Generative engine optimization: the playbook
How answer engines pick what to cite, what a quotable page looks like, which crawlers to allow, how to measure it, and a 90-day plan built on evidence.
Generative engine optimization is the work of being cited when someone asks an assistant instead of searching. It is earned the same way a good search result is — crawlable pages, clear answers, credible sources — plus two things SEO never needed: third-party validation the engines trust, and measurement that counts citations rather than positions.
On this page
What is generative engine optimization?
It is the practice of getting your pages, and your brand, into the answers produced by assistants and AI search experiences. The output you are optimising for is a sentence in a generated answer with a link next to it, rather than a blue link in a ranked list. It is what our GEO services are built to do, and the rest of this guide is the method behind them.
The distinction that matters is between being mentioned and being cited. A mention is the assistant naming your brand. A citation is the assistant linking your page as a source. They are not the same, and the gap between them is large and measurable: in a study of 13,184 citations across ChatGPT, Claude, Gemini and Perplexity, brands were mentioned far more often than they were cited, with gaps of up to 31 percentage points on some engines [1].
Both matter, and they are earned differently. Mentions come from being known — from what the model absorbed, and from other pages talking about you. Citations come from having a page that is the most convenient thing to quote at the moment the answer is assembled. If you only chase one, chase mentions, because a mention with no citation still sends the buyer looking for you.
There is a counterintuitive consequence of that split, and it catches people out when they first buy a tracking tool. The better known a brand is, the wider its mention-to-citation gap tends to be: the model recommends it from memory and sources the answer from somebody else's page. In the tracking study above, a well-known software company was named in the overwhelming majority of relevant answers while its own pages were cited far less often, and in some runs not at all [1]. A citation dashboard therefore understates visibility for a famous brand and tracks it almost one-to-one for an unknown one. Know which end of that you are on before a vendor shows you a low number and calls it an emergency.
The second thing to understand up front is that none of this is deterministic. Ask the same question twice and you will get different sources, sometimes a different recommendation. Everything below is about raising the probability of being chosen, not about controlling it, and any agency describing it differently is describing something that does not exist.
Do the engines cite the same sources?
Almost never, and this is the single most important planning fact in the field. Across 1,792 query-and-domain combinations, all four major models agreed on citing the same domain for the same question thirty times — 1.7% [1].
1.7%
They also have distinct sourcing habits. Perplexity is the most generous citer at over nineteen sources per answer and leans heavily on professional networks, video and community sites. Gemini sits around eight and favours community posts and roundup listicles. ChatGPT is sparse and skews towards primary sources — official documentation, research papers, vendors' own pricing and help pages. Claude is sparsest of all and leans on review platforms and curated lists [1].
The practical consequence is that a per-engine strategy is mostly a trap. The habits differ, but the assets that satisfy them overlap: a clear page of your own, a presence on the review and listing sites, and something worth mentioning. Build those, and you are eligible everywhere.
It is worth resisting the standard advice here, because a great deal of it is wrong for business buyers. A large share of published guidance recommends building a presence on community forums, and for consumer questions that is sound. For business-to-business buying questions the tracking data goes the other way: across seventy-two buyer questions, community forum citations were effectively absent from two of the four engines entirely [1]. What was cited instead were niche listicles, review platforms and comparison posts. If your buyer is a company rather than a consumer, spend the effort there.
| Engine | Sourcing habit | What it rewards |
|---|---|---|
| ChatGPT | Few sources, primary ones | Official docs, research, your own pricing and help pages |
| Perplexity | Many sources per answer | Professional profiles, video, community discussion |
| Gemini | Moderate, community-leaning | Roundups, listicles, forum threads |
| Claude | Fewest sources | Review platforms and curated lists |
| Google AI features | Query fan-out across subtopics | Ordinary indexed pages that answer a sub-question |
Does ranking in Google get you cited?
Not reliably. In the same tracking study, only a minority of the domains the models cited also appeared in Google's top ten for the matching query, and at the level of exact URLs the overlap collapsed to a few per cent for two of the four engines [1]. Good rankings help, because they are evidence the page is useful, but they are not the mechanism.
