrete to attribute: a figure with a named source, a quotation from a named person, a link to the primary document. Google's guidance agrees from the other side, advising non-commodity content with a unique point of view. Original data is the strongest version, because nobody else can cite a number you measured.

Cover the sub-questions on one strong page

Because a buyer question fans out into related searches, answer them inside one well-organized page. Do not generate a thin page per variation. Google says doing that primarily to manipulate rankings or AI responses violates its scaled content abuse policy.

Let the right crawlers in

Retrieval requires access. Each major vendor runs separate bots for search, user-triggered fetches and model training, and you can treat them differently.

Two failure modes are common. A robots.txt written years ago, or a blanket block copied from a template, can shut out search bots along with training bots. And your CDN or firewall can block them even when robots.txt allows them: Perplexity's documentation walks through firewall allow-listing, OpenAI asks you to allow its published IP ranges, and Cloudflare has said it changed its default to block AI crawlers unless they pay creators, so a site behind such a service can be unreadable to AI search without anyone deciding that. Our audits ran into the same wall from the outside: many of the sites we tried to scan blocked our own crawler, and an AI crawler could meet it too. Whether each AI crawler runs JavaScript is not documented consistently, so we also check that key text is in the initial HTML.

What about schema markup?

Schema is worth having for classic search, but the best causal evidence says it is not a citation lever. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 matched control pages. Citations on AI Overviews fell 4.6% relative to controls (statistically significant but small), while AI Mode (+2.4%) and ChatGPT (+2.2%) were indistinguishable from zero. Meanwhile, AI-cited pages were almost three times as likely to have JSON-LD, which is exactly the kind of correlation that misleads. Google also says structured data is not required for its generative AI search.

Does llms.txt help you get cited?

The evidence that llms.txt improves AI visibility is thin to nonexistent. It is a real proposal and some tools use it, but Google says Search ignores it, and an SE Ranking analysis of nearly 300,000 domains found no relationship with citation frequency. Publish one if it is cheap for you. Do not build a strategy on it.

What is established: Jeremy Howard of Answer.AI proposed llms.txt in September 2024, a Markdown file at a site's root that gives AI agents a curated overview and links to clean versions of key pages. It is a community convention, not a formal standard, and the proposal says it is used most heavily for software documentation, where coding agents follow it to find API references.

What is not established: any effect on AI search visibility. Google's guide says you do not need llms.txt or similar files to appear in Google Search, including its generative features, and that creating one will neither help nor harm you there. SE Ranking found the file on 10.13% of the nearly 300,000 domains it analyzed. In its statistical tests and an XGBoost model, having one showed no effect on citation frequency, and removing the variable improved the model's accuracy. That is a domain-level correlation on one sample, so it cannot rule out an effect on a single assistant, but it gives no reason to expect one. OpenAI's crawler documentation covers robots.txt controls and does not describe llms.txt as an input to search results, though OpenAI does publish llms.txt files for its own developer docs, the documentation use case the proposal describes.

Our position matches the evidence: one low-cost lever among several, with no assistant guaranteed to use it.

How do you measure GEO?

Measure GEO by asking AI assistants realistic buyer questions that never contain your name, running each many times, and tracking how often you are named. Add referral traffic and branded search. Do not track a position, because the lists change on almost every run.

SparkToro and Gumshoe.ai (an AI tracking vendor, which they disclose) found that nearly every response differed in which brands appeared, their order and their number. Their conclusion: rank in AI answers is close to meaningless, but visibility, the share of runs in which a brand appears, looked reasonable when averaged over many runs. They recommend asking the same question at least 60 to 100 times.

A workable method:

1. Write constructed buyer questions of at least three types: category ("best X for Y"), problem-first (the symptom, not the solution) and comparison. Our audits use five per business, including one local question. If a question includes your name, it proves nothing.
2. Run each one repeatedly across the assistants your buyers use. Record whether you are named, who is named instead and which sources are cited.
3. Track the rate, not the rank. "Named in 40% of runs on problem-first questions" is usable. "Ranked third" is not.
4. Add downstream signals: Search Console's Generative AI performance report for Google's features, AI assistants set up as their own referral channel in analytics, and branded search volume as the honest indirect signal.
5. Repeat on a schedule with the questions fixed, so changes mean something.

