Generative Engine Optimization: The Complete Guide

Generative Engine Optimization (GEO) is the practice of making your content easy for AI answer systems, such as ChatGPT, Perplexity, Gemini and Google’s AI Overviews, to find, understand and cite when they write a response. SEO competes for a place in a list of links. GEO competes for a mention inside the answer itself.

This guide covers what the research behind GEO showed, how these engines choose sources as far as anyone can document, what to do about it (ordered by how well each action is supported), and how to tell whether it worked. Want the short version first? Read What is GEO? and come back.

What is Generative Engine Optimization, and why does it matter now?

A generative engine retrieves information and then writes a response in natural language, usually with links to the sources it used. ChatGPT search, Perplexity, Gemini and Google’s AI Overviews and AI Mode are examples. People read one composed answer instead of scanning ten links. So the question for a site owner changes from “do I rank?” to “is my page one of the sources behind that answer?”

You’ll also see people search for “GEO SEO”. The two belong together because the work overlaps heavily with classic search optimization, as the comparison table below shows. GEO is SEO extended to a new kind of result, not a separate discipline that replaces it.

Why now? Use is large, and it’s spread across several products:

People use more than one AI product, and the mix is moving. Optimizing for a single engine is a risky bet. For more figures, with their methods, see our AI search statistics page.

One more fact sets expectations. Cloudflare sits in front of a large share of websites, and it measured how many times AI companies’ crawlers fetch pages for each visit they send back:

The figures come from Cloudflare’s network only, and they mix crawling for training with crawling for search. Still, AI answers return far fewer visits per page read than a classic search engine does. GEO is less about collecting clicks and more about being present, and accurately described, in the answer.

Where does GEO come from?

The term comes from research, not marketing. In November 2023 Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande posted “GEO: Generative Engine Optimization” on arXiv, the open archive where researchers share papers before and after review. It was published at KDD 2024, a major data-science conference.

What did the original GEO paper actually test?

The paper did three things:

• It named the problem: content creators have no way to know how to improve their visibility inside answers written by generative engines.
• It built GEO-bench, a benchmark (a standard set of test queries) of 10,000 queries spanning about 25 domains.
• It ran black-box optimization experiments. The authors changed the content of web pages and measured how often the changed pages appeared in the engines’ responses, without any view of how those engines work inside.

The headline result: GEO methods can boost visibility in generative engine responses by up to 40%, and effectiveness varies by domain. Both halves matter. “Up to 40%” is a ceiling reached by the best methods in the best conditions, not an average. “Varies by domain” means a tactic that helps in one subject area may do little in another.

Of nine methods tested, adding quotations, adding statistics and citing sources performed best. Keyword stuffing gave little or no improvement.

What did later studies and critiques add?

Several later papers complicate the picture. They’re also the source of most of what can be said with evidence today.

Finding	Study	Limit
Citations are often unreliable. Only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to.	2023 Stanford-led study (Liu, Zhang and Liang): people evaluated four generative search engines	These engines have changed since. Even so, a link next to a sentence doesn’t prove the sentence is right.
Answer engines struggle with verifiability. Participants identified 16 limitations. An automated evaluation of three engines found frequent hallucination (made-up content) and inaccurate citation.	2024 study led by Narayanan Venkit of Penn State, with co-authors from Salesforce AI Research: usability study with 21 participants	Small sample, 2024 versions of the engines.
News citation is error-prone. The chatbots collectively gave incorrect answers to more than 60% of queries. Perplexity was wrong 37% of the time and Grok-3 94%.	Tow Center for Digital Journalism at Columbia, February 2025: 1,600 queries across eight AI search tools (200 article excerpts from 20 publishers, asked of each tool once)	The authors caution against generalizing, and the task was identifying a source, not writing a good answer.
Earned media dominates. The first preprint reports a “systematic and overwhelming bias towards Earned media”, meaning independent third-party coverage, over brand-owned and social content, unlike Google’s more balanced mix. It also finds that engines differ in domain diversity, freshness, stability across languages and sensitivity to how a question is phrased.	Two preprints from the same Toronto research group (Chen, Wang, Chen and Koudas) compare AI search with Google. The second, updated in May 2026, compares source domains, content types, query intent and freshness between Google Search and leading generative AI services.	Both abstracts give no numbers, and both are preprints that have not been through full peer review.
Page-level features correlate with citation. Metadata and freshness, semantic HTML and structured-data features were most associated with citation.	2025 preprint (Kumar and Palkhouski): 1,702 citations across 1,100 unique URLs, using 70 product-intent questions put to Brave Summary, Google AI Overviews and Perplexity	Observational, English-language B2B software pages only, shows association, not cause.

