Generative Engine Optimisation: How ChatGPT, Perplexity and AI Overviews Pick Citations
AI answers cite sources differently from ranked links. Structure, dates, sourcing and mentions on sites you do not own decide whether you appear.
Part of our guide to SEO, AEO & GEO.
Generative engine optimisation is the work of making a business citable by AI assistants such as ChatGPT, Perplexity and Google's AI Overviews. These systems retrieve a few pages, read them, and write one answer with citations. They favour content that is clearly structured, dated, sourced and corroborated by other sites. GEO is mostly good SEO plus extractability plus mentions you do not control.
Key takeaways
- An AI answer is one synthesised reply with a handful of citations, so the question is not where you rank but whether you are among the pages the assistant read.
- Assistants weight content that is clearly structured, states facts plainly, carries a visible date and cites its own sources.
- Being described accurately on sites you do not own matters more than anything on your own site, because assistants cross-check before they cite.
- State the machine-readable facts once, plainly and consistently: legal name, what you do, where you serve, and prices where you publish them.
- llms.txt is a proposed plain-text summary for assistants, not a ranking signal. Allowing the AI crawlers in robots.txt is the prerequisite for any of this.
- There is no separate trick. Pages that are indexed, helpful, extractable and corroborated are the pages that get cited.
How an AI answer differs from a ranked link
A traditional search result is a list. Ten links, ranked, and the user chooses. An AI answer is a paragraph. The assistant takes the question, runs its own searches, retrieves a small number of pages, reads them, and writes a single reply with a few citations attached. The user may never see a list at all. Whether your business appears depends on whether your page was among the handful retrieved and whether the assistant found something in it worth attributing.
This is true across the systems people now use. ChatGPT with browsing, Perplexity and Google's AI Overviews all work on a retrieve-then-write pattern, and all show sources. Google's documentation on AI features in Search is explicit that there is nothing special to do beyond the normal requirements for appearing in Search, and that the same crawler and the same snippet controls apply. The difference is in what happens after retrieval: the assistant is choosing passages to build an answer from, not pages to list.
For a small Dubai company the consequence is blunt. Position four on a results page still gets seen. Being the fourth-best source for an assistant that cites two means you do not exist in that answer. Generative engine optimisation is the work of being one of the two.
What assistants weight when they choose a source
Nobody outside these companies can see the exact selection logic, and it changes. But the behaviour is consistent enough to describe. Assistants prefer pages where the fact they need is stated plainly, in a sentence they can quote, rather than implied across a page. They prefer pages that show when they were written or updated, because a dated page can be judged current. They prefer pages that cite their own sources, because a claim with a reference is easier to trust than a bare assertion. And they prefer claims that other pages agree with.
Structure does the first part. A heading that names the topic, a paragraph beneath it that answers in one or two sentences, and a table where the question is a comparison give the assistant something clean to extract. Our answer engine optimisation guide covers that writing discipline in detail and we will not repeat it here; everything in it applies to assistants as much as to featured snippets.
Dates and sources are the part business sites usually skip. A services page with no date, no author and no reference to anything outside itself is a page the assistant cannot place in time or check. Add a visible last-updated line, name the standards or vendor documentation you rely on, and link to them. The references in this article are there for that reason as much as for the reader.
Why being mentioned elsewhere matters more than your own pages
The uncomfortable truth of GEO is that the strongest signal is one you do not control. When an assistant is deciding whether to describe a company as an office IT provider in Dubai, it looks for agreement: does the free-zone directory say so, does a supplier's partner page say so, does a trade publication say so, do the map listing and the company's own site give the same name, the same services and the same location. A single site describing itself is a claim. Several unrelated sites describing it the same way is a fact.
This is corroboration, and it is why two companies with similar websites can be treated very differently. One is mentioned, consistently, in many places it does not own. The other exists only on its own domain. The assistant cites the first and is cautious about the second, however well the second site is written.
The work, then, is partly off-site and partly editorial. Make sure every listing, directory entry and partner page describes the business the same way, using the same legal name and the same short description of what it does. Where a description is wrong or outdated, get it corrected. Where you have a genuine relationship with a supplier, association or publication, ask to be listed accurately. None of this is new; it is the same corroboration that moves a new domain through Google's crawl queue. Assistants simply lean on it harder.
Machine-readable facts: say the plain things once and consistently
Assistants are frequently asked factual questions about businesses: what does this company do, where is it, who is it for, what does it cost. Most business websites answer those questions nowhere in particular. The services are implied by a hero image, the location is in a footer graphic, and the pricing is a conversation. An assistant cannot cite a conversation.
Write the facts down in plain sentences, on the pages where a reader would expect them, and keep them identical everywhere. Our own custom CRM and web development pages, for example, state that projects are quoted between AED 2,500 and AED 30,000 depending on requirement. That is a sentence an assistant can lift and attribute. A page that says contact us for a quote gives it nothing.