Google's own AI features are the exception, because they run on Google's index. The company states plainly that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimisations necessary — a page must simply be indexed and eligible to be shown with a snippet [2]. What changes is the question being answered: those features use a query fan-out technique, issuing several related searches across subtopics, so a page that answers one narrow sub-question can be pulled into an answer about something broader.
That is the most actionable thing in Google's documentation. Write for the sub-questions, not only the head term. GEO vs SEO vs AEO sets out where the three disciplines actually diverge if you want the short version.
The practical way to use fan-out is to stop thinking in keywords and start thinking in question trees. Take the commercial question your buyer actually has — should I buy this, from whom, at what price — and write out the eight or ten sub-questions they would need answered to decide. Price, timeline, what is included, what is not, alternatives, risks, what happens if it fails, who else uses it. Each of those is a heading, and each heading's first two sentences are a candidate passage. A page built that way is eligible for a dozen fan-out queries instead of one.
It also has a pleasant side effect: it is simply a better page for a human. The reason this discipline is not as alien as it sounds is that the engines are, crudely, rewarding the things a well-edited trade publication has always done.
What makes a page quotable?
Assistants lift passages, not pages. The unit of success is a two-to-three-sentence block that answers one question completely enough to be pasted into a reply without editing.
That reframes almost every content decision. Length stops being a target, because a long page is simply a page with more candidate passages in it, not a stronger one. Keyword density stops mattering entirely. What matters is whether a specific claim on your page can be lifted and stand alone: if a sentence needs the two paragraphs above it to make sense, it will not be lifted, and if it is lifted it will be lifted wrongly.
The test is mechanical and takes a minute. Read any section of your page aloud starting from its first sentence and stop after three. If what you have said is a complete, correct answer to the heading above it, that section is quotable. If you are still setting up context, it is not, and no amount of further editing below that point will fix it.
The experimental work on this is now several years old and still the clearest guidance available: adding statistics, quotations from named people, and citations to sources all measurably raised a page's visibility in generated answers, as did simply writing more fluently [3]. None of that is exotic. It is what a good trade publication does by default.
Answer in the first paragraph
Two to three sentences, no preamble, no throat-clearing. If an assistant has to summarise you, it may summarise you wrongly.
Make every H2 a question
Specifically the sub-questions a fan-out would generate: price, timeline, deliverables, alternatives, risks. Then answer each one in its first two sentences.
Put a table on the page
Comparisons are the easiest structure to lift and the hardest to paraphrase badly. One good table does more than three paragraphs of the same content.
Cite real sources, visibly
Numbered references with publishers and dates. It is the cheapest credibility signal available and it is the one most pages skip.
Sign it with a real person
A named author with a page of their own, and a named reviewer who is not the author. Anonymity is a liability in every guideline published since 2026.
Which crawlers have to be allowed in?
All of the ones you want to be cited by, and they are not the same bots as the ones that train the models. OpenAI alone publishes three, with different jobs: one for training, one for live fetching during a conversation, and one that builds the index used to surface citations [4]. Block the wrong one and you remove yourself from the answers while leaving yourself in the training data, or the reverse.
Two decisions worth making deliberately. First, whether to allow training crawlers: allowing them puts your brand into what the model knows without being asked, which is where mentions come from, and it is a genuine trade against control of your content. Second, avoid snippet-suppressing directives on pages you want cited — a page that forbids a snippet has forbidden the quote.
The same logic applies to anything that makes the page hard to read without a browser. Content that only appears after a script runs, gated behind an email form, or hidden until a tab is clicked is content that may not exist as far as a retrieval crawler is concerned. This is not new advice — it is the same rendering discipline that has always applied to search engines — but the cost of getting it wrong is now paid twice.
| Crawler type | What it does | Cost of blocking it |
|---|---|---|
| Search index | Builds the index behind citations | You disappear from cited answers |
| Live fetch | Reads a page during a conversation | You cannot be quoted in real time |
| Training | Feeds the model's general knowledge | You are absent when nobody searches |
| Edge or firewall rule | Blocks all of the above silently | Everything, with no error message |
Where do citations actually come from?
