Generative engine optimization, or GEO, is the practice of getting your pages cited in the answers that ChatGPT, Perplexity, Gemini, Claude and Google’s AI Overviews write. It matters because the answer now sits between your buyer and your site. When Google shows an AI summary, people click through to a regular result on about 8% of visits, against 15% when there is none, according to Pew Research Center’s analysis of 900 US adults’ browsing in March 2025. Ranking still gets you into the pool. Whether the answer names you is the newer question, and the harder one to answer from the outside.
Most guides to GEO are lists of tactics, and most of the tactics were written before anyone measured anything. On October 28, 2026 we put ten questions a B2B software buyer might ask to five engines and logged every source their 50 answers cited: 440 citations, and a clear enough picture of who the engines trust and why.
What GEO is and where it came from
The practice is writing and structuring pages so that an AI answer engine retrieves them, quotes them and cites them. The engines are ChatGPT, Perplexity, Gemini, Claude, and the AI Overviews and AI Mode that Google now places above its results. Classic SEO earns you a ranked link; GEO earns you a mention inside the answer, usually with a link and sometimes without.
The term is younger than it sounds. Pranjal Aggarwal and five co-authors coined it in November 2023 in a paper called “GEO: Generative Engine Optimization”, which proposed a black-box framework for raising a source’s visibility in generated answers and reported that it could “boost visibility by up to 40%” on GEO-bench, a benchmark the authors built for the purpose. The 40% has travelled a long way since, usually without its condition. The gains were measured for sources that were already in the answer’s context, which for a B2B company is most of the problem.
Google’s view arrived in May 2026, when Search Central published its own guide to generative AI features and updated it in July: “optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” No new machine-readable files, AI text files, markup or Markdown are needed to appear, and Google Search ignores llms.txt-style files altogether. Other services may read an llms.txt file, so having one is optional. The same guide says you don’t need to write a second version of a page for machines, or cut it into fragments: “There’s no requirement to break your content into tiny pieces for AI to better understand it.” If a vendor has told you otherwise, the guide has a section of myths written for that conversation.
You will also meet answer engine optimization, LLM SEO and AI search optimization, which are the same job under different names. AEO started life around featured snippets and voice assistants; LLM SEO is named after the models. GEO is the word on this page because it names the thing that changed: the engine writes the answer instead of listing pages.
How an AI answer decides what to cite
An AI answer is built in two steps: the engine runs a search and pulls a set of pages, then writes an answer from that set with citations back to it. A page outside the retrieved set cannot be cited, no matter how well it is written, so the first job is being fetchable. Each engine has a fetcher of its own with its own rules, and the vendors document them.
| Engine | Fetcher for answers | Training crawler or token | What opting out of the fetcher does, in the vendor’s words |
|---|---|---|---|
| ChatGPT | OAI-SearchBot | GPTBot | “Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers.” “Each setting is independent of the others.” |
| Perplexity | PerplexityBot | none stated | PerplexityBot “is designed to surface and link websites in search results on Perplexity” and “is not used to crawl content for AI foundation models.” |
| Claude | Claude-SearchBot and Claude-User | ClaudeBot | Blocking Claude-SearchBot prevents indexing for search; blocking Claude-User stops the fetch when a user asks. |
| Google AI Overviews and AI Mode | Googlebot | Google-Extended, a robots.txt token for Gemini training and grounding rather than a crawler | A page must be “indexed and eligible to be shown in Google Search with a snippet.” Google-Extended “does not impact a site’s inclusion in Google Search nor is it used as a ranking signal.” |
The rows come from OpenAI’s overview of its crawlers, Perplexity’s crawler guide, Anthropic’s help center as updated in April 2026, and Google’s pages on AI features and your website and its common crawlers, all read on October 29, 2026.
Blocking a training crawler does not block the search fetcher from the same vendor, and OpenAI says so directly: “Each setting is independent of the others.” The fetchers that act on a user’s request are different again. OpenAI says robots.txt rules “may not apply” to ChatGPT-User, and Perplexity says Perplexity-User “generally ignores robots.txt rules”, because a person asked for the page.
For Google there is nothing extra to build. Its documentation says that “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”; a page has to be indexed and eligible to show with a snippet, the same bar as ordinary search. What is new is the measurement. Since June 2026 Search Console has carried a Generative AI performance report with impressions by page, country, device and date for AI Overviews and AI Mode. It counts impressions, not clicks, and Google rolled it out to a subset of sites first.
What the engines cited
Protocol ran this study on October 28, 2026. We put ten buyer questions to five AI engines and logged the 440 sources their 50 answers cited. The engines cite very different amounts and very different kinds of page. And none of the seven pages that rank in Google’s top ten for “generative engine optimization” is among the 50 pages they cited most.
How we ran it
The ten questions are ones a B2B software buyer might ask, on AI search, on hiring for go-to-market and on sizing a market, including the one this page answers. We wrote them ourselves and put each to five engines, once, from the US, keeping every source each answer cited in the order it appeared:
- How do I get my B2B software company cited in ChatGPT and Google AI Overviews?
