Generative engine optimization: the practical guide

Generative engine optimization takes its name from a research paper first published in November 2023 and accepted at the KDD 2024 conference. It described generative engines as systems that retrieve sources and summarise them with a language model, and it proposed ways for the people who write those sources to raise their visibility inside the answers^[1]. The aim is a place in the answer, as a cited source or a named brand, rather than a position in a list of links.

Google’s view is that the label changes little. Its guide to generative AI features says that, from Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO, and it treats answer engine optimization the same way^[2]. For its own AI features, AI Overviews and AI Mode, Google says there are no additional requirements to appear and no special optimizations needed^[3].

Both statements are true, and the useful position sits between them. The foundation is shared: a page that a search engine cannot crawl, index and trust will not be quoted by the AI features built on that engine. What changes is practical: where the result shows, which crawlers, such as PerplexityBot, need access, what a good passage looks like and how success is counted. GEO, in this guide, is that difference, done on top of sound SEO and never instead of it.

Every engine that cites sources works in two stages: it retrieves pages, then a model writes an answer grounded in them. Google describes its version plainly. Its generative AI features rely on retrieval-augmented generation, also called grounding, which uses the core Search ranking systems to fetch relevant pages from the Search index, and on query fan-out, a set of related searches the model issues at the same time to gather more results^[2]. Google’s own example: a question about fixing a lawn full of weeds fans out into searches about herbicides, removing weeds without chemicals and preventing weeds^[2].

Microsoft describes the second stage from the content side. Assistants such as Copilot break a page into smaller pieces, a process it calls parsing, judge each piece for authority and relevance, and assemble an answer from pieces of several sources^[4]. Bing’s reports for site owners speak of grounding queries: the phrases the AI used when it retrieved content that was then cited^[5].

ChatGPT and Perplexity retrieve with crawlers of their own. OpenAI says OAI-SearchBot is used to surface websites in ChatGPT’s search features, and that sites opted out of it will not be shown in ChatGPT search answers^[6]. Perplexity says PerplexityBot is designed to surface and link websites in its search results, and recommends allowing it for a site that wants to appear there^[7].

Two consequences follow, and the rest of this guide is built on them. First, retrieval is a gate: a page that is not in the index an engine searches, or that its crawler may not fetch, cannot be quoted however good it is. Second, the answer is assembled from passages, and fan-out turns one question into several searches, so a page is judged by whether one of its passages answers one of those searches cleanly.

Answers also vary from run to run and from tool to tool, so GEO is judged across many answers, never from a single one. How to measure GEO results shows how to hold everything else still while you measure.

The paper that named GEO tested nine ways of rewriting a source page on GEO-bench, a benchmark of 10,000 queries drawn from nine sources, 80% of them informational^[1]. Visibility was scored inside the generated answer: how many of the answer’s words came from a source, weighted towards the top of the answer, and a second, model-judged impression score^[1].

Read the limits before acting on the numbers. The answers came from a two-step test engine that fetched the top five Google results for each query and had one model write the answer; the check on Perplexity used a subset of 200 queries, with the sources uploaded as files^[1]. Visibility meant a share of the answer’s words, not visits or sales. The paper measures what a model does with pages it has already been given. It says nothing about how a page gets into that set in the first place, which is the harder half of the work.

What the findings support is a direction, not a recipe. The edits that helped all make a passage more specific and easier to check: a number with its source, a quotation from someone who knows, a clean sentence. Keyword stuffing, the classic shortcut, did not help^[1]. That matches Google’s advice to write non-commodity content with a point of view of its own, such as a first-hand review, rather than a restatement of what is already online^[2]. Microsoft says the same in other words: anchor claims in measurable facts, and avoid words such as innovative that mean little without specifics^[4].

Retrieval comes first because nothing else counts without it. For Google, 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^[3], and the site must be included in Search generative AI features in Search Console^[2], which is the default setting^[8]. For ChatGPT and Perplexity, their search crawlers must be allowed to fetch the page^[6][7]. Six checks cover most sites:

1. 1Read robots.txt on every host you serve, not only the main domain. The file is read from the top of each scheme and host, so blog.example.com needs its own^[9]. Check the groups that apply to Googlebot, Bingbot, OAI-SearchBot and PerplexityBot.
2. 2Check that robots.txt itself answers. When the file returns a server error, the standard tells crawlers to assume the whole site is disallowed^[9].
3. 3Check the layer in front of the site. Google lists crawling allowed by any CDN or hosting infrastructure among the fundamentals^[3], and Perplexity documents firewall rules that let its crawlers through^[7]. Whether the rules hold is a test in the server logs, one of the checks in the GEO audit checklist.
4. 4Check indexing and snippets for the pages that matter. The URL Inspection tool shows the HTML Googlebot received, and a nosnippet, data-nosnippet, max-snippet or noindex rule limits what Google can show from a page, in its AI features as elsewhere^[3].
5. 5Keep important content in text. Google asks for it in textual form^[3]; Microsoft warns that answers hidden in tabs or expandable menus, in images or only in PDFs may be skipped or read less accurately^[4].
6. 6Decide about training separately. Blocking a training crawler such as GPTBot is a different choice from blocking search, and OpenAI keeps the two settings independent of each other^[6].

Many sites close this gate without meaning to. In the October 2026 edition of QueryRev Research’s State of AI crawler access study, 14.2% of the 6,294 sites read shut at least one AI crawler out of the whole site by name, and 2.7% shut out OAI-SearchBot, the crawler ChatGPT search depends on^[10]. Some of those rules are deliberate. Others were written years ago for a crawler that has since changed its job.