You can try the method on your own site with our free AI visibility checker. It is a single snapshot, so read it as a starting point and compare it with the industry results on our AI visibility research page.

What are the most common GEO mistakes?

Most mistakes come from treating GEO as a trick instead of a content and reputation problem.

• Stuffing keywords or turning up the sales tone. The research found no gain from the first and little from the second.
• Betting on llms.txt or schema alone. Neither has evidence behind it as a citation lever.
• Blocking the wrong bots. A blanket robots.txt or firewall rule can remove you from AI search while doing nothing useful about training.
• Mass-producing variant pages. Google treats this as scaled content abuse, and it does not improve retrieval.
• Ignoring everything off your site. If directories, reviews and third-party pages describe you inconsistently or not at all, polishing your own pages will not close the gap.
• Inventing statistics or chasing a rank. The paper's test could not tell true additions from false ones, and a fabricated figure is a liability for you. Paying to track a position in AI answers measures noise.

What does a practical GEO plan look like?

Work in this order, because each step makes the next one measurable.

1. Baseline. Build your constructed question set, run it repeatedly and record your visibility rate, the competitors named and the sources cited.
2. Fix access. Review robots.txt, your CDN and your firewall rules for the search bots above, and decide separately about training bots.
3. Define the entity. Write the canonical description and facts, align your site, structured data and key profiles, and correct stale listings.
4. Upgrade the pages that should answer the questions. Lead with the answer, add sourced statistics and named quotations, link primary documents, and cover sub-questions on one strong page.
5. Earn the off-site presence your category's answers actually cite. Your baseline's cited-sources list tells you which directories, publications, communities and video channels to target.
6. Re-measure and adjust. Keep the questions fixed, compare rates and drop anything that does not move them.

If you would rather have it done for you, GEO is part of the work described on our AEO and GEO service page.

FAQ

Related reading

• AEO vs GEO vs SEO: what actually separates them
• What is answer engine optimization (AEO)? The complete guide
• How to get cited in ChatGPT, Perplexity and AI Overviews
• Outbound lead generation: the complete guide
• We ran a real AI visibility audit on 29 B2B companies
• We audited 30 B2B SaaS companies for AI visibility
• AI visibility research: every audit in one place
• AEO and GEO documentation

Sources

• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, GEO: Generative Engine Optimization, KDD 2024 (DOI: 10.1145/3637528.3671900)
• Google Search Central, Optimizing your website for generative AI features on Google Search
• Google Search Central, Organization structured data
• Google Search Central, Google's common crawlers
• OpenAI Help Center, Searching the web with ChatGPT
• OpenAI, Overview of OpenAI crawlers
• Anthropic, Does Anthropic crawl data from the web, and how can site owners block the crawler?
• Perplexity, Perplexity crawlers and user agents
• Cloudflare, Content Independence Day: no AI crawl without compensation
• Profound, AI Platform Citation Patterns: How ChatGPT, Google AI Overviews, and Perplexity Source Information
• Ahrefs, Top Brand Visibility Factors in ChatGPT, AI Mode, and AI Overviews (75k Brands Studied)
• Ahrefs, We Tracked 1,885 Pages Adding Schema. AI Citations Barely Moved.
• Chen, Wang, Chen & Koudas, Generative Engine Optimization: How to Dominate AI Search (arXiv preprint)
• SE Ranking, LLMs.txt: Why Brands Rely On It and Why It Doesn't Work
• Jeremy Howard, The /llms.txt file
• SparkToro, New research: AIs are highly inconsistent when recommending brands or products
• Deligatr, AI visibility research and the audit posts linked above (law firms, insurance agencies, marketing agencies, 29 B2B companies)