Taken together: the original 40% figure is real but narrow. Engines are inconsistent. Third-party presence appears to matter a great deal. And nobody has published a controlled, repeatable test showing that a specific on-page change on your site will win you citations in every engine.

How do generative engines pick and cite sources?

No AI company publishes its full source-selection rules. What we can say comes from official documentation and a few studies, read with their limits in mind. There are four layers.

1. Can the engine reach your page?

A crawler is an automated program that reads web pages. A robots.txt file is a plain text file at the root of your site that tells crawlers what they may read. Each AI company runs more than one crawler, with different jobs:

• OpenAI documents GPTBot (content that may be used to train models), OAI-SearchBot (surfacing sites in ChatGPT search) and ChatGPT-User (fetches triggered by a person’s request). OpenAI says sites that opt out of OAI-SearchBot will not be shown in ChatGPT search answers, with about 24 hours for a change to apply. For ChatGPT-User, OpenAI says robots.txt rules may not apply.
• Anthropic documents three bots, one each for training, user-directed fetching and search, and says they respect “do not crawl” signals in robots.txt.
• Perplexity says PerplexityBot is for surfacing sites in search and is not used to train foundation models. It says Perplexity-User, which acts on a person’s request, generally ignores robots.txt rules.
• Google uses Google-Extended as a control for whether content is used to train future Gemini models and for grounding. Google says it does not affect a site’s inclusion in Google Search and is not a ranking signal.

Whether bots obey is a separate matter.

Treat robots.txt as a statement of your wishes, not a guarantee.

2. Is the page eligible?

For Google, the documentation is direct. A page must be indexed and eligible to be shown in Search with a snippet to appear as a supporting link in AI Overviews or AI Mode. Google says there are no additional requirements, and that no special schema.org markup (structured data that labels page content for machines) is needed.

The controls are the same snippet controls used in Search: nosnippet, data-nosnippet, max-snippet and noindex. Traffic from AI features is counted in Search Console under the “Web” search type, so you can’t separate it from the rest.

3. What makes a page more likely to be chosen?

Here the evidence is thinnest. The strongest signals are the ones above: the original GEO experiments, the earned-media bias and the GEO-16 association between page features and citation. Google’s own guidance cuts the other way on markup, as the playbook below explains.

One more result is indirect. “Lost in the Middle” (Liu and colleagues, 2023) found that language models often use information best when it sits at the beginning or end of the text they are given, and worse when it’s in the middle. That’s an experiment on text fed to models, not on search results. Read it as an analogy for putting the key answer first, not proof about citation order.

4. Is the citation correct?

As covered above, being cited is not the same as being represented accurately. That’s why any GEO program needs a step where a person reads what the engines say about the business.

GEO vs SEO vs AEO: what is the difference?

SEO (search engine optimization) aims at ranking in a list of search results. AEO (answer engine optimization) is an informal label for winning the direct answer a system gives, such as a featured snippet or a voice answer. GEO aims at being used and cited inside a written answer composed by a generative engine.

Our SEO vs AEO vs GEO comparison goes deeper. This table is the summary.

	SEO	AEO	GEO
Full name	Search engine optimization	Answer engine optimization	Generative engine optimization
Goal	Rank in a results list	Be the direct answer	Be cited inside an AI-written answer
Typical surface	Google and Bing result pages	Featured snippets, voice answers	ChatGPT, Perplexity, Gemini, AI Overviews, AI Mode
What you earn	A click from a link	A displayed answer	A mention or citation in the response
Main inputs	Crawlability, relevance, links, page experience	Short, clear, direct answers	Access for AI crawlers, clear answers, third-party coverage, accuracy
How it is measured	Rankings, impressions, clicks	Snippet ownership	Citations and mentions across a fixed set of questions
Status of the term	Established	Informal	Coined in a 2023 paper; still evolving

The overlap matters most. Indexing, clear writing and trustworthy content serve all three. For a wider view of the AI side of SEO, see our AI SEO guide.

