Generative engine optimization

Generative engine optimisation is the work of making your content the source an AI assistant reaches for when it answers a question in your category. Not ranking a link that someone might click - being named, quoted and cited inside the answer itself, which is increasingly where the decision gets made.

The surfaces in scope are the assistants people ask directly - ChatGPT, Perplexity, Claude, Copilot, Gemini - plus the AI-generated answers now sitting on top of ordinary search results. They behave differently enough that a tactic which works on one may do nothing on another, which is why this page describes workstreams rather than tricks.

Two things happened at once. Assistants became a normal place to ask a buying question, and search results started answering questions without sending anyone anywhere. An Ahrefs study published in December 2025 found that zero-click searches rose from about 56% to 69% in a year - meaning roughly two in three searches now end without a click to any website at all.

That number is the whole argument. If most of the demand in your category resolves inside an answer box or a chat window, then the only positions worth holding are the ones inside those answers. Ranking third for a query that never produces a click is a rosette for a race nobody finished.

The phrase itself came out of academic work rather than an agency pitch deck, which is mildly reassuring, and it is about two years old, which is not. Anyone presenting a settled playbook is extrapolating from a handful of months, and you should read them - and this page - accordingly.

The overlap is bigger than the GEO-is-a-whole-new-discipline crowd admits, and the divergence is bigger than the it's-just-SEO crowd admits. Both camps are selling simplicity you can't afford.

What's shared: crawlability, server-rendered content, sane architecture, genuine usefulness and site health. Google's own line is that preparing for AI surfaces is largely still ordinary SEO, and on the technical layer that's true. If your fundamentals are broken, GEO is not a workaround - it's the same wall, painted differently.

This is the whole job. Everything sold as a GEO service is some subset of these five, and they run roughly in this order, because each is worthless if the one above it fails.

1 · Technical retrievability

Can a machine fetch your content and read it without executing JavaScript? Client-rendered pages are the most common silent failure on the modern web: the site looks perfect to you and is empty to a crawler. Check the raw HTML, check your robots.txt for the AI user agents, check your pages carry visible dates. An afternoon of work, and it gates everything else.

2 · Entity clarity

Can the machine tell what you are, who you're for, and what category you belong to - from plain declarative sentences rather than inference? "We help teams move faster" is unusable. "Invoice tracking for freelance designers who bill hourly" is a fact a model can retrieve and match to a question. Structured data helps here for the same reason: it removes ambiguity rather than adding magic.

3 · Quotable formats

Models extract passages, not pages. What gets lifted is direct answers placed immediately under the question, comparisons, and original numbers that exist nowhere else. A paragraph that only makes sense in the context of the three before it is nearly impossible to quote; one that carries its own claim and support drops into an answer intact. The citation mechanics are broken down here.

4 · Off-site presence

The hardest workstream and the one most people skip, because it doesn't feel like optimisation. Assistants synthesise how the web describes your category, so being genuinely mentioned in the roundups, forum threads and comparisons that already get cited moves you more than any on-page change. If nobody independent has written about you, there is nothing to retrieve, and no amount of schema markup invents it.

5 · Measurement

Fix a set of twenty to forty prompts a real buyer would type, run them on a schedule across the assistants your buyers actually use, and record who gets named rather than just whether you did. Competitor share of voice is the more useful number, and the prompt set has to stay frozen between runs or you lose comparability, which is the only thing that makes the exercise worth doing at all.

• Anything researched before it's bought. Software, tools, services, anything with a "best X for Y" query attached to it. The assistant is now doing the shortlisting that a comparison article used to do.
• New sites with no authority. The genuinely good news: retrieval weights relevance and extractability over domain age, so a six-month-old site can get cited next to a household name - an opening classic search hasn't offered in a decade.
• Anyone whose informational traffic is already sliding. If your how-to content is losing clicks while impressions hold, you are watching the answer get given without you. See the AI Overviews page for how to confirm that diagnosis.
• Not everyone. If you sell locally, by referral, or through a channel with no research step, this is a low priority. Fix your actual bottleneck first.

Every honest practitioner here is working from behavioural observation, not documentation. The list of open questions is longer than the list of answers, and pretending otherwise is how people end up paying for cargo cult tactics.

• How sources are weighted. No assistant publishes its selection criteria. We can see what gets cited; we cannot see why one source was chosen over an equally good one.
• Whether on-page changes cause citation gains. Correlation is all anyone has. The systems update without notice, so a gain after a change may be the change, or may be a model refresh that shipped the same week.
• Whether llms.txt matters. The proposal exists, publishing one is cheap, and no major assistant has committed to consuming it as an input. Treat it as insurance, not strategy.
• How durable any of this is. Tactics tuned to 2026 retrieval behaviour may be irrelevant by 2027. The workstreams above are stable only because they reduce to "be findable, be clear, be worth quoting" - the tactics underneath them are not.
• What the traffic is worth. Assistant referrals are low in volume and high in intent for most sites. Anyone quoting a confident revenue multiple is guessing with a straight face.

Establish a baseline before you change anything, or you'll never know whether the work did anything. Write your prompt set, run it, record who gets named. Then fix retrievability, because it's cheapest and it gates the rest. Then publish one thing only you could publish - your own numbers.

That last one is the most underrated move available to an operating business. Every company sits on data nobody else has, and a specific number attached to a date is infinitely more quotable than a well-argued opinion. Ours, published because it makes the point: across four sites launched in 2026, first-weeks impressions ran between 2,200 and 16,500, click-through rates 0.1% to 0.6%, average positions 12 to 49. Unglamorous, but real, dated, and ours.

The GEO Audit runs your category's real buying prompts across the major assistants, records who gets named instead of you, and returns the specific fixes - with a 90-day recheck so you find out whether they worked.

Or get the 30-day post-build playbook Send it