GEO: how to get your product cited by ChatGPT, Claude and Perplexity
Assistants answer questions by lifting passages, not by ranking pages. That changes what a page has to be: a self-contained, verifiable, quotable answer under a heading that matches the question. Eight practices follow from that — including the unglamorous ones like giving every page a real URL, prerendering it, and writing the number instead of the adjective.
Somewhere in the last two years a meaningful share of the “which tool should we use for X” research moved from a search results page to a conversation. That does not make search engine optimisation obsolete — the assistants are reading indexed pages — but it changes what a good page is. A ranking mechanism rewards pages. A generative one lifts passages.
Everything below is what this site does. That is deliberate: advice about findability that the advising site does not follow is one “view source” away from being embarrassing.
1. Write the passage that answers the question
A model answering “what is AI-native DevOps” needs a self-contained paragraph that defines it, under a heading that looks like the question. Not a paragraph that builds towards a definition over three sections, and not one that requires the sentence before it to make sense. Each answer should survive being lifted out of its page.
2. Put the number in instead of the adjective
“Highly configurable” is unquotable. “Nine statuses, sixteen transitions and eighteen conditions” is a fact with a shape, and a model repeating it can be checked. Specific, verifiable claims get cited more because they are the kind of thing worth citing.
The corollary is uncomfortable and worth accepting: do not publish numbers you cannot defend. A claim that turns out to be wrong does more damage in a generated answer than in a brochure, because it arrives without your branding attached to warn the reader.
3. Give every page a real URL, and prerender it
This is the least interesting and most frequently fatal item. A single-page application where every page lives behind the same URL is, to an index, one page. Every documentation article and every post needs its own address, and that address needs to return the full content as HTML without running JavaScript.
4. Answer questions in the words people use
People ask assistants full questions — “how much does an AI coding agent cost per user story”, “can an agent merge to main” — rather than typing keywords. Headings and FAQ blocks written in that voice match what is being asked. This is also the single highest-value overlap with paid search: the query you are bidding on and the heading on your landing page should be the same sentence.
5. Mark it up so a machine does not have to guess
Structured data is not a ranking trick; it is a translation of what your page already says into a form nothing has to infer. Article and TechArticle for content, HowTo for procedures, FAQPage for question blocks, BreadcrumbList so a page knows where it sits, Organization and SoftwareApplication so the entity is unambiguous. The rule that matters: the markup must state what the visible page states, or you are describing a page that does not exist.
6. Publish an llms.txt with the facts, not the pitch
A file at /llms.txt that summarises what your product is and links to the pages that matter is cheap to maintain and easy to read. It is worth writing as a fact sheet rather than as copy: what the product does, what it deliberately does not do, what is enforced, what is not built yet. The honest limitations are the part that makes the rest credible.
7. Be comparable on purpose
A large share of assistant queries are comparative: X versus Y, alternatives to Z. If you have not written the comparison, something else is answering it about you. Write it fairly — concede what the alternative is genuinely better at, and state where you do not compete — because a comparison that reads as advertising gets discounted by the model and by the reader.
8. Say what is not there yet
Counter-intuitive and consistently effective. A page with a section on what the product does not do yet reads as a source rather than a brochure, and it protects you from being cited for something you cannot deliver — which is a support ticket and a refund, not a lead.
A short checklist
- Every page has its own URL and returns complete HTML without JavaScript.
- Each page answers one question, in a heading phrased the way it is asked.
- Facts are specific, numeric where possible, and defensible.
- FAQ blocks are visible on the page and mirrored in structured data.
- A sitemap that lists every page, generated from the same list the navigation uses.
- An
llms.txtthat states the facts, including the limitations. - A comparison page written fairly, naming the alternatives.
- robots.txt allows the crawlers you want to be quoted by.
What is GEO (generative engine optimisation)?
Making your content usable as a source by AI assistants that answer questions directly — through self-contained, verifiable, quotable passages under question-shaped headings, marked up so nothing has to be inferred. It complements SEO rather than replacing it, because the assistants read the indexed web.
Does structured data help with AI assistants?
It helps by removing ambiguity: FAQPage, HowTo and Article markup state explicitly what a page already shows, so a machine does not have to infer the structure. It is not a ranking trick, and markup that describes something other than the visible page is worse than none.
Should I write a page comparing my product with competitors?
Yes — comparative queries are a large share of assistant use, and if you have not written it, other sources are answering it about you. Write it fairly, concede what the alternative does better, and say where you do not compete; a comparison that reads as advertising gets discounted by readers and models alike.