Generative engine optimization
Be the page an answer engine can quote.
GEO — also called AEO, or LLM SEO, depending on who is selling it — is not a new set of tricks. It is the older discipline of answering one question clearly, on a page that renders, with facts a reader can verify and metadata a machine can parse. Here is what that means in practice, and what the pipeline does about it.
One query per postFacts in markupNo citation guarantees
What the term actually asks of a page.
Strip the acronym away and these six requirements are the whole of it — none of them new, all of them easy to get wrong at scale.
- One question per page
- An answer engine quotes a page that answers one thing clearly. A page covering six topics gives it six half-answers and no reason to attribute any of them to you.
- The answer, early and intact
- A sentence high on the page that survives being lifted out of context: a specific claim, in the words the question was asked in, rather than a paragraph of scene-setting.
- Specifics a model can repeat
- Numbers, names, dates, jurisdictions, sizes, prices. Vague adjectives are unquotable; a detail that can be checked is what makes a page worth citing.
- Facts stated in markup
- Structured data, a canonical URL and a date on every post, so an engine can tell what the page is, who published it and when — rather than guessing from prose.
- A page that renders without help
- Real HTML, readable with one request. Content that only appears after client-side rendering is content the reader may never see at all.
- Provenance you can stand behind
- A named publisher, a real domain, a sitemap and a plain-text index. Attribution goes to whoever is easiest to identify correctly, and being identifiable is a publishing decision.
How a page becomes quotable.
Five steps, in the order they matter. The first two are editorial decisions and the last three are engineering, which is roughly the split you would expect.
- 01
Start from a question, not a keyword
Each planned post exists because a search does — with the intent behind it. Long, specific questions are where a quoted answer comes from, not head terms with a thousand competitors.
- 02
Commit the outline before writing
The H2 and H3 structure is approved with the entry, which is what stops an article wandering across four topics and answering none of them.
- 03
Say the answer in the first screen
The post opens with what the reader came for, then earns the rest of the page. Nothing important is buried under a definition of the problem.
- 04
Emit the furniture with the post
Meta title, meta description, canonical URL, structured data and sitemap entry are generated with the post rather than retrofitted a quarter later.
- 05
Make the domain legible
Sitemap, robots.txt and llms.txt come from the same published set, so a crawler that wants the whole picture can take it in one request.
What the pipeline does about it.
Not a GEO module bolted onto a writer: the plan, the outline and the metadata are the same workflow, which is why the pages come out structured rather than fixed up later.
Every post starts from a query the domain does not answer yet, so the page exists because a question does. The outline is approved with the entry, the answer is written to the top of the piece, and the metadata is generated with the article instead of by whoever remembers to open the CMS.
Multiply that by the publishing days you picked and by each market you sell into, and the readable surface of the domain grows on a schedule: one answer per page, in each language, planned from that market's own searches.
- One target query per post, with the intent behind it
- An outline reviewed before a word is written
- Meta title and description written for the result page
- Canonical URL, structured data and sitemap entry per post
- Cover and in-article images with alt text
- Internal links into the pages that convert
- One editorial line per language, each planned separately
- Crawlable HTML on a domain you own
The tactics that do nothing.
GEO has attracted a folklore of its own. This is the honest comparison between that folklore and the boring version that works.
GEO theatre
- Writing for the model, in a voice no reader would choose
- Stuffing FAQs that answer nothing a buyer asked
- Publishing near-duplicate pages on every phrasing of a keyword
- Keyword-shaped headings that leave the page unexplained
- Hidden text, invented author bios, unverifiable numbers
The boring version
- One question per page, answered in the words it was asked in
- Real FAQs: the objections a buyer actually raises
- One canonical page per question, updated rather than multiplied
- Headings that describe what the section contains
- Named publisher, real dates, checkable specifics
How to tell whether it is working
There is no citation dashboard to open, so the honest signals are indirect: impressions in Search Console for long, specific queries; the same questions arriving in sales conversations earlier than they used to; branded and direct traffic growing rather than only referral clicks.
And the signals stay indirect by design: no engine tells a site it was quoted. That is why the plan is judged on what it publishes — one question per page, answered plainly, in the language it was asked in — rather than on a citation count nobody outside those companies can verify.
More on content
- llms.txt
The summary file for language models, and what ours contains.
- For AI search
Be the source the answer cites.
- For AI Overviews
Publish the page the block can summarise.
Questions about GEO.
Including whether it is different from SEO, whether llms.txt matters, and how anyone claims to measure it.
How is GEO different from SEO?
It is mostly the same work seen from a different end. Search engines rank a page and a person clicks it; answer engines read a page and may quote it. The things that make a page readable — one answer, stated early, with specifics and machine-readable metadata — are the things that made it rankable for a long-tail question in the first place.
Do I need to write differently for AI?
No, and writing for two audiences at once is how content becomes unreadable. Write for the person who typed the question, put the answer near the top, and be specific. A model summarising your page for a buyer is reading the same sentences the buyer would.
Does structured data get me quoted?
It removes ambiguity, which is not the same as earning a mention. Markup tells an engine what the page is, who published it and when; whether the page is worth citing is decided by its content and by the engine. No markup and no supplier can promise a citation.
Is llms.txt part of this?
It is one layer of it: a plain-text index at the root of the domain, generated from what is published. It is a proposal rather than a standard, so it is treated here as a convenience on top of the sitemap, canonical URLs, structured data and crawlable HTML — not as a substitute for them.
How do I measure whether GEO is working?
Honestly: indirectly. There is no dashboard of AI citations to read, so the usable signals are Search Console's impressions for long, specific queries, a rising share of direct and branded traffic, and sales conversations that arrive further along than they used to. Anyone offering a precise citation-count metric is measuring something else.
Is GEO a new discipline, or an old one renamed?
Renamed, mostly, and that is the useful part. One answer per page, stated early, with specifics and machine-readable metadata, is what long-tail SEO already asked for; what changed is who reads the result. A vendor selling GEO as a separate module is charging for the acronym.
Should we publish more, to be quoted more?
Volume without a plan produces near-duplicates that compete with each other and add nothing to cite. The pipeline publishes a post per publishing day with one query behind each, which is a slower and more defensible answer than flooding a topic with variations.
Who owns GEO — marketing or engineering?
Both, and the split is real: choosing one question per page and committing the outline is editorial, while the canonical URL, structured data, dates and sitemap are engineering. A pipeline that treats them as one workflow is why the pages come out structured; a GEO module bolted onto a writer's desk is why most do not.
One answer per page
Publish the pages worth quoting.
Scan a domain and read the plan: the questions it does not answer yet, the outlines behind them, and the dates each post goes live on.
Structured data on every post20€ / site / language / month3 days free, no card