What is generative engine optimization (GEO)? A complete guide
GEO & AI Optimization

What is generative engine optimization (GEO)? A complete guide

2026-03-30 14 min read

What GEO means, where the term comes from, what the research and the vendors actually say, and how to build and measure a GEO strategy step by step.

Generative engine optimization (GEO) is the work of making a website readable, understandable and worth citing for AI services that answer questions using the web: ChatGPT, Claude, Perplexity, Copilot, and Google's AI Overviews and AI Mode. The goal is for your content to be retrieved, used and linked as a source in the answer, not only to rank in a list of results. Most of the work is the same as good search engine optimization, with a few additions that concern the AI services' own crawlers and the way they lift individual passages out of a page.

In brief

  • The term comes from a research paper posted in November 2023 and presented at the KDD 2024 conference. It measured visibility inside generated answers in a test setup, not traffic or revenue.
  • AI services that search the web first retrieve pages through a search index and then write an answer from them. You have to be findable and readable before anything else matters.
  • What you can influence is access, structure, clarity, evidence, identity and freshness. No single technique guarantees a citation.
  • Google states explicitly that no special markup or special files are required to appear in its AI features, and that from its perspective GEO is still SEO.
  • Since 2026 there is first-party data: a generative AI report in Google Search Console and AI Performance in Bing Webmaster Tools.

What GEO means and where the term comes from

GEO stands for generative engine optimization. A "generative engine" is a search service that does not just list links but has a language model write a single answer from several sources, usually with citations.

The term was introduced in the paper GEO: Generative Engine Optimization by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande. The first version was posted to arXiv on 16 November 2023 and the paper was accepted to KDD 2024. The authors were affiliated with IIT Delhi and Princeton University, with two listed as independent.

What the study actually measured

The researchers built their own test engine: for each query, the top five results were fetched from Google, and GPT-3.5 Turbo wrote an answer with citations from them. They then had a language model rewrite one of the five sources using nine different methods and measured how that source's visibility in the answer changed. The test set, GEO-bench, contains 10,000 queries drawn from nine sources.

Visibility was measured in two ways: the number of words in the answer attributed to the source, weighted by how early they appear (position-adjusted word count), and a rating produced by a language model (subjective impression). The main results:

  • The best methods were adding quotations from relevant sources, adding statistics and citing sources. They improved visibility by up to 41 percent on the first metric and 28 percent on the second, compared with the unmodified text.
  • Keyword stuffing, repeating the query terms, gave little to no improvement.
  • In a smaller test against Perplexity, quotations improved the first metric by 22 percent and statistics improved the second by up to 37 percent. Keyword stuffing performed 10 percent worse than the unmodified text.
  • Lower-ranked sources gained the most. With source citations added, visibility for the page ranked fifth rose by 115.1 percent, while the page ranked first lost 30.3 percent on average.
  • Which method worked best varied by subject area.

The study's limits

The "up to 40 percent" figure is often quoted as if it referred to traffic, or to citations in ChatGPT. It does not. It refers to share of a generated answer in a test setup with GPT-3.5 and five sources, measured in 2023. The authors themselves note that they tested only two engines, that the methods may need to adapt as the engines evolve, and that they did not evaluate how the methods affect search rankings. The rewrites were also produced automatically by a language model. The reasonable takeaway is that concrete evidence makes a text more useful as a source, not that numbers and quotes should be sprinkled in for their own sake.

We cover later work in our overview of GEO research.

GEO and SEO: same foundation, partly new goal

GEO does not replace SEO. AI services that search the web depend on a search index, and a page that cannot be crawled and indexed cannot become a source in an answer either. In its guide to optimizing for generative AI features, published in May 2026, Google writes that from Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, "and thus still SEO".

SEOGEO
GoalRank in the results and earn the clickBe retrieved, used and linked as a source in an AI answer
Unit selectedThe pageOften a single passage or claim on the page
CrawlersGooglebot, BingbotThe same, plus the AI services' own, such as OAI-SearchBot, Claude-SearchBot and PerplexityBot
MeasurementPositions, impressions, clicksImpressions and citations in AI answers, AI-referred traffic

For a fuller comparison, see SEO vs GEO: differences and similarities. If you are wondering about neighbouring acronyms such as AEO, LLMO and SGE, we sort them out in the terminology article.

How AI answer engines retrieve and cite

An AI answer can draw on two things: what the model learned during training and what it retrieves from the web at the moment of asking. When the service searches, it happens in three steps.

  1. The question is rewritten into searches. Google calls the technique query fan-out: its AI features issue several related searches across subtopics to build the response. Pages can therefore become sources for queries the user never typed word for word.
  2. Pages are retrieved from an index. The technique is called retrieval-augmented generation, or grounding. Google retrieves from its regular Search index. OpenAI, Anthropic and Perplexity crawl the web with their own bots for their search features.
  3. The model writes the answer and links to its sources. How the selection works in detail is not public, but it is reasonable to assume that text where the answer is stated plainly, and makes sense without the rest of the page, is easier to use.