Frequently not from your own site. Analyses of what assistants cite most consistently put professional networks, community platforms, video, encyclopaedic references and review or listing sites near the top [5], and the tracking study above found review platforms to be the single most-cited third-party source on at least one engine [1].
That has an uncomfortable implication for a marketing budget: some of the highest-return GEO work is not content work at all. It is getting listed accurately on the platforms your category is reviewed on, keeping professional profiles current and complete, and being included in other people's comparison posts.
It also explains why publishing your own methodology-led comparison is one of the highest-leverage pages you can write. Listicles are lifted more readily than ordinary articles, and a list you wrote with a transparent methodology is a legitimate source rather than an advertisement — provided you disclose where you sit in it.
Rank the third-party work by how quickly it pays. Review and listing profiles in your category come first: they are cheap, they are permanent, and they are what the sparsest-citing engines reach for. Professional company and personal profiles come next, because they are cited heavily for questions about firms and people and they cost an afternoon. Inclusion in other people's comparison posts is third, because it depends on someone else saying yes. Original data you publish yourself is last to pay and largest when it does, because it is the only one of the four that other people will cite without being asked.
None of this is link building in the old sense. Most of these profiles carry no followed link at all, and the value is not the link. It is that the page exists, says the right thing about you, and sits on a domain the engines already trust enough to read.
How do you measure AI visibility?
By tracking a rate over time, never by checking an answer. The same question asked twice, two days apart, produced substantially different citations from every engine in the tracking study — one model retained under a third of the URLs it had cited forty-eight hours earlier [1]. A single answer is noise.
| Measure | What it tells you | Trap |
|---|---|---|
| Mention rate per engine | Whether the model knows you | Rises with fame, not with effort |
| Citation rate per engine | Whether your pages are quotable | Understates visibility for known brands |
| Share against named rivals | Whether you are gaining or losing | Needs the same prompt set every run |
| Referral sessions from assistants | Whether any of it sends people | Small numbers, high variance |
| Which URL was cited | Which page is doing the work | Often a third party rather than you |
Fix the prompt set before you start and do not change it, because changing the questions changes the answer and you lose the trend. Twenty to forty questions per brand is directional, not statistical, and that is fine — you are looking for movement, not proof. An AI visibility audit is how we set that baseline before any content work starts.
Two free sources are worth having alongside any paid tool. Both Google and Microsoft now report on how their AI surfaces interact with your site inside their own webmaster tools, and that data is first-party: it is about your pages rather than about a sample of prompts somebody chose. It will not tell you what an assistant said about you, but it will tell you whether anything is reaching you from those surfaces at all, which is the check that costs nothing.
Write the numbers down monthly, with a sentence explaining each movement, and keep the sentences. Six months of one-line explanations is worth more than any dashboard, because it is the only record of what you changed and what happened next.
What does a 90-day GEO plan look like?
Foundations first, because they are the only part with no waiting. Confirm the assistant crawlers can reach the site and nothing at the edge is blocking them. Get the organisation and author entities correct and consistent everywhere — one legal name, one address, one phone number, spelled identically on your site and on every profile. Entity confusion is invisible to you and expensive: two spellings of a company name are two companies as far as a machine is concerned. Then set a prompt set and take a baseline.
Then the pages. Rewrite the five commercial pages that already get traffic so each opens on a direct answer, carries a table, cites real sources and is signed by a real person. Publish one piece of original data if you have any, because original numbers are the thing other people cite and quoting is how mentions spread.
Then the third parties. Claim and complete the review and listing profiles for your category, get your professional pages current, and pitch inclusion in the comparison posts that already rank. This is the slowest part and the one with the longest tail, which is why it starts in month one rather than month three.
| Weeks | Work | What you are waiting for |
|---|---|---|
| 1-2 | Crawler access, entity consistency, prompt baseline | Nothing; this is all under your control |
| 2-6 | Rewrite five commercial pages as answer pages | Recrawl and reindex, a week or two |
| 3-12 | Claim profiles, pitch inclusions, publish one data piece | Other people, which takes months |
| 6-12 | Second measurement run, compare against the baseline | Enough runs to separate trend from noise |
Three things to resist during those ninety days. Do not add a second tool because the first one showed a bad week. Do not rewrite the prompt set because the answers were unflattering. And do not start publishing volume on the theory that more pages mean more chances; the engines are not counting pages.