Which crawlers to allow is a decision with trade-offs, made crawler by crawler in GPTBot, OAI-SearchBot and robots.txt. One file can wait. An llms.txt, a proposed Markdown summary of a site for language models^[11], is ignored by Google Search, and Google says keeping one will neither help nor harm a site there^[2]. Whether your site needs one depends on which other systems read it.

Once a page can be retrieved, the question is whether a passage of it is the best available answer to one of the searches behind a prompt. Three habits do most of the work, and none of them is new to good editors.

1. 1Answer first, under the question. Put the question a buyer asks in a heading and the answer in the first one or two sentences under it, then the detail. Microsoft notes that direct questions with clear answers mirror the way people search, and that assistants can often lift such pairs word for word^[4].
2. 2Write sentences that stand alone. A passage lifted into an answer loses the paragraph around it, so name the subject instead of writing “it”, and keep the condition with the claim: “The Team plan includes five users” rather than “It includes five”. Microsoft calls this self-contained phrasing^[4].
3. 3Be specific and checkable. A figure with its source, a price with its date, a limit stated as a number, a quotation from a named role. These are the edits the GEO paper found to raise visibility most^[1], and the kind of claim Bing asks publishers to support with examples, data and cited sources^[5].

Two things not to do. Do not cut pages into fragments for machines: Google says there is no requirement to break content into tiny pieces, and no ideal page length^[2]. And do not publish a page for every variant of a question: Google says that creating separate content for fan-out queries mainly to manipulate rankings or AI answers breaks its scaled content abuse policy^[2].

Microsoft’s advice to make answers snippable and Google’s advice not to chunk are less opposed than they sound. A page organised as sections, each answering one question in full under a heading that states it, satisfies both. It is also the page a person finds easiest to scan, which is the test Google keeps returning to.

Fan-out also changes what a complete page looks like. If one question becomes searches about cost, alternatives and how to start, a page that answers the obvious next question as well as the first is useful for more of those searches. Before you write a new page, check whether an existing one should carry the answer. Restructuring the pages you already have, so that each answer sits under its question, is often the quickest gain, and it avoids two pages competing for one question.

An answer that names a brand has to be sure which brand it means and what is true of it. That is entity work: the same name, description, products, prices, places and policies, stated the same way on your own pages and wherever else a model might read about you. A brand whose pricing page, about page and directory listings disagree gives an answer three versions to choose from.

• One fact, one wording. Write the basic facts once (what you sell, to whom, where, at what price and under which policies) and use that wording on the home page, the about page, product pages and profiles. Bing advises aligning text, images and video so that they represent the same entities and products^[5].
• Organization structured data on the home page. Google says it helps disambiguate an organization from others, and that it has no required properties^[12]. Its sameAs property lists pages that unambiguously indicate who you are, such as a Wikidata entry or an official website^[13]. Whatever the markup says must match the visible text^[3].
• Business and product listings. Google names Merchant Center feeds and Business Profiles as ways for products and local details to appear in AI responses^[2]; Bing names Bing Places for Business for addresses, hours and contact details^[5].
• Mentions you earned. Google’s AI features can show what is said about products across the web, in blogs, videos and forums, but Google warns that seeking inauthentic mentions is not as helpful as it might seem, because its core systems favour quality and other systems block spam^[2]. The sources an engine cites for your prompts show where a mention would count: open them, and earn a place in the ones that fit.

Structured data deserves a sober place on this list. Google says it is not required for generative AI search and that no special schema.org markup is needed^[2]. Use it because it states your facts in a form machines read without guessing, not as a key to AI answers.

There is no single AI ranking to track. An answer is written fresh for each prompt, so measurement starts with a fixed set of prompts: the questions your buyers ask, worded as they would ask them, asked the same way on a schedule. For each answer, record whether the brand is named, whether one of your pages is an AI citation, where in the answer it appears and which competitors are named. Across the set, those records add up to your AI search visibility.

The engines now report part of this themselves: Search Console counts impressions of your links in AI Overviews and AI Mode^[14], and Bing Webmaster Tools counts citations in Copilot and Bing’s AI summaries^[5]. Each report sees one engine and none sees a mention without a link, so the prompt set stays the spine of the measurement; How to measure GEO results compares the reports and sets out the scoring, step by step.

GEO work competes for the same hours as the rest of SEO, so it needs an order. Put each job through three questions: what could it bring in, what has to happen before it, and can it be done with what you have. A blocked crawler comes before any rewrite, because a better page that cannot be fetched earns nothing.

A 90-day plan sets a sequence, not a promise of citations by a date. Answers vary between runs and engines change their models, so judge the quarter on the trend across the whole prompt set and on what the work changed that you can see: crawlers let in, pages rewritten, facts aligned. The GEO audit checklist turns the first two steps into tests, with a priority for every finding.

A new field attracts shortcuts. Google’s own guide lists several things you can ignore for Google Search^[2]:

• llms.txt and other special files, markup or Markdown for AI: Google Search does not use them.
• Chunking content into tiny pieces: Google’s systems can find the relevant part of a page that covers several topics.
• Rewriting content just for AI: AI systems understand synonyms and general meaning, so you do not need every long-tail variant of a phrase.
• Inauthentic mentions: placed mentions are not as helpful as they seem.
• Overfocusing on structured data: it is not required for generative AI search, though it still helps with rich results.

To that list the research adds keyword stuffing, which gave little to no gain inside AI answers^[1]. And treat tool scores with care. Google warns against third-party tools that promise ranking success or claim to use internal Google metrics, because no third-party tool has access to its internal ranking or AI systems^[2]. A visibility score you cannot trace back to the answers it counts is not evidence. Keep the answers, and the score becomes something you can check.