showing up in AI answers: a working definition of GEO

A growing share of buying decisions now start with a question typed into an assistant rather than a search box. "Who does penetration testing for small healthcare companies?" returns a short paragraph naming three or four firms. If you aren't one of them, you were never in the running, and nothing in your analytics will show you the loss.

Optimising for that is being called generative engine optimisation, or GEO. It overlaps with SEO but rewards different things, and the differences are where the work is right now.

How the answer gets assembled

Simplifying, but not by much, there are three sources feeding any given answer:

1. Training data: what the model absorbed before it shipped. You can't edit this, and it lags reality by months to years.
2. Retrieval: live search results the assistant pulls in while answering. This is the fast-moving part, and it's the one you can influence this quarter.
3. The model's own summarisation: how it compresses those sources into a paragraph. This strongly favours content that already reads like a clear, factual answer.

The practical consequence: being retrievable and being quotable are two separate jobs, and most sites fail the second one even when they pass the first.

The five things that move it

1. Be crawlable by the AI crawlers specifically

They are not Googlebot and they don't all respect the same rules. Check your robots.txt for GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended. A default deny in a security config or a CDN bot rule blocks them silently, so the only way to know is to go and look.

Also: much of this retrieval doesn't execute JavaScript reliably. If your content only exists after hydration, assume it isn't being read.

2. Answer the question in the first two sentences

Models quote passages that stand alone. A page that opens with "In today's fast-paced digital landscape…" gives them nothing liftable. A page that opens with "A penetration test for a single web application with two user roles typically takes three to five days and costs between X and Y" can be quoted as it stands, and gets cited.

On most sites this is the biggest single improvement available, and the fix is a writing job.

3. Publish specifics you're willing to commit to

Numbers, ranges, timelines, named trade-offs, explicit "don't use this if…" sections. Vague marketing copy is unquotable by construction: there's no proposition in it to extract. Specificity is what makes a page a source rather than a brochure.

4. Structure it so a machine can parse the claim

Proper heading hierarchy, real <table> markup for comparisons, FAQ schema where you're answering real questions, and Organization / Service schema describing what you do and where. None of this is new advice. It just matters more when the reader is a parser.

5. Get cited somewhere that isn't your own site

Assistants weight third-party corroboration heavily: directories, comparison articles, forums, industry press, and community answers. A claim about you on your own domain is weak evidence. The same claim on three independent sites becomes the consensus the model reproduces.

Measuring it

You can't use rank tracking, because there's no ranking. Track prompt coverage instead: take the twenty or thirty questions a buyer would plausibly ask, run them across the major assistants on a schedule, and record whether you're named, in what position, in what sentiment, and which sources the answer cites. Do the same for three named competitors.

That gives you a baseline you can move deliberately, and it shows you where the failure is. Usually the assistants are citing a directory listing or a four-year-old comparison post, and both of those are fixable.

What not to bother with

• Stuffing prompts or instructions into your page copy. It doesn't work, and it reads badly to humans.
• llms.txt. An interesting proposal with minimal adoption. Publishing it costs nothing, but don't count it as work done.
• Mass-generating pages. Thin content is what these systems are best at discounting.
• Chasing every new assistant. Cover the ones your buyers use, because the mechanics transfer.

The short version

Be readable by the crawlers, answer the question immediately, publish specifics, mark it up properly, and get corroborated off-site. It's mostly the same discipline as good SEO, applied to a reader that summarises instead of clicking.

We run this as part of our marketing practice: baseline, fixes, and monthly tracking against named competitors. Ask for a baseline and we'll tell you where you stand today.