AI SEO, GEO and AEO: the same practice under three names

AI SEO is what most people call this, and the measured demand reflects it: 6,600 searches a month, and a search page that is half explanation and half product listing. The precise names for the same practice are Generative Engine Optimization (GEO) and Answer Engine Optimization (AEO).

Both are the descendants of traditional Search Engine Optimization (SEO). They describe the work of publishing content so that generative answer engines — ChatGPT Search, Perplexity, Claude, and Google's AI Overviews and AI Mode — can read it accurately, quote it, and recommend the business behind it.

In traditional SEO the objective is a position among ten blue links. In GEO and AEO the objective is being the source the answer was synthesized from. That is a different thing to be good at, and it is what the rest of this page is about.

Why there are three names for one practice

The explanations on that search page agree with each other. What they disagree about is the name.

The names are a vocabulary, not a taxonomy. One practice; three names circulating for it. What follows is what the practice consists of.

The three pillars, and what the evidence says about each

All three are about the same moment — what the engine finds when it reads the page:

text
  the engine retrieves pages
            │
            ├── it reads the llms.txt index  ← pillar 1
            ├── it reads the JSON-LD graph   ← pillar 2
            └── it waits for the first byte  ← pillar 3
            │
            ▼
      it synthesizes one answer
            │
            ▼
      it cites the sources it used

1. Machine-readable context: the /llms.txt proposal

The llms.txt proposal — published by Jeremy Howard in 2024 and maintained at llmstxt.org — gives a site a clean, Markdown-formatted index of its purpose, its services, its pricing and its documentation.

It is worth being precise about what that is worth, because the popular version of the claim is not true. The file is a proposal, and the evidence says it is not a citation factor. Google's own guide to generative AI features in Search states that creating one "will neither harm nor help" a site's visibility, and that AI Overviews and AI Mode need nothing beyond being indexed and eligible for a snippet. Two large studies agree: one measured citation frequency across roughly 300,000 domains and found no relationship between publishing the file and being cited; the other measured average citations across 37,894 AI-cited domains and found 6.8 with the file against 6.7 without it — a gap of one tenth of a citation that no test separates from noise. No AI service has said it uses the file.

What the file does help is the narrow case it was written for: coding assistants and retrieval agents that fetch documentation on request. The format, the second version's changes, and the measurements themselves are in What is llms.txt.

2. High-fidelity Schema.org graphs

A JSON-LD block states your entity type, your services, your opening hours, your FAQ pairs — in a form a machine can lift without paraphrasing your prose. That is its whole value: it makes the site legible to a machine, and it keeps your pages from quietly disagreeing with each other.

What it is not is verified truth. The markup is a declaration by the site owner, not a certification, and a declaration is worth exactly what the site's honesty is worth — Google's guidance is explicit that structured data must reflect content a reader can see on the page. It is not a ranking lever either; it is a labelling. And it does not create the answer: the answer has to exist on the page before it can be marked up. FAQ schema covers one block in detail, including the two dates that changed what it is worth, and LocalBusiness schema covers the one that matters most to a local business.

3. Crawlability and the first byte

AI crawlers fetch with aggressive timeouts. A site built as a client-rendered single-page application makes the crawler wait for a script to run before anything readable arrives, and some of them stop waiting. Serving pre-compiled static HTML from the edge removes the one thing a crawler cannot wait for.

That is a failure mode removed, not a citation earned. It is also the cheapest of the three, because it is mostly a property of how the site is built. What is Time to First Byte covers the measurement, and What is Anycast DNS covers what "the edge" means in practice.

What an answer engine does before it cites you

The three pillars are artefacts. The four steps are the process, and they are where a citation is actually won or lost:

1. It decides, per crawler, who may fetch you. A crawler that is refused never sees the page.
2. It looks for facts it can lift. This is where the markup earns its place, and where a claim written only in the marketing copy is lost.
3. It checks whether it may use what it found. Usage permission is a different question from access.
4. It fetches the text. A promise of text is not text — a manifest that describes a page an agent never reads is housekeeping, not a lever.

And the part no markup can do: the answer has to be quotable and attributable to you, which means the page has to say something a machine can quote and a reader can check. How to get cited by ChatGPT is the step-by-step version.

What is only a well-meant claim

Worth stating plainly, because the practice's own marketing is full of these:

• "Add schema and you will be cited." Schema states facts. It does not request a citation, and no published measurement credits it with one.
• "Publish llms.txt and the engines will cite you." Four independent sources say the opposite, and one of them measured 300,000 domains.
• "Fast servers make you rank in AI." Speed removes a timeout. It is not a request for a citation, and it is not what makes an answer worth quoting.
• "AI SEO is a new discipline." The comparison pages that rank for it are titled "AEO vs GEO vs SEO" — the industry's own way of saying they are three names for one thing — and the top answer on their biggest thread is "AEO is SEO".

None of this makes the practice pointless. It makes it smaller than the marketing suggests: three things a site can do, and one thing it cannot buy.

The questions this practice splits into

Each of these is a separate piece of work with its own evidence, and each is worth reading on its own terms rather than as a bullet on this page:

Where Zeso sits

Every site on Zeso is set up for the practice by default rather than by a line of copy:

• an /llms.txt manifest, generated with the site and pointing at the site's own pages;
• Schema.org emitted from the same details the page already shows, so the markup and the page cannot drift apart;
• pre-compiled static HTML delivered from the edge, so a fetch never waits on an origin.

That is the whole claim: the practice's three pillars, done by default. Whether an engine cites you is still a question about your page, and no configuration answers it.