Most brands that are invisible in AI answers are not losing a content race. They are missing from the three or four third-party pages the assistants keep reading, and no amount of blogging fixes that.
What does not work?
Filing an llms.txt and waiting. There is no published evidence that answer engines use it for retrieval or citation, and treating it as a visibility tactic is a way of feeling busy.
Chasing one engine. The sourcing habits differ, the overlap is tiny, and a strategy built around whichever assistant you personally use is a strategy built on a sample of one.
Publishing volume. Google's position is that the ordinary quality bar applies and there is nothing extra to do for its AI features [2]; everything else is judged by the same standards as any other content. Pages produced in bulk to catch citations are the same risk they have always been, with a newer justification.
And reacting to single answers. Ask an assistant about your category on a Monday and you will get a different set of sources on Wednesday. Track the rate, act on the trend, and keep getting cited by ChatGPT as a monthly measurement rather than a weekly emotion.
Questions people also ask
Is GEO just SEO with a new name?
The foundations overlap almost completely — a page has to be crawlable, indexed and useful before any engine can cite it. What differs is the target, the formats that get lifted, and the measurement, which is a citation rate across engines rather than a position in a list.
Does llms.txt help with AI citations?
There is no published evidence that any answer engine uses it for retrieval or citation. Publish one if coding agents consuming your docs would benefit, but do not budget for it as a visibility tactic.
Should I block AI crawlers to protect my content?
Only if you do not want to be cited. Blocking the retrieval crawlers removes you from the answers; blocking the training crawlers removes you from what the model knows without being asked. Both are legitimate choices, but they are choices to opt out of visibility.
How long does it take to see AI citations?
Weeks for pages that are already indexed and easy to lift from, longer for anything that depends on third-party sources mentioning you. Expect to be tracking a rate over months rather than checking a single answer.
Which engine should I optimise for first?
The one your buyers use, which you find out by asking them rather than by reading a market-share chart. In the absence of that, work on the foundations, because they are what every engine shares.
Can I just pay a tool to fix this?
A tool tells you where you stand. It does not write the page, earn the third-party mention or fix the crawling, which is where the work is. Budget for the tool as measurement, not as a solution.
Sources
- I tracked 13,184 AI citations across 4 LLMs: here's what I learned — Orbit Media Studios (checked 11 September 2026)
- AI features and your website — Google Search Central (checked 11 September 2026)
- GEO: generative engine optimization — Aggarwal et al., KDD (checked 11 September 2026)
- Overview of OpenAI crawlers — OpenAI (checked 11 September 2026)
- Top 10 sources LLMs cite most in 2026 — Contently (checked 11 September 2026)
Want to know where you currently stand in AI answers?
Book a free 30-minute consultation with Ibrahem and Jeremy. We look at your site while we talk, and you leave with a prioritised plan.
Who wrote this
Ibrahem HelalCo-founder
Leads strategy and delivery.
Works on: SEO, GEO
Reviewed before publishing by Jeremy Zava, Co-founder.
- How to get cited by ChatGPTChatGPT cites few sources and prefers primary ones. What that means for your pages, your crawler rules and the third-party profiles it reaches for instead.
- GEO vs SEO vs AEO: what is the difference?Three labels, one foundation and two real differences. What each term means, where the work diverges, and which of them you should be buying separately.
- How to choose an SEO agency in 2026A buyer's guide to hiring an SEO agency: what the work is, what it should cost, which proposal lines matter, and the questions that separate two quotes.
- GEO for B2B CompaniesWhy B2B buying questions lean on comparison posts and vendor pages more than forums, and what a vendor should do about its own AI visibility.