A growing share of searches now end without a click. Someone asks a question, an AI system assembles an answer from several sources, and the person acts on that answer. If your business is not among the sources it drew from, you are absent from the conversation entirely — not ranked tenth, simply