GEO: how to get cited by AI answer engines

Search has long been played out on a results page: ten links, an order, a click-through rate. That frame still exists, but it now shares the stage with another surface. A conversational assistant, a summary generated above the results, an answer engine: all of them render text and attribute only a few citations. Attention no longer spreads across ten positions, it concentrates on the handful of sources picked to compose the answer.

The acronym GEO — generative engine optimization — names that shift. The word is new; the discipline much less so. Most of what works comes down to editorial and technical practices that were already known, applied this time to a reader who does not click, does not scroll, and forgives no ambiguity.

What genuinely changes

Three things, and no more.

• The useful unit is no longer the page but the passage. A model lifts a paragraph, a definition, a list; it rarely takes a whole document.
• Selection owes as much to clarity and verifiability as to the domain’s standing. A sentence that needs its context to be understood is a sentence that cannot be used.
• The visit may never happen. Being cited without being visited becomes a result in itself, and the citation link is the only thing left that can still bring anyone back.

That last consequence is the uncomfortable one: part of what a piece of content produces becomes invisible in the usual reporting. We would rather say so than manufacture a reassuring number.

What does not change: the technical foundation

Good news — the foundation still rules, and it is the part that can be measured. A page that cannot be reached, cannot be indexed or is slow to respond will not be cited by an answer engine any better than it was ranked by a search engine. On this ground, the crawlers feeding assistants are cruder than the ones behind the established engines, not more forgiving.

1. Stable URLs, a single canonical version, no redirect chains.
2. A robots.txt that explicitly allows what you want to see cited, and a sitemap kept current.
3. A short server response time, and content already present in the HTML that comes back.
4. A heading hierarchy that describes the content, not the layout.
5. One subject per page, treated in full.

Nothing spectacular. Yet this foundation is what is most often missing when a page stays invisible, and no amount of editorial work compensates for a site the crawlers cannot read properly.

Answer the question in the first paragraph

The most profitable rule is also the most ordinary: the page answers one precise question in its first two sentences, then develops. No introduction taking a run-up, no context spread over three paragraphs. An extractor that reads only the beginning has to be able to leave with a complete, self-contained, attributable answer.

A page you have to read in full to understand will be neither cited nor read in full.

In practice: a title that poses the question or states a checkable promise, a standfirst that answers it in one or two sentences that stand up outside their page, then a body that justifies, qualifies and details. The same paragraph then serves as a short answer to an assistant and as an opening for a human reader.

Structure in questions and answers

Subheadings phrased as questions are not a writing tic: they hand the model the phrasing a user will employ. Each section then has to stand on its own — no “as we saw above”, no pronoun pointing three paragraphs back, no conclusion deferred to the end of the page.

One idea per section

A section covering two subjects will be cited for the wrong one. So we cut short, even at the cost of more subheadings, and we keep lists for content that genuinely is a list: steps, criteria, options. An enumeration manufactured to break up the page is spotted immediately and dilutes the passage instead of making it extractable.

Publish dated, attributable facts

An answer engine has to cite. It therefore favours what it can attribute without risk: a dated claim, tied to an identifiable entity, consistent with the rest of the site. A visible publication date and update date are worth more than a “recently” that never ages well.

We rule out numbers whose source cannot be named. A percentage with no origin is a handicap: it makes everything around it unverifiable, and a system that has to cite would rather do without a page it could not defend.

Get structured data right, and promise it nothing

schema.org markup settles ambiguities that text alone leaves open: which organisation is this, who wrote this article, when was it modified, which page does this breadcrumb belong to. It is a service rendered to the machine, not a ranking lever.

Two rules are enough. The markup describes only what the page already displays — markup that claims something else is at best ignored, at worst penalised. And it grants no right to any rich display: engines decide what they use, and change their minds without notice.

Stay fast and readable without JavaScript

Many crawlers read only the HTML the server returns. Whatever appears only after a script runs, after a tab is clicked or after scrolling, does not exist for them. The test takes one operation: turn JavaScript off, reload the page, read what is left. If the answer has gone, it was never published.

That condemns no modern interface: server rendering, accordions whose content is already in the document, and lazily loaded images are enough to satisfy both requirements. The only constraint is to treat the initial HTML as the content, not as a starting point.

Maintain your mentions elsewhere on the web

A model synthesises what it finds, and what it finds is not only your site. A business listing, serious trade directories, sector pages, public documentation, argued answers in specialist communities, press coverage: these mentions confirm your existence and your fields of work to systems that have no reason to take your word for it.

Consistency matters more than volume. A name written three different ways, two contradictory addresses, or a main domain that shifts from listing to listing manufacture exactly the doubt a citation system is trying to avoid. So we start by aligning what already exists before creating more of it.

What is not guaranteed

No position can be guaranteed, and no citation can be bought. Two runs of the same question can produce two different answers, resting on different sources, on the same day. A supplier promising you a place in generated answers is promising something they do not control.

Measurement, for its part, is still immature. There is no serious equivalent of rank tracking for this channel: referral traffic sometimes shows up, often badly attributed, and Search Console reports do not cover answers generated elsewhere. What remains workable is manual, and honest: keep a list of target questions, ask them periodically, note who is cited, and check whether the pages concerned are progressing by other measures.

It is little, but it is true. And it is enough to decide what to write next, which is the only thing this work has to make possible.