The whole chain is described in How AI search engines work and in the article on how language models choose their sources.

What you can actually influence

1. Access

The AI companies use different crawlers for different purposes, and you control each one separately in robots.txt. The details below are as of November 2026 and come from each company's own documentation.

CompanySearch and answersTrainingUser-requested fetching
OpenAIOAI-SearchBotGPTBotChatGPT-User
AnthropicClaude-SearchBotClaudeBotClaude-User
PerplexityPerplexityBotNot used for training, according to PerplexityPerplexity-User
GoogleGooglebotGoogle-Extended (controls training and grounding in Gemini, not Search)

Some details are easy to miss. OpenAI states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, although they can still appear as navigational links, and that a robots.txt change can take about 24 hours to take effect. Anthropic describes three separate bots along the same lines. Perplexity says that Perplexity-User generally ignores robots.txt rules, because the fetch is made at a user's request. Google-Extended, according to Google, affects neither a site's inclusion in Google Search nor its ranking.

So you can say no to training and yes to search. This is what it looks like for a site that wants to appear in answers without contributing training data:

User-agent: OAI-SearchBot
User-agent: Claude-SearchBot
User-agent: PerplexityBot
Allow: /

User-agent: GPTBot
User-agent: ClaudeBot
User-agent: Google-Extended
Disallow: /

Sitemap: https://www.example.com/sitemap.xml

Also check that your firewall and CDN are not blocking the bots, and that the important content is present in the HTML without JavaScript having to run. More in the guide to robots.txt for AI crawlers and the technical factors for AI visibility.

2. Structure

A clear heading hierarchy, short paragraphs that cover one thing at a time, tables for comparisons and lists for steps make it easier to find the right part of a page, for people and machines alike. In its AI Performance announcement, Bing recommends clear headings, tables and FAQ sections. Google, for its part, says there is no requirement to break content into tiny pieces for AI. The two fit together: structure for the reader, and do not chop the text up for a machine. See content architecture for AI visibility.

3. Clarity

Write so that the answer can be lifted out. Open the section with the answer, name the thing you are talking about instead of writing "it" and "this", and state units, dates and conditions.

Before:

Our solution is market-leading and delivers fast results for most customers.

After:

Example Store ships orders within the UK in two to four working days. Delivery costs £3.95 and is free on orders over £40. You can return an item within 30 days of receiving it.

The second version (an invented example) answers three common questions, can be checked, and works without context. There are more examples in Writing content AI wants to cite.

4. Evidence

This is the part with the strongest research support, subject to the caveats above. Give a source for numbers, link to the primary source, state the year a figure refers to and quote named people accurately. Bing gives the same advice: back claims with examples, data and cited sources. Google singles out original perspectives and first-hand experience. It calls this non-commodity content, material that goes beyond common knowledge, and expects it to matter more over time than any other suggestion in its guide.

5. Identity and authority

It should be clear who is behind the content: organisation, author, contact details. Use the same names and details everywhere, and describe the organisation with structured data that matches the visible text. Google's guidance on helpful content says that E-E-A-T is not itself a specific ranking factor, but that its systems use a mix of factors to identify content with good E-E-A-T, and that trust is the most important aspect. Mentions on other websites can show up in AI answers, but Google advises against chasing inauthentic ones. Read on about E-E-A-T in the AI era, digital authority and JSON-LD and structured data.

6. Freshness

Keep pages up to date, show when they were last changed and remove outdated details. Bing recommends regular updates so that AI systems reference the current version, and the IndexNow protocol for notifying search engines of changes. Do not change the date without changing the content.

What the vendors themselves say

Google is the most explicit. Its documentation AI features and your website says there are no additional requirements to appear in AI Overviews or AI Mode: the page has to be indexed and eligible to be shown in Google Search with a snippet. In the May 2026 guide, Google lists things you do not need to do on its account: llms.txt and other special AI files (Google Search does not use them), splitting text into small chunks, rewriting text just for AI systems, seeking inauthentic mentions, and overfocusing on structured data. Structured data is not required for the AI features, but it remains useful for rich results and should always match what is visible on the page.

Microsoft uses the word GEO itself. When AI Performance in Bing Webmaster Tools launched in public preview in February 2026, it was described as an early step toward GEO tooling, and the post contains the advice on structure, evidence and freshness cited above.

OpenAI, Anthropic and Perplexity mainly document their crawlers. What they say about visibility concerns access: if you shut out the search bot, you cannot expect to appear in search answers.

Note that Google's statements apply to Google Search. That Google does not use llms.txt says nothing about other services, and equally, advice from Bing is no promise about how ChatGPT or Claude behave.

