Skip to content
← All posts
8 min read

Generative engine optimisation: making a SaaS site citable

Generative engine optimisation is the practice of writing a site so that an AI answer engine can quote it. It is not search engine optimisation with a new name, and the difference is not cosmetic: the two disciplines optimise for different mechanics and can pull a page in opposite directions.

This is what the checks look like in practice, why each one exists, and what nobody in this field can honestly promise you.

How is this different from SEO?

A search engine returns a ranked list of links. The competition is for position, and the user still clicks through to your page, where your copy gets to do its work.

An answer engine returns prose. It assembles that prose from sources, decides what to say about you before you get a word in, and may name you or may not. The competition is not for position. It is for being the text the model reaches for when it needs a sentence about your category.

The practical consequence: a page can rank beautifully and be entirely unquotable. A hero that says "Ship faster. Sleep better." with a screenshot underneath, and a features grid of six icons, may convert human visitors at a respectable rate and will contribute nothing to an assistant asked what tools exist for your problem. There is no sentence in it that survives being lifted out of context.

The other consequence is that you never see the absence. A missing search result shows up in Search Console as an impression you did not get. A missing citation in somebody's private conversation with an assistant leaves no trace anywhere you can look.

What does a citability audit actually check?

Our own audit runs eight checks against a crawl of the site. Each is a yes or a no with a weight attached, and the score is the weighted pass rate rounded and clamped to 0 to 100. The checks are grouped into four questions.

Does the site say what the thing is? The opening of the landing page is searched for a sentence of the form "X is a …". Not a promise, not a benefit: a definition with a copula in it. If there is no such sentence, an answer engine has nothing it can quote as a definition, and it will either skip you or improvise one, which is worse. This is the heaviest single check on the list.

Are headings written as questions? Headings across the crawled pages are matched against question words in both English and French. Fewer than three across the whole site fails the check. The reasoning is mechanical: assistants match on the question a user asked, and a heading that is already that question with the answer underneath it is the easiest thing in the world to retrieve.

Is the question and answer section machine-readable? A separate check looks for FAQPage JSON-LD. Question-shaped headings are good; question-shaped headings that are also marked up as structured data mean no parser has to infer where a question ends and its answer begins.

Is the product a resolvable entity? The crawl collects every @type in every JSON-LD block, including inside @graph arrays, and requires at least two of Organization, SoftwareApplication, Product, FAQPage, HowTo, Article or WebSite. Without them an assistant can read your page and still not be able to say what kind of thing you are or who makes you. A lighter check asks whether the product's own name appears in the title tag or the H1, which is a surprisingly common omission on sites whose H1 is a slogan.

Is it compared to its alternatives? Two checks. One looks for pages whose URL or title contains a comparison word: "vs", "versus", "alternative", "comparison". One looks for whether any page mentions a competitor or the word "alternative" at all. The first matters because "what are the alternatives to X" is one of the most common shapes of question put to an assistant, and the pages that get quoted in that answer are comparison pages. The second matters because a site that never names what it is competing with reads as evasive to readers and models alike.

Is there anything substantial to quote? The last check counts pages of 400 words or more and wants at least three. Not because length is a virtue, but because a model extracting a claim needs a passage that survives extraction, and a 90 word feature blurb does not contain one.

What does a good definition sentence look like?

The single highest-value edit on most sites, so it is worth being concrete.

Bad:

The fastest way to turn your data into decisions.

This has no subject. It does not say what the product is, who it is for, or what category it belongs to. There is no clause here an assistant can pull into an answer, because pulled out of context it describes nothing. It also contains a superlative that nobody verified.

Good:

Meridian is an error monitoring tool for Python teams that groups exceptions by root cause instead of by stack trace.

Name, category, audience, distinguishing mechanism, in one sentence with a copula. Lift it out of the page and it still means something. That is the entire test.

Two mechanical details that matter more than they should. Put it near the top: our check only looks at the first 800 characters of the page's visible text, because a definition below the fold is a definition the crawler that fed the model may never have reached. And write it as prose, not as a headline. A headline is a slogan in a large font; the definition sentence can sit underneath it in normal body text and do the load-bearing work.

What does a good FAQ answer look like?

Same principle, applied to a different block.

Bad question and answer:

Is it secure? Security is our top priority. We use industry-standard encryption and follow best practices to keep your data safe.

Nobody typed "is it secure" into anything. And the answer contains no fact, so there is nothing to extract. A model summarising this says "the vendor states that security is a priority", which is worth less than nothing to a reader.

Good:

Where is my data stored, and who can read it? Data is stored in the European Union, in Frankfurt. It is encrypted at rest with AES-256 and in transit with TLS 1.3. Our support staff can read event metadata but not payloads, which are encrypted with a key held in your account.

Written as somebody would type it. The answer arrives in the first sentence and the specifics follow. Every clause is checkable, which is exactly what makes it quotable: a model reproducing it is repeating facts, and facts are what it is trying to assemble.

The rule that generalises: answer in the first two sentences, then elaborate. Do not build to the answer. An assistant reading your page may only take the first part.

Does any of this guarantee a citation?

No, and be suspicious of anyone who tells you otherwise.

This is a young discipline. The retrieval behaviour of the major answer engines is undocumented, changes without notice, and differs between them. Some of what is confidently written about GEO today is extrapolation from a handful of observations, and some of it is SEO advice with the letters changed.

What can be said with more confidence is the negative: a site with no definition sentence, no structured data, no question-shaped headings and nothing over 400 words is very unlikely to be quoted, because there is nothing in it to quote. Fixing that does not buy a citation. It removes the reasons you certainly will not get one.

The other honest point in its favour is that none of this work is wasted if the premise turns out to be overstated. A page that says plainly what the product is, answers the questions buyers actually ask, and compares itself fairly to the alternatives is a better page for humans too. That was true before answer engines existed. It is the rare optimisation with no downside if the channel never materialises.

If you want these eight checks run against your own site, with the findings ordered by impact, that is what our audits do.

See your own numbers first

Connect Stripe and your real revenue is on screen in five minutes, before an agent has said a word. Create your account, answer the questions, and your strategy gets written.