Generative engine optimization (GEO): the complete guide

Generative engine optimization is the work of making a site legible, quotable and attributable to systems that answer questions instead of listing results. That is the whole definition, and most of what follows is an argument about which parts of it are supported by evidence and which are being sold ahead of it.

This guide is written from what CiteSite measures across real sites. It is deliberately explicit about the difference between the parts that are mechanically true — a blocked crawler cannot cite you — and the parts that are reasonable inference. A field this young produces a great deal of confident advice, and the confidence is usually the first thing to check.

What changed, and what did not

Search used to end with a list. You ranked, someone clicked, and the click was the outcome you optimised for. A generative engine ends with a paragraph. It may name three brands, cite two sources, and produce no click at all — and the brand it names is not reliably the one ranking first.

What did not change is the substrate. Every major answer engine reaches pages through crawlers, and most reach them through a search index that already existed. Which means most of classical SEO still applies, and the people telling you to throw it out are wrong. GEO is not a replacement discipline. It is a set of additional requirements layered on the same infrastructure.

How an answer actually gets made

Simplifying, but not in a way that misleads, there are three stages, and knowing which one you are failing at is most of the work:

1. Retrieval. The engine turns a question into searches, fetches candidate pages, and reads them. This is where robots.txt, rendering and index presence decide everything. A page that is not retrieved does not exist at any later stage.
2. Synthesis. The model reads the retrieved passages and composes an answer. This is where the shape of your prose matters: which passages are self-contained enough to be carried into the answer, and which have to be paraphrased into anonymity.
3. Attribution. The engine attaches sources. This is where being a recognisable entity, with an author and a date, decides whether the sentence it lifted from you carries your name or not.

Almost every real GEO problem is a failure at one of those three, and the fixes have nothing to do with each other. Diagnosing which comes first.

The six things that decide it

1. Crawler access

The only item on this list that is absolute. An engine whose crawler is disallowed cannot read your page, so no amount of anything below produces a citation from it.

The mistake is almost never deliberate. It is a robots.txt tidied up months ago, where a block of AI user-agents went in together —GPTBot, CCBot, and Google-Extended with them. The first two are training crawlers and cost you nothing. The third removes the site from Gemini and Google’s AI answers while leaving every ranking intact, so nothing looks broken and no report says otherwise.

The three categories are worth learning once, and the crawler checker reports them agent by agent, naming the line responsible for each block.

2. Rendering

The quiet one. If your body text is assembled in the browser rather than delivered in the response, most crawlers see an empty page. Google runs JavaScript, eventually and inconsistently; the answer-engine crawlers largely do not.

This is worth checking before anything else on the list, because it invalidates every other measurement. A site whose pages arrive with forty words of text has one problem, and it is not its schema.

3. Entity clarity

When an assistant reads your name on your site, in a review, and in a forum thread, it has to decide whether those are one company. Everything it later says about you is said about the answer it settles on.

The common failure is not an absent declaration. It is three that disagree: Acme in the schema.org Organization, Acme Software in og:site_name, Acme — project tracking for agencies in the title. Each is correct, written by a different person at a different time. To a machine matching strings, they do not match.

The fix is unglamorous and cheap: pick one name, declare it in a schema.org Organization, use it in every declared field, and add sameAs — LinkedIn, Crunchbase, GitHub, Wikipedia if you have one. sameAs is the only item in this entire guide that recruits sources you do not control.

4. Passages that survive being lifted

A quote is one passage, alone, on someone else’s screen. A paragraph that opens “This means that…” refers to something the reader cannot see, so the engine paraphrases it — and a paraphrase carries no link back to you.

This is the uncomfortable part of GEO, and it is a fact about engines rather than about writing: a clearly structured page making a weak argument is more quotable than a brilliant one written as a single unbroken essay. The recommendation is not to write worse. It is to open paragraphs with their subject, break the page under subheadings, and put anything enumerable in a list or a table — the formats that come out of a page in one piece.

Specificity matters more than length. A page that says a process is “significantly faster” gives an answer nothing to carry; one that says it takes four minutes gives it a sentence. If there is no number behind a claim, that is the more useful finding.

5. Structured data

Markup does not make a page worth quoting. What it does is remove inference about facts a machine would otherwise guess — who published this, when, what it is called, what it costs — and stop a page that is worth quoting from being misattributed or misdated.

Four types do most of the work: Organization with sameAs, Article with an author and both dates, Product with an Offer, and BreadcrumbList. More markup is not better markup, and the most common failure is not a missing type at all — it is a block that does not parse. Those are invisible in a browser, present in your source, and skipped by every consumer, which is why they survive on pages for years. The types that earn their place covers the ordering.

6. Corroboration

The part you cannot do on your own site, and the part that matters most over time. An assistant deciding whether to name a company is influenced by whether other sources it already reads say the same things — comparison pages, review sites, forum threads, industry write-ups.

This is where GEO stops being technical work and becomes ordinary marketing under a new name: get mentioned accurately in the places your market already reads. What is new is only the reason. It is no longer about a link passing authority; it is about a claim being corroborated somewhere the model looks.

What the evidence does not support

Three claims are made routinely and none of them stands up.

That anyone can guarantee a citation. Generated answers vary between users, sessions and days, and the systems producing them are neither documented nor stable. Certainty about them is not available to be sold.

That llms.txt is a ranking factor. No major engine has committed to reading it. There is still a reasonable case for publishing one — it is a page on your site, it is crawled like any other, and it states your entity and your key URLs in the flattest available form — but the honest case is narrower than the one being made for it.

That GEO replaces SEO. The retrieval stage of most answer engines runs on a search index. Ranking well is still a large part of being retrieved, and a site that has abandoned classical SEO to chase GEO has usually weakened the first stage to improve the second.

The order to do the work in

Roughly in descending order of consequence, and the order is the advice — most of the value is in the first three, and most published GEO advice starts at the fourth.

1. Check crawler access. Five minutes, and it stays correct for months.
2. Check that your body text arrives in the HTML. If it does not, stop here.
3. Declare one entity, consistently, with sameAs.
4. Fix any structured data that does not parse, then add Article and Product where they apply.
5. Rework your most important pages for self-containment: subheadings, subjects at the start of paragraphs, numbers instead of adjectives.
6. Work on corroboration off-site, which is slow and does not stop.

The free GEO audit scores a site across exactly these six and ranks what to fix; the AI visibility checker reads the same crawl as a chain and tells you where it breaks.

Then measure it

Everything above is readiness — whether a citation is possible. Whether one happens is a different question, and it can only be answered by asking the engines what they say and repeating it on a schedule. The measurement guide covers how to do that without over-reading the result, which is easier to get wrong than the technical work.