The table below lists the facts we recommend every small business site states explicitly. Structured data such as Organization and LocalBusiness markup can carry the same facts to machines, and we add it, but the visible sentence is what gets quoted.
| Fact | Where to state it | Why assistants need it |
|---|---|---|
| Legal and trading name | Footer, about page, contact page, every listing | Matching names across sites is the core of corroboration |
| What you do, in one sentence | Home page opening, about page, directory descriptions | A quotable description of the business |
| Where you serve | Contact page, service pages, map listing | Location questions are among the commonest |
| Who you serve | Service pages, about page | Assistants match businesses to the asker's situation |
| Pricing, where you publish it | Service pages, pricing or cost pages | A stated range can be cited; a quote form cannot |
| Last updated date | Every article and service page | Lets the assistant judge whether the fact is current |
What llms.txt is, and what it is not
llms.txt is a proposal, published at llmstxt.org, for a plain Markdown file at the root of a website that gives language models a short, curated summary of the site: what it is, which pages matter, and where to find the authoritative versions of key information. The idea is sensible. A model with a small context window benefits from a map rather than having to infer one from navigation menus and cookie banners.
What it is not is a ranking signal or a guarantee of anything. It is a convention that some tools read and most assistants, at the time of writing, do not publicly commit to. Adding one does not make a site citable; it makes a citable site slightly easier to summarise. Treat it as cheap housekeeping that may pay off, not as the heart of a GEO programme. If the facts in the previous section are not stated on the site, an llms.txt file pointing at them has nothing to point to.
The practical details of writing the file, what to include and how it interacts with robots.txt are covered in our separate guide to llms.txt, AI crawlers and robots.txt. The short version: keep it honest, keep it short, and keep it in step with the pages it describes.
Allowing the AI crawlers in robots.txt
None of the above matters if the assistants cannot read the site. Each major provider publishes the names of its crawlers and how to control them. OpenAI's crawler documentation distinguishes between the bot that gathers training data, the bot that indexes for ChatGPT search, and the agent that fetches a page when a user asks. Perplexity's bot documentation does the same for its crawler and its user-triggered fetcher. Anthropic's crawler documentation explains how site owners can allow or block its crawlers.
The distinction between training and search matters for the decision. A business may reasonably not want its content used to train models while very much wanting to be cited when a model answers a question about office IT in Dubai. The documentation above lets you make that choice per crawler. What you should not do is copy a blanket block list from a forum post, which is how many sites have quietly removed themselves from every AI answer without noticing.
Google's AI Overviews are different again: they draw on the ordinary Search index, so the normal Googlebot and the usual snippet controls govern inclusion, as its AI features documentation sets out. Check your robots.txt once, deliberately, and record the decision.
Generative engine optimisation is mostly good SEO, plus extractability, plus corroboration
Strip away the acronym and generative engine optimisation reduces to three things, two of which you were already supposed to be doing. The first is ordinary SEO: a site that is indexed, technically sound, and trusted enough to be retrieved when an assistant searches. A new domain still sitting in the crawl queue is not in any assistant's answer, and our guide to the first year of SEO for a small Dubai business is where to start if that is you.
The second is extractability: facts stated plainly, dated, sourced and structured so that a passage can be lifted and attributed. The third is corroboration: the same facts about the business appearing on sites you do not own, consistently, so that the assistant can confirm what your site claims. llms.txt and crawler permissions are the plumbing that lets the first two be read at all.
That is how we approach it in our SEO, AEO and GEO service: one programme, not three, with the AI-specific work being a layer of editing and listing hygiene on top of fundamentals that have to be right anyway. If you would like to know which of the three your site is missing, get in touch and we will tell you plainly, including if the honest answer is that the fundamentals come first.
Frequently asked questions
Generative engine optimisation is the work of making a business citable by AI assistants such as ChatGPT, Perplexity and Google's AI Overviews. These systems retrieve a few pages, read them and write one answer with citations. GEO means being among the pages retrieved and having facts stated plainly enough to be quoted: structured, dated, sourced content on your site, matched by consistent descriptions on sites you do not own.
They run searches, retrieve a small set of pages and prefer passages that state the needed fact plainly, show a date, reference their own sources and agree with other pages. Exact logic is not published and changes, but the pattern is stable: clear structure, currency, sourcing and corroboration. A page that is indexed, extractable and described consistently elsewhere is far more likely to be cited than one that only describes itself.
Only marginally, and only if the site is already citable. llms.txt is a proposed plain-text summary that helps a model find the pages that matter. It is not a ranking signal and most assistants do not publicly commit to reading it. Add one because it is cheap and may be used, but put the effort into stating facts plainly on the pages themselves and into consistent mentions elsewhere.
Decide per crawler rather than blanket-blocking. OpenAI, Perplexity and Anthropic each document separate crawlers for training, for search indexing and for fetching a page when a user asks. You can refuse training use while allowing search and user fetches, which is what most businesses that want to be cited actually mean. Google's AI Overviews use the normal Googlebot, so standard snippet controls apply there.
Usually because the competitor is described consistently on more sites it does not own: directories, supplier pages, associations, publications and map listings all saying the same thing. Assistants cross-check before they cite, and several agreeing sources outweigh one well-written site. Fix your own listings so they match, pursue genuine mentions, and make sure your pages state facts plainly with a visible date.
Not in any useful sense. GEO is ordinary SEO, which gets the site indexed and trusted, plus extractability, which is the answer-first writing covered in our AEO guide, plus corroboration from sites you do not own. The AI-specific additions are crawler permissions, a plain statement of key facts and optionally an llms.txt file. We run it as one programme because the fundamentals have to be right first.
Written by Srinivasan Ramachandran, Founder, Listonics · Published Oct 7, 2026 · Last reviewed October 11, 2026 · 2,198 words