The GEO playbook, ordered by strength of evidence

Most GEO advice online doesn’t say how strong its support is. Here every tactic carries one of three labels:

Label	What it means
Measured in a study	A published study tested or observed it. The study may have limits, which we state.
Official guidance	An engine’s own documentation recommends or describes it.
Reasonable practice, not yet measured	Sensible and low-risk, but we found no study that measures its effect on AI citations.

Summary table

#	Tactic	Evidence label	Main source and limit
1	Earn independent third-party coverage	Measured in a study	Toronto preprints; abstracts give no numbers
2	Add specifics, quotations and citations to content	Measured in a study	Original GEO paper; 2023 engines, up to 40%
3	Keep metadata fresh and use clean semantic HTML	Measured in a study	GEO-16 preprint; observational, B2B software pages
4	Let the right crawlers in	Official guidance	OpenAI, Anthropic, Perplexity, Google docs
5	Keep key pages indexable and snippet-eligible	Official guidance	Google Search Central
6	Make interactive elements properly labeled	Official guidance	Chrome Lighthouse (experimental)
7	Write natural content for people	Official guidance	Google Search Central
8	Lead each page with a direct answer	Reasonable practice, not yet measured	Indirect analogy only
9	Cover the related sub-questions	Reasonable practice, not yet measured	No citation study found
10	Audit what engines say about you	Reasonable practice, not yet measured	Motivated by accuracy studies
11	Publish an llms.txt file	Measured in a study (no effect found)	SE Ranking, Ahrefs; Google neutral

Measured in a study

1. Earn independent third-party coverage. In the Toronto group’s controlled experiments, AI search showed a systematic and overwhelming bias towards earned media over brand-owned and social content. If that holds for your market, reviews, press, trade publications, directories and expert mentions may carry more weight in AI answers than your own pages. So find out which independent sites the engines cite for your topic, and work to be accurately described there.

2. Add specifics, quotations and citations to your content. The original GEO paper tested content-rewriting methods and reported gains of up to 40% in visibility, varying by domain. For a small business, that means making each page more concrete and checkable.

Use real figures with their source. Quote named people who agreed to be quoted. Link to primary sources.

Never invent any of it. An invented statistic is worse than none, and the whole approach depends on content that deserves to be cited.

3. Keep metadata fresh and the HTML clean. The GEO-16 preprint found that metadata and freshness, semantic HTML (proper heading, list and table tags rather than styled boxes) and structured data were the page features most associated with citation, across 1,100 unique URLs. An example threshold of high scores aligned with substantially higher citation rates in its data.

11. llms.txt: tested, with no effect found. An llms.txt file is a proposed plain-text file placed at the root of a site that lists and summarizes key pages for AI systems. Our llms.txt guide explains it in full, and you can check yours with the llms.txt checker. Here we only report what has been measured.

• SE Ranking’s study of nearly 300,000 domains found no correlation between having the file and being cited by AI engines, and a model without that variable predicted citations better.
• Ahrefs, studying 137,210 domains, found that 97% of the llms.txt files it saw received zero requests in May 2026, and that no AI bot requested non-existent paths.
• Google says llms.txt “will neither harm nor help” visibility in Search.

Adoption figures differ sharply because each study measured a different population:

Common Crawl’s authors say their figure is not comparable with the Web Almanac’s.

Official guidance

4. Let the right crawlers in. Check your robots.txt, and decide bot by bot. Allowing a search bot such as OAI-SearchBot is what documentation ties to appearing in ChatGPT search answers. Blocking a training bot such as GPTBot is a separate choice.

Your hosting or security setup can also block bots without your knowing, so test from outside. And remember the compliance limits above.

5. Keep key pages indexable and snippet-eligible. Google says a page must be indexed and eligible to be shown with a snippet to appear in AI Overviews or AI Mode. Make sure important pages return normal responses, aren’t marked noindex, and aren’t restricted by snippet controls you didn’t mean to set.

6. Label interactive elements properly. Google’s Lighthouse tool, which audits web pages, added an experimental “Agentic Browsing” category. Its documentation says agents rely on the accessibility tree as their primary data model. The accessibility tree is the structured version of a page that screen readers use: names, roles and states of buttons, links and form fields. The audits check names and labels, tree integrity and visibility of interactive content.