- What is generative engine optimization and does it work for B2B?
- What is the difference between AEO and SEO?
- What is a done-for-you SEO and AI search service for B2B SaaS?
- Which agencies do AI search optimization for B2B software companies?
- Should I hire a GTM engineer or use a fractional GTM team?
- What does a GTM engineer do?
- How should a B2B SaaS company approach SEO in 2026?
- How do I calculate total addressable market for a B2B SaaS product?
- What is the difference between TAM, SAM and SOM for a B2B software company?
The engines were ChatGPT, Perplexity, Gemini, Claude and Google AI Overviews. We ran the questions through OpenSEO on DataForSEO’s answer endpoints, from the US, in a single run: 50 answers in all. Answers move between runs, and a July 2026 survey of 45 GEO studies found that commercial audits of the same market often barely overlap in the sources they report. This run is a baseline we will measure against again.
The 50 answers carried 440 citations between them, though seven cited nothing at all, four of those from Gemini and three from ChatGPT. The engines differ enormously in how much they cite. Perplexity averaged 19.3 sources an answer (range 15 to 30), AI Overviews 8.4 (7 to 11), Claude 6.3 (4 to 7), ChatGPT 5.8 (0 to 26) and Gemini 4.2 (0 to 8), each over its ten answers.
Who was cited
For the question this page answers, the engines went to Google first, and the two GEO questions drew the following.
| Question | What the answers cited most |
|---|---|
| “What is generative engine optimization and does it work for B2B?” (49 citations) | Google’s AI features page and its generative AI optimization guide, Google’s May 2026 announcement and its guidance on AI-generated content, Broadcastwell’s “2026 State of GEO” (listed as cited, not verified by us), and the 2023 and 2025 GEO research papers |
| “How do I get my B2B software company cited in ChatGPT and Google AI Overviews?” (43 citations) | The same two Google pages, Google’s Dataset structured-data page, Citera’s study of 350,000 B2B SaaS articles (listed as cited, not verified by us), OpenAI’s help center; and in AI Overviews, LinkedIn and YouTube |
None of the seven organic pages in Google’s top ten for “generative engine optimization” appears among the 50 pages or the 50 domains the engines cited most across the run. The other three slots in that top ten were search features rather than pages. The engines and the ranking are reading different things.
Nobody owns these answers. The most-cited domains in the whole run, developers.google.com and LinkedIn, each appeared in seven of the 50 answers, every one of LinkedIn’s in AI Overviews. YouTube appeared in six, again all in AI Overviews, and GitHub in six, five of them in Perplexity. After that came an SEO agency’s site with six, and HG Insights, Apollo and an agency’s list of agencies with five each. A domain’s count is the number of the 50 answers in which it appeared at least once.
The engines have different tastes, and the differences are large enough to plan around. AI Overviews cited LinkedIn on seven of the ten questions and YouTube on six. Perplexity leaned on GitHub and on Google’s documentation. Claude cited small agencies’ listicles, and ChatGPT a mix of Google’s documentation and agency blogs. If your buyers ask on Google, a LinkedIn post and a short video are part of the page’s job.
What the cited pages have in common
We then read the two most-cited pages for each of the ten questions with claude-seo’s page renderer. That came to 12 distinct pages once the six that topped more than one question were counted once, and 11 of them could be read; one Medium post returned a 403.
They run from 892 to 5,224 words of body text, with a median of 2,041, and carry anywhere from 4 to 24 H2 headings. Of the 11, 9 define their subject within the first 60 words, 7 name an author, and all 11 show a date in 2025 or 2026. A FAQ block appears on 7, at least one table on 6, Article or BlogPosting markup on 6, and a BreadcrumbList on 8. The pages AI answers cite define the topic early, run about 2,000 words, name an author and show a recent date.
The seven organic pages that rank for “generative engine optimization” look similar on the surface. Six of the seven run 1,000 to 2,550 words of body text and the Quora thread’s answers run to about 3,500; five of them have 9 to 12 H2s, and four have a FAQ. What they lack is evidence. Of the six that are not the Quora thread, one links Google’s guide, one cites the paper that coined the term, and none shows a measurement of its own.
Does GEO work for B2B?
Yes, in a narrower sense than most vendors claim. The answers do take clicks. In Pew Research Center’s browsing data from 900 US adults in March 2025, people who saw an AI summary clicked a regular result on 8% of visits, against 15% when there was none. A link inside the summary itself was clicked on just 1% of visits. Google’s reply, in a May 2025 post on succeeding in its AI experiences, is that the clicks that do arrive are “higher quality, where users are more likely to spend more time on the site”, a claim it makes without a number. Fewer visits and better ones can both be true at once.
