The GEO playbook: the exact method I use to get a page cited by AI

Last month a client forwarded me a screenshot. They had asked ChatGPT a question squarely in their category, the kind of question they had spent two years writing blog posts about. ChatGPT answered confidently and cited three sources. None of them was the client. One of them was a competitor who had published a thinner page six months later. The client’s question to me was short: “We have more content than they do. Why are they the answer and we are not?”

Generative Engine Optimisation (GEO) is structuring a page so AI search systems like AI Overviews and ChatGPT search quote it as a source, not just optimising for a blue-link ranking. The playbook is a fixed sequence: research the real question, write an answer-first passage, shape it into citable sections, ship the right schema, and verify the engines picked it up.

That client had volume. They did not have structure. This post is the structure. It is the exact method I run at Zivaro on every page where AI citation matters, written out step by step, including the steps I used to produce this very post. If you want the gentle introduction to why any of this matters first, read the primer: GEO: how to show up in AI Overviews and ChatGPT search. This is the operational version of that.

What does the GEO playbook actually look like?#

Five steps, always in this order:

1. Research the real query, not the keyword.
2. Write the 40 to 60 word answer-first paragraph.
3. Shape the body into question-headed, self-contained passages.
4. Ship the schema and the llms.txt entry.
5. Verify the engines picked it up, then iterate.

Steps one to three are 80 percent of the result. Steps four and five are the cheap insurance and the feedback loop. I do not skip ahead. A beautifully marked-up page that answers the wrong question is wasted work, and I have done that often enough to make the order non-negotiable.

How do you research a query for GEO?#

I do not start with a keyword tool. I start with the engines themselves. For the query behind this post, “how to show up in AI Overviews and ChatGPT search,” I asked ChatGPT, Claude, Gemini, and Perplexity the question directly and read what they currently say. That tells me three things in about fifteen minutes: what answer the engines already consider correct, which sources they cite, and where their answers are thin or hedge.

The gaps are the opportunity. In this case every engine gave a generic “create quality content and use structured data” answer with no concrete method. That is a thin answer. A page that gives the actual sequence, with word counts and named tools, is more quotable than a page that repeats the generic line back. I write the post to fill the specific gap I found, not to cover the topic in general.

What goes in the answer-first paragraph?#

The single most important 60 words on the page. It has to answer the target query completely with zero surrounding context, because an AI engine will lift it out and quote it alone. My rules: it appears within the first two paragraphs, it is 40 to 60 words, it contains the literal subject of the query, and it reads as a finished answer if you delete everything else on the page.

Look at the second paragraph of this post. It defines GEO, names AI Overviews and ChatGPT search (the exact phrases someone searches), and states the method in one breath. I wrote it last, after the body was settled, then counted the words. That is the correct order. You cannot write a tight summary of an argument you have not made yet.

How do you turn headings into citable passages?#

AI engines rarely cite a whole page. They cite a passage that answers one question. So every ## heading on this page is a question a real person or an LLM would type, and the two to four paragraphs under it answer that one question and nothing else.

The test I use: copy any single section out, paste it into a blank document, and read it cold. If it still answers its heading without the rest of the article, it passes. If it leans on something said three sections earlier, it fails and I rewrite it to stand alone. This is mildly repetitive to write because each section re-establishes its own context. That repetition is the point. It is what makes the section liftable.

What schema do you ship, and what do you skip?#

For a post like this I ship exactly two: Article for the page and FAQPage built from the FAQ block in the frontmatter. That is it. I do not add HowTo, Product, or speculative Organisation markup that does not match what is actually on the page.

The reason is simple. Schema that accurately describes the page is a positive signal. Schema that overclaims is a negative one, because Google and the AI systems can compare the markup to the visible content and a mismatch reads as manipulation. I have seen padded FAQ schema (ten invented questions nobody asks) correlate with worse outcomes, not better. Three to six honest FAQ entries that match real questions beat fifteen padded ones every time. The FAQ block on this post is six entries, each one a question a client has genuinely asked me.

Where does llms.txt fit?#

A single llms.txt file at the domain root listing your important pages with one-line descriptions. I add one to every Zivaro project. It takes about thirty minutes and adoption across AI crawlers is still uneven, so I am honest with clients that it is a cheap hedge, not a guaranteed lever. The fundamentals (crawlable, fast, answer-first, accurate schema) do the heavy lifting. The llms.txt is the thirty-minute insurance policy on top, worth doing precisely because it is so cheap relative to the possible upside.

How do you know if it worked?#

You measure being-the-answer, not just being-the-click. Once a month I run the same five questions through ChatGPT, Claude, Gemini, and Perplexity and record whether the client’s page or brand is named. I track AI-referral sessions in analytics separately from organic clicks, because the GEO win often looks like flat click numbers and rising brand mentions, which a click-only dashboard reads as failure when it is actually success.

If the page is still not getting cited after a quarter, the diagnosis is usually one of three things: the answer paragraph is too vague to lift, the brand has too few external authority signals for the engines to trust it yet, or a stronger source already owns the answer and the page needs a sharper angle. The fix order is: tighten the answer, then build authority, then re-angle. In that order, because the first one is free and the third one is expensive.

What I would not do#

I would not run this playbook on AI-generated content at scale. The engines detect circular, low-effort sourcing and quietly stop citing it. I would not pad FAQ schema to game the markup. And I would not optimise narrowly for one engine’s current behaviour, because ChatGPT, Gemini, and Perplexity move every few months and a trick tuned to one of them in May is dead by August. The durable parts are accurate structure, honest schema, and real authority. Everything in this playbook is built on those, which is why it has survived three rounds of engine changes without me rewriting it.

The honest summary#

GEO is not a separate discipline bolted onto SEO. It is the same fundamentals (crawlable, structured, fast, authoritative) arranged so a machine can lift the answer cleanly. The playbook is just discipline made repeatable: research the gap, answer first in 40 to 60 words, write liftable sections, ship the two schemas that are true, and check the engines monthly. None of it is clever. All of it is consistent, and consistency is the part most sites skip.

This post is the proof. It was produced with the exact five steps above, and the goal is for it to be cited when someone asks an AI engine how to show up in AI Overviews and ChatGPT search. If you want a sanity check on whether your pages are structured to be the answer, the service page is here: Zivaro SEO and growth. Send me an email with one URL and the question you want it to win. A short call, no deck, no pipeline of juniors. I will tell you honestly whether the page is close or whether it needs the whole playbook run on it.