7. Write natural content for people. Google’s guidance for generative AI features says there’s no requirement to break content into tiny pieces and that structured data is not required for generative AI search. It also says you don’t need new machine-readable files, AI text files, markup or Markdown to appear in Google Search. The Search Engine Journal noted that Lighthouse’s experimental audit checks for llms.txt while Search says it isn’t needed, and tied the difference to Search visibility versus browser-agent readiness.

Reasonable practice, not yet measured

These steps are cheap, low-risk and consistent with the sources above, but we found no study that measures their effect on AI citations.

8. Lead each page with a direct answer. Put a plain two-to-three sentence answer under each heading. Write sentences that make sense on their own, with the subject named, so they can be quoted without the paragraph around them. The only related evidence is the “Lost in the Middle” experiment, which concerns text given to a model, not ranking in search.

9. Cover the related sub-questions. If you sell a service, answer the questions a buyer asks around it: cost, process, risks, alternatives, local details. Real questions from your customers and sales team are the best source. No study we reviewed tests this against citation rates.

10. Audit what engines say about you. Ask the engines your customers use about your business and your category, and read the answers critically. Given the accuracy findings above, expect errors. The fix is usually on the sources the engines draw from, including your own pages.

How do you measure GEO?

There’s no standard GEO metric. The original paper used its own visibility measures, and commercial tools use others. Any figure a vendor shows you depends on its method, so ask for the method. Two complementary measurements are practical.

1. AI visibility tracking

Pick a fixed set of questions, perhaps 20 to 50 that your customers genuinely ask, and run them on a schedule:

Search Console won’t isolate this for you. Google counts traffic from AI features within the “Web” search type, combined with ordinary results. In other engines, check your analytics for referrals, but remember the crawl-to-referral ratios above: no clicks doesn’t mean no citations.

Be careful with click data in general. Studies disagree on how much AI summaries cost websites in clicks, and the reasons are method. These don’t contradict each other so much as measure different things:

Source	What it found	Method note
Pew	A link click on 8% of Google visits with an AI summary against 15% without	Browsing panel
Ahrefs	The top result’s click-through rate with an AI Overview about 58% lower than forecast	2026 update that reused its earlier method over a longer window (December 2023 to December 2025) and updated its 2025 figure of 34.5%
Google	Total organic click volume “relatively stable” year over year	A statement without published numbers; measures total volume across all sites
Seer Interactive	Brands cited in AI Overviews got about 35% more organic clicks than uncited ones	Its authors do not claim cause

2. Site structure and readiness

The other half is whether engines and agents can navigate and read your site at all. That means crawl access, indexability, a clear internal-link structure so important pages are reachable, descriptive headings, and accessible markup. Our guide to agentic navigability explains this layer in detail.

GraphoRank is a tool for this second measurement: it crawls a site, respecting robots.txt, and scores how well AI agents can navigate, read and cite it on a 0-100 scale. How it works shows the method and its limits.

Common mistakes

#	Mistake	Why it hurts
1	Treating GEO as a replacement for SEO	Google’s own documentation ties AI features to indexing and snippet eligibility. If search engines can’t index a page, GEO effort on it is wasted.
2	Blocking every AI bot by default	Training, search and user-triggered crawlers are different. Blocking search bots can remove you from the answers you want. Blocking training bots is a separate, legitimate choice.
3	Trusting robots.txt to stop everything	Documentation says some user-triggered fetchers may not follow it, and a study found AI search crawlers rarely check it.
4	Chasing one engine	Usage is split across products and the mix shifted within a year.
5	Inventing proof	Made-up statistics, fake reviews and planted mentions are the opposite of what the evidence rewards, and Google says inauthentic mentions are not as helpful as they seem.
6	Believing guaranteed-uplift claims	The strongest published number is “up to 40%” from 2023 experiments. If a vendor promises a fixed gain, ask for the method and the sample.
7	Adding markup because a blog said so	Google says structured data is not required for its AI features, and a large study found no citation link for llms.txt. These files are cheap and harmless, but they shouldn’t displace work on access, content and independent coverage.
8	Never reading what the engines say	Accuracy studies show errors are common. Check, and correct at the source.
9	Measuring once	Results vary by engine, by run and by phrasing. One screenshot is not a measurement.
10	Ignoring accessibility	Markup that screen readers can’t understand may also be hard for agents to use, according to Google’s Lighthouse documentation.