Search stopped being a list of links while most sites were still optimising for one. When someone asks ChatGPT which free PDF tool to use, or asks Perplexity how to strip GPS data from a photo, an answer comes back written in prose, with a handful of sources named underneath. Being the site named in that paragraph is a different problem from ranking third on a results page, and it has its own name: generative engine optimization.
What is generative engine optimization?
Generative engine optimization (GEO) is the practice of structuring a site so that large language models cite it as a source when they answer a question. Where SEO competes for a position and a click, GEO competes to be the sentence the model repeats — whether or not anyone clicks at all.
That difference in target changes what you optimise. A ranking is won by a page. A citation is won by a claim — a specific, quotable, attributable statement that a model can lift and stand behind.
SEO vs AEO vs GEO
The three are layers, not rivals. They are frequently presented as a succession in which each replaces the last, which is the single most misleading thing written about the subject.
- SEO optimises a page to rank in a list of links. The unit of work is the page; the unit of success is a position.
- AEO — answer engine optimization — structures a specific passage so it can be extracted into a featured snippet or a direct answer. The unit of work is a passage; the unit of success is being the extracted one.
- GEO builds the wider evidence that makes a generative model name your brand inside its own prose: clear claims, structured data, entity consistency, and corroborating mentions elsewhere. The unit of work is a claim; the unit of success is attribution.
They stack because each depends on the one beneath it. A page nobody can crawl cannot be extracted from. A passage no model can parse will not be quoted. And a claim with nothing corroborating it is one a cautious model will decline to attribute.
Does GEO replace SEO?
No — and the mechanics make this fairly clear. Every major AI search surface retrieves from a conventional index before it generates anything. ChatGPT search, Perplexity and Google's AI Mode all fetch candidate documents first and write second. A page that cannot be crawled, or that ranks nowhere for the underlying query, is simply not in the candidate set when the writing starts.
GEO is a layer on top of working SEO. If the foundation is broken, the layer has nothing to sit on.
What actually changes in the work
1. Write claims, not copy
Models quote sentences. They do not summarise your positioning. A sentence that survives extraction is self-contained, factual, and names its subject explicitly — "Quill Tools processes every file in the browser, so no document is uploaded to a server" rather than "we take your privacy seriously." The first can be lifted into someone else's paragraph and still mean something. The second cannot be quoted at all.
2. Answer one question per heading
Phrase the heading as the question a person would actually type, then answer it completely in the 40 to 60 words directly beneath. That block has to work with no surrounding context, because being pulled out of context is precisely what you are optimising for. Everything else on the page can be as discursive as you like — the answer block is the part designed to travel.
3. Let the crawlers in
Retrieval crawlers are the ones that decide whether you are eligible for citation, and they are distinct from training crawlers. OAI-SearchBot, Claude-SearchBot and PerplexityBot fetch pages at answer time; GPTBot, ClaudeBot and Google-Extended relate to training and grounding. Those are two separate decisions, and plenty of sites block both by accident while intending to block neither. The robots.txt Generator tests a URL against a rule set so you can confirm which bots you are actually admitting.
4. Mark up the structure
Structured data does not make a model like your page, but it removes ambiguity about what the page contains. FAQPage pairs questions with answers, HowTo preserves the order of a procedure, and Organization binds claims to a named entity — which matters more than it sounds, because a citation needs something to attribute to. The Schema Markup Generator produces valid JSON-LD for each of these, and our AEO structured-data guide covers which type to use where.
5. Publish something only you have
A model has no reason to cite the fourth restatement of a widely known figure. It has every reason to cite the only source of a number. First-hand measurements, tests you ran yourself, and data drawn from your own operations are the most durable citation advantage available, because they cannot be aggregated away.
6. Keep the page visibly current
Recency is a real input for anything time-sensitive — pricing, comparisons, or a query containing the current year. Two otherwise identical pages are not treated equally when one carries a recent update and current framing and the other reads as stale.
Add an llms.txt
A short Markdown file at your domain root, describing what the site is and which pages matter, costs almost nothing and gives models a curated summary instead of one inferred from your homepage. It is a convention rather than a ratified standard, and it carries one sharp failure mode: a stale file is worse than no file, because its claims get repeated verbatim with full confidence. Our llms.txt guide covers the format — and the stale-file problem we ran into on this very site.
How to measure any of this
There is no Search Console for AI answers, so the measurement is manual and that is fine. Write down the ten to twenty questions a real visitor to your site would ask. Ask each one in ChatGPT with browsing enabled, in Perplexity, and in Google's AI Mode. Record three things: whether you were named, in what context, and which sources were cited instead of you.
That last column is the useful one. The sites being cited in your place are a direct, unambiguous list of what the model currently considers authoritative on your subject — which is considerably more actionable than a rank-tracking chart. Re-run the set monthly.
What to do first
- Confirm retrieval crawlers are allowed — check with the robots.txt Generator.
- Add an answer block under each question-shaped heading on your top ten pages.
- Mark those pages up with FAQPage or HowTo schema.
- Publish or generate an llms.txt, and wire it to a source of truth so it cannot go stale.
- Run the twenty-prompt baseline and write down who gets cited instead of you.
None of this is exotic. Most of it is the thing good technical SEO was always supposed to be, aimed at a reader that quotes rather than clicks.
Frequently Asked Questions
Is GEO just SEO with a new name?
There is real overlap — crawlability, structure and authority matter to both. The genuine difference is the target. SEO optimises for a position in a list; GEO optimises for a sentence being quoted and attributed, which rewards self-contained factual claims in a way ranking never specifically did.
Do I need separate content for AI search?
No, and duplicating content for it tends to backfire. The same page should serve both: answer blocks near the top for extraction, depth below for the reader who stays. Writing a thin "for AI" version produces exactly the shallow aggregator content that models are least inclined to cite.
Which AI surface should I optimise for first?
The retrieval-based ones — Perplexity and ChatGPT search — because they fetch live and can reflect a change within days. Anything dependent on training data moves on the model release cycle instead. Optimise for the part you can actually measure.
Building the technical side? The SEO tools here cover schema markup, robots.txt, sitemaps, meta tags and readability — all free, and all running in your browser.