The measured gains are for pages already in the room. The 2023 paper’s “up to 40%” was measured on sources already in the answer’s context, and the July 2026 survey of 45 studies is blunt about what follows from that. Topical relevance and position within the context are the most reproducible levers. Generic heuristics transfer poorly between settings. No reviewed technique shows a stable, long-run, cross-platform causal effect on whether a page gets found in the first place. Rewording a page changes how it is cited once it has been retrieved, and does nothing to get it retrieved.
AI search also favors other people’s pages about you. A September 2025 study by Mahe Chen, Xiaoxuan Wang, Kaiwen Chen and Nick Koudas compared ChatGPT, Perplexity and Gemini with Google and found that AI search strongly favors third-party, authoritative sources over brand-owned and social content. Our run showed something close: on both agency questions the most-cited page was a list of agencies, though agencies’ own pages were cited too.
What a B2B software company can control is being retrievable (indexed, fetchable by the search bots and eligible for a snippet), being the kind of page that gets cited, and being present on the third-party pages the engines already use for its buyers’ questions. Generating more text is not on that list. The engines already generate; what they are short of is sources worth citing. That puts what matters upstream of generation: in the context a page is written from, the evidence it carries and the person who stands behind it. None of it guarantees a citation, and nobody can offer one. As Google’s guide puts it, “No third-party tool has access to our internal ranking or AI systems.”
Where to start, and in what order
Measure first
Write down ten questions your buyers ask before they buy, put each one to ChatGPT, Perplexity, Gemini, Claude and Google, and log every source each answer cites by domain and URL. It is a morning’s work by hand or an automated run on DataForSEO’s answer endpoints, and at the end of it you know who is cited, on which engine, and whether you are. Everything that follows is ordered by that list.
Let the fetchers in
Check robots.txt for OAI-SearchBot, PerplexityBot and Claude-SearchBot and decide each one deliberately, remembering that blocking a training crawler is a separate decision from blocking a search fetcher. Make sure the pages you want cited are indexed and not blocked from showing a snippet. A decision can be three lines:
User-agent: OAI-SearchBot
Allow: /
User-agent: GPTBot
Disallow: /
That keeps the site in ChatGPT search answers while opting its pages out of training, because OpenAI treats the two settings as independent. Each engine’s owner documents which fetchers it sends, and we go through the five one at a time in our guide to getting cited, engine by engine.
Write in the shape that gets cited
Define the thing in the first sentence. Go as long as the question needs, which for a buyer’s question is usually about 2,000 words. Name a person as the author and show the date. Use a table where there is a comparison and a numbered list where there is an order, and mark the page up as an Article with a Person, an Organization and a BreadcrumbList. Keep a FAQ for questions people really ask, but not for the markup: Google’s documentation changelog recorded on May 8, 2026, the day after it took effect, that FAQ rich results “will no longer appear in Google Search starting May 7, 2026”.
Put a person on every page
Of the 11 pages the engines cite, 7 name a person and all 11 show a date, and we don’t think that is a markup trick. A named author is someone who stands behind the claims; a date is a promise that the page was true when it was written. Everything a customer reads should pass a person before it ships, an editor who checks every claim against the product and the docs and cuts what a model invented. A model can draft a page in an afternoon. What gets the page cited is that someone who knows the product has read it and put their name to it. That matters more with AI writing, not less. The engines are looking for sources they can trust, and a page nobody checked is not one.
Say something only you can say
Google’s phrase for it is “non-commodity content”, meaning a unique viewpoint or first-hand expertise. For a B2B company that usually means your own data, a worked example from your product, or a position you are prepared to defend. A page that restates the top ten results is retrievable but gives the engine nothing worth citing.
The hard part is rarely having something to say. Most B2B companies know things about their product and their buyers that no competitor can copy; the knowledge lives in people’s heads and old decks, so every new page starts from a blank chat and comes out generic. A company should own its context: positioning, voice, product and customer knowledge, and the terms it uses, written down once in a repo the team owns. Then every page and every AI tool starts from it instead of from nothing. That is how Aptible’s team drafts from its own product data, and it is the one thing no vendor can sell you off the shelf.
Earn the third-party pages
Go back to the list from the first step. The lists, directories, LinkedIn posts and videos the engines cite are the places to be, so pitch inclusion with facts you can stand behind. Which outside pages get cited differs by engine, and we list them for each. Don’t buy mentions; Google’s guide names “seeking inauthentic mentions” among the myths to ignore.
Measure again every month
For Google, use the Generative AI performance report in Search Console. For everything else, re-run the ten questions and count citations per engine. Shares move between runs, so compare months rather than days. Count the citations yourself and take no tool’s visibility score as ground truth: no third party can see inside the engines, and the 2026 survey found little overlap between commercial audits. The loop earns its keep when something ships every cycle: a measured list, and the pages it calls for.
Can a tool do this for us?
A tool will show you the numbers on this page for your own questions, and that is worth having. The parts it leaves to you are deciding which pages to write, writing them from what your company knows, and checking them. Those are the parts we run in the Search Visibility System, every business day: research, strategy, drafts, fixes and reporting, written from your company context and checked by an editor. Nothing goes live without you.