Getting started: a GEO strategy in seven steps

  1. Choose the questions. List 20 to 30 questions your customers actually ask and where you want to be the source. Take them from sales calls, support tickets and search data, phrased as full questions.
  2. Check access. Read your robots.txt against the table above, confirm in your server logs that the bots receive a 200 response, and make sure the pages are indexed in both Google and Bing.
  3. Record a baseline. Ask the questions in the services your customers use and note which sources are cited. Open the AI reports in Search Console and Bing Webmaster Tools.
  4. Fix the technical basics. Content in the HTML without JavaScript dependency, reasonable load time, canonical, sitemap, and no accidental noindex or snippet restrictions. Our GEO checklist walks through the points.
  5. Rewrite the most important pages. Start with the five to ten pages that should answer your questions. Answer first, concrete details, sources, a visible author and a date.
  6. Make the publisher clear. About page, author pages, contact details, and Organization schema that matches what is visible.
  7. Review monthly. Repeat the same questions, compare with the baseline and update the pages. Expect answers to vary from one run to the next.

There is sector-specific advice for e-commerce, local businesses, B2B and multilingual sites. The most frequent errors are collected in common GEO mistakes.

How to measure

  • Google Search Console. The Generative AI performance report shows impressions of your links in AI Overviews and AI Mode, broken down by page, country, device and date. According to Google it has been available to all websites since 31 August 2026. Today it shows impressions only, not which questions were asked.
  • Bing Webmaster Tools. AI Performance shows the number of citations in Copilot and Bing's AI summaries, which pages are cited, and a sample of the phrases used during retrieval. Bing points out that the numbers say nothing about placement within the answer or about a page's importance.
  • Server logs. Filter on the bots' user agents to see which pages are actually fetched and how often.
  • Web analytics. Track visits referred from the AI services' domains and what those visitors do on the site.
  • Your own prompt tests. The same questions, the same services, at regular intervals. Run each question several times: answers are not deterministic, and the researchers behind the GEO study generated five responses per query for exactly that reason.

Metrics and methods are covered in Measuring GEO results.

What is hype or unproven

  • Guarantees. Nobody can promise placement in an AI answer. Google's guidance on third-party tools says that no external tool has access to its internal systems and that a tool's predictions are the tool's own opinion.
  • llms.txt as a shortcut. llms.txt is a proposal by Jeremy Howard from September 2024, not an adopted standard. Google says Google Search does not use the file. It is cheap to add and may be useful for agents and developer tools, but there is no evidence that it produces more citations. See the llms.txt guide.
  • Structured data as a miracle cure. Schema markup makes facts unambiguous and is required for rich results, but Google says it is not needed for the AI features.
  • "40 percent more visibility". The figure is a laboratory result from 2023, not a forecast for your site.
  • Tool scores. A GEO score, ours included, measures how well a page meets known and plausible prerequisites. It does not measure whether you will be cited.
  • Influencing training data. What the models have already learned cannot be changed after the fact, and how much weight a single website carries in training is not public.

What is well supported is less exciting: pages that can be fetched, that answer a question clearly and that back up their claims are better placed than pages that do not.

To see where your own page stands, you can run a free GEO scan of a single URL without an account. It checks bot access, structured data, heading structure and publisher details, among other things, and returns a prioritised list of fixes. Terms used here are explained in the GEO glossary.

Frequently asked questions about GEO

What is generative engine optimization?

Generative engine optimization is the work of making web content easy to retrieve, understand and cite for AI services such as ChatGPT, Perplexity, Claude, Copilot and Google's AI Overviews. People also write "GEO optimization", which repeats the last word, but means the same thing.

What is the difference between SEO and GEO?

SEO aims for a position in the results and a click. GEO aims for your content to be used and linked as a source in an AI-generated answer. The technical foundation is the same. What GEO adds is the AI services' own crawlers, more weight on individual passages being clear and supported, and different metrics.

How do you optimize a website for AI search engines like ChatGPT and Perplexity?

Allow the search bots OAI-SearchBot and PerplexityBot in robots.txt, make sure the content is in the HTML, answer concrete questions directly and with evidence, and make it clear who is behind the content. Then measure by asking the same questions at regular intervals and tracking referred traffic.

How long does GEO take to show results?

There is no documented timeframe. Technical changes can take effect quickly: OpenAI cites about 24 hours for robots.txt changes. Content changes only show once the page has been recrawled and reindexed, and what models learned in training does not change until a new model is trained. Measure monthly and plan in quarters.

Kristoffer Pihlfelt

Developer and founder, Pihlfelt & Vänner

Developer and founder of the Swedish web agency Pihlfelt & Vänner since 2014. Builds Visicite and writes the GEO guide.

About the author

Share:

Related Articles