Key takeaways
- AI search tools cite pages from a search index, so a page has to be crawlable, indexed and allowed to show a snippet before its writing matters.
- Write passages worth quoting: the answer in the first lines, then specific facts, figures and their sources.
- Google says AI Overviews and AI Mode need no special files, markup or schema; structured data should match the visible page.
- Allow the crawlers that cite: Googlebot, Bingbot and OAI-SearchBot. Blocking GPTBot, OpenAI’s training crawler, is a separate choice.
- Research teams: publish each report’s findings on an HTML page, not only in a PDF.
- Measure with a fixed list of questions checked monthly, plus Search Console, Bing Webmaster Tools and referral traffic.
“Is SEO dead?” is the wrong question, but it comes from a real worry. Buyers now ask ChatGPT, Copilot or Google’s AI Mode and get an answer with three or four sources beside it, and the business that isn’t one of those sources may never be seen. Marketing teams want to know how these tools choose their sources, what actually helps, and which of the new advice is hype. This guide answers from each engine’s own documentation, the original research paper that named the idea, and the method this site is built with.
What is generative engine optimisation?
Generative engine optimisation is the work of making your pages easy for AI search tools to find, understand and cite when they write an answer. The term comes from a 2023 research paper, GEO: Generative Engine Optimization by Pranjal Aggarwal and colleagues, published at KDD 2024. The authors built a benchmark of queries, GEO-bench, and tested nine ways of rewriting web content to see which made it more visible in the answers.
Three findings are worth knowing, with one caution:
- Adding quotations, statistics and citations to sources worked best. In their tests these methods improved the paper’s visibility measure by 30 to 40 percent relative to unchanged pages; the abstract sums it up as “up to 40%”.
- Keyword stuffing didn’t help. It scored below the unchanged pages in the main results.
- Results varied by subject, so the authors argue for different methods in different domains.
- The caution: the tests ran on an engine the authors built on a GPT-3.5 model, plus a smaller test on Perplexity. Today’s engines are different products. Treat the paper as evidence that specific, sourced writing helps, not as a formula.
Some people call the same work AEO, answer engine optimisation. The name matters less than the habit: write passages that answer a real question so clearly that a machine can lift them and a person can trust them.
How do AI search engines choose which sources to cite?
Each engine picks sources from a search index first, so a page has to be crawlable, indexed and eligible to be shown before anything else matters; then the engine chooses the passages that best answer the question. None of them publishes a full formula. What each one does publish, in its own documentation, is below.
Two points follow from the table. First, technical access comes before writing: a page blocked to Bingbot or OAI-SearchBot, or marked nosnippet, can’t be quoted however good it is. Second, Google’s AI features can split one question into several related searches, which it calls “query fan-out”, so a page that answers a narrow sub-question clearly can be cited for a broad one.
Want to be the cited source?
Tell us the questions your buyers or readers ask, and we’ll look at which of your pages can answer them well enough to be quoted.
What actually helps a page get cited, and what is hype?
What helps is unglamorous: an answer in the first lines, specific facts with their sources, structured data that matches what the page shows, and pages the engines are allowed to read. Most of the hype is about special files and tricks that the engines’ own documentation says you don’t need.
Google is the clearest on this. Its guide to AI features and your website says there are no additional requirements to appear in AI Overviews or AI Mode, that “you don’t need to create new machine readable files, AI text files, or markup to appear in these features”, and that there is no special schema.org structured data to add. Its advice on structured data is to make sure it matches the visible text on the page.
That puts llms.txt in perspective. It is a proposal by Jeremy Howard, first published in September 2024: a Markdown file at the root of a site, or of a section of it, that gives AI tools a short summary and a list of the pages that matter, mainly for use when an assistant is answering a question rather than for training. It is cheap to publish and some tools may read it. Google’s guide says you don’t need one for its AI features, so don’t expect it to move rankings or citations.
How does this site apply GEO?
This site applies GEO through the way its pages are built, not through special files. Here is what its code does:
- Every page opens with a direct answer. Each blog post’s answer passage, like the one under this post’s title, is checked at build time to be 40 to 60 words, and the site’s writing rules ask for each question-shaped heading to be answered in its first sentence.
- One block of structured data per page, built from one source. Each page carries a single JSON-LD graph, with the organisation, the website, the page and, on posts, the article and its author, generated from the same data as the visible page. Build tests check, for example, that no price and no unconfirmed placeholder text reaches the markup.
- An llms.txt file, generated from the same list of routes that sets every page’s title and description, so it updates when a page is added. It is there because it is cheap, not because it is expected to rank anything.
- AI search crawlers allowed by name. The robots.txt file names the crawlers that fetch pages to cite them, among them OAI-SearchBot, ChatGPT-User, PerplexityBot, Claude-SearchBot, Applebot and Bingbot, and allows each one.
How do research organisations make their reports findable in AI search?
Research organisations make their reports findable by publishing the findings as a web page, in plain HTML, answer-first, with the method, the data and a citation beside them, and not only inside a PDF. If an AI answer cites a news story about your report instead of the report itself, treat it as a sign that the findings were easier to read in the news story than on your own page. One likely reason is that the news article states the finding in two clear sentences on an indexable page, while the report keeps it on page 34 of a download.
PDFs can be indexed, so keep them: they are what people print, cite and archive. But give each report a page of its own that does the work an AI engine, and a busy reader, needs.
Structured data helps when it describes what the page shows: the report’s title, authors, publication date, and a link to the dataset if there is one. Don’t mark up findings the page doesn’t state in its text. For the wider site, from navigation to accessibility, see the research website checklist, and for how a research website is planned and built, research websites.
How do you measure whether AI search cites you?
Measure AI citations with a fixed list of questions checked on a schedule, plus the reports the engines provide and the referral traffic in your analytics. No single number tells the whole story, and AI answers vary between runs, so look for trends over months, not changes from one day to the next.
Keep a simple change log beside the question list: the date, the page and what changed. When a citation appears or disappears, the log tells you whether it followed something you did or something the engine did. The SEO and AEO services page describes this kind of tracking as part of the service. For the writing itself, see how to optimise website content.