Generative Engine Optimisation: Getting Cited by AI Search

Your best page may be invisible to the tool your customer is using

A buyer who once typed a query into Google and scanned ten links now often asks an assistant and reads one paragraph. If that paragraph cites a competitor, you were not in the running. There was no page two to be on.

That shift is what generative engine optimisation is about. The name is newer than the practice, and a fair amount of what is being sold under it is ordinary SEO with a fresh label. But there is a real difference in emphasis, and it is worth understanding before you pay anyone to do it.

How AI search tools actually pick sources

Most AI search products work in two stages. First they run a search, either against a conventional web index or their own, and pull back a set of candidate pages. Then a language model reads chunks of those pages and writes an answer, citing the ones it drew from.

So there are two gates. You have to be retrieved, which still depends largely on classic ranking signals: relevance, links, crawlability, reputation. Then you have to be quotable, which depends on whether the model can find a clean, self-contained statement on your page that answers the question.

1. Retrieval. The page must be indexed, fast, readable without heavy client-side rendering and relevant to the query wording.
2. Chunk selection. The system splits pages into passages. A passage that answers the question on its own beats one that needs three paragraphs of context.
3. Trust. Where sources disagree, models lean towards those that look authoritative and are corroborated elsewhere.
4. Citation. The answer cites the passages it used. Being retrieved but not quoted earns you nothing.

What to change on your pages

The most effective change we make is boring. Put a direct, two or three sentence answer immediately under the heading that asks the question. Then elaborate. Models extract that opening passage far more often than a conclusion buried at the end.

• Question-shaped headings. Write H2s the way people ask: How much does a commercial boiler service cost in Leeds? rather than Our Pricing.
• Specific, checkable facts. Price ranges, lead times, service areas, minimum order quantities. Vague copy gives a model nothing to quote.
• Tables for comparisons. Structured comparisons survive chunking well and are easy to lift accurately.
• Named authors and dates. A page with a visible author, a real updated date and a clear business behind it is easier to trust.
• Structured data. Organisation, Product, FAQ and LocalBusiness markup does not guarantee citation, but it removes ambiguity about who you are.
• Server-rendered content. If your key text only appears after JavaScript runs, some crawlers will not see it at all.

Our technical SEO checklist covers the crawl and rendering side in more detail. None of it is optional for GEO; it is the first gate.

GEO versus SEO: where they differ

The overlap is large. If someone proposes a GEO programme that ignores your existing search performance, be sceptical. A page that is not retrieved cannot be cited.

The off-site part most businesses skip

Language models form a picture of your business from everything they can find, not only your site. If your Google Business Profile says you close at five, your website says six and a directory says you moved premises in 2023, the model has to pick one, and it may pick the wrong one or leave you out.

Tidy the consistent facts first: name, address, phone, opening hours, service list, prices where you publish them. Then look at where your category gets discussed. Trade directories, industry comparison pages, relevant forums and review platforms all feed the picture. We cover the recommendation side of this in how AI assistants choose which businesses to recommend.

How to measure it without fooling yourself

This is the weakest part of GEO today. There is no equivalent of Search Console for most AI assistants, answers vary between runs, and they differ by user location and history.

What we do is pragmatic. Pick 20 to 40 real questions your customers ask. Run them through the main assistants on a schedule, record whether you are mentioned or cited, and track the rate over months. A handful of tools now automate this sampling; they are useful, but treat their scores as trend lines, not truth.

• Referral traffic from AI assistants in your analytics, which is small but growing and easy to segment
• Branded search volume, which tends to rise when you are mentioned in answers even without a click
• Enquiries where the customer says they found you through an assistant; add it as an option on your form

When GEO is not worth your money yet

If your site has thin content, no organic traffic and technical faults, fix those first. GEO sits on top of fundamentals and cannot replace them.

It is also a weak bet for businesses whose customers do not research before buying. An emergency locksmith gets far more from the map pack and phone calls than from being quoted in a long assistant answer. And anyone promising guaranteed citations is selling something they cannot control. At SpiderHunts we treat GEO as an extension of search work we would do anyway, and we say so when a client would be better off spending the budget elsewhere.

One useful side effect: pages restructured into clear, self-contained answers also make excellent source material for an assistant on your own site. If that is on your list, the same content work feeds straight into AI chatbot development, so it pays twice.

Frequently asked questions