How to get cited in AI answers — AEO and GEO, without the hype

Two things are true at once. The way people ask questions is changing, and almost everything sold as "AI SEO" is the old advice with new labels.

This is the part that is actually different, the part that is unchanged, and the specific work worth doing.

What actually changed

In a list of links, the click is the goal. In a generated answer, the answer is the product and your page is a source that may or may not be named.

That shifts what a page has to do. It is no longer enough to be relevant to a topic; the specific claim has to be extractable, attributable and correct enough to survive being quoted next to a competitor's.

Three consequences follow, and they are the whole of AEO and GEO worth the acronyms:

1. Structure beats prose. A system pulling an answer needs a passage that stands alone. A meandering introduction that gets to the point in paragraph six is invisible to it.
2. Facts beat adjectives. "Fast, reliable service" is unquotable. "Verification review takes up to five business days" is a sentence an assistant can lift and attribute.
3. Entity clarity beats keyword coverage. If a model cannot tell which company you are, it cannot attribute anything to you.

What did not change

Crawlable, server-rendered HTML. Fast responses. Content that answers a real question. Internal links that establish which page is the authority on a topic. Accurate structured data. Consistency between what your site says and what every other profile says about you.

If a page cannot be indexed by a traditional crawler, it is not going to be cited by an assistant either. The indexing checklist is the prerequisite, not an alternative.

The crawlers, and what each one controls

This is where most advice goes wrong, because the tokens do different jobs and blocking the wrong one costs visibility you wanted.

Two exact points, because they are the ones people get backwards:

• Google states that "Google-Extended does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search". Blocking it is a training and grounding decision, not a search decision.
• OpenAI states that disallowing GPTBot "indicates a site's content should not be used in training generative AI foundation models", and directs publishers to use OAI-SearchBot to manage search opt-outs separately. Blocking the training crawler does not remove you from ChatGPT's search results.

Anthropic and Perplexity publish their own tokens as well. Check their current documentation before writing rules — these lists change, and a stale robots.txt is how sites quietly opt out of something they wanted.

The work that makes a page quotable

Answer first, under the question. Use the question as the heading, then answer it in the next two sentences. Everything else — nuance, caveats, process — goes below. This is also better for humans, which is the tell that it is real advice rather than a trick.

Write extractable facts. Numbers, thresholds, timelines, limits, definitions. Vague copy cannot be cited because there is nothing in it to cite.

Keep facts verifiable and dated. Say where a figure comes from and when it was checked. Anything that expires — a price, a policy, a version — needs a visible date and a real review cycle.

Add an FAQ where questions genuinely differ from the body. Not a duplicate of the article in question form; the follow-ups a reader would actually ask.

Use tables for comparisons. A comparison expressed as prose is far harder to extract than the same comparison as rows.

Make the entity unambiguous. One Organization node, one stable @id, one name, one description, the same sameAs profiles everywhere. If your company name is also a common word or another business's name, this is most of the battle.

Serve it as HTML. Assistants generally do not execute your JavaScript. If the content only exists after hydration, assume it does not exist.

What to ignore

"AI visibility scores." Nobody outside the model providers can measure citation share reliably. Tools that claim to are sampling prompts and extrapolating.

Keyword-stuffed "AI optimised" pages. Generated filler written for machines reads exactly as it is, and the systems it is aimed at are trained to be sceptical of it.

llms.txt as a mechanism. The proposal is fine and costs nothing to publish, but no major assistant documents reading it. Treat it as a cheap bet, not a channel.

How to measure it

Three things are actually measurable today:

1. Referral traffic from assistant domains in your analytics — small numbers, usually high intent.
2. Branded search volume in Search Console. Assistants create demand they do not always send: people read an answer, then search the company name.
3. Whether the answer is correct. Ask the assistants what your company does, and what it costs. If the answer is wrong, that is a content and entity problem you can fix.

Google does not break out AI Overview impressions separately in Search Console, so anyone quoting you an exact AI Overview click count is estimating.

Where this fits

AEO and GEO are not a separate project. They are what good technical SEO, honest writing and a clean entity look like when the consumer of the page is a model rather than a person scanning ten links.

We build sites this way by default — web development for the rendering and structure, SEO for the entity and content architecture. If you want to know how a specific site reads to a machine today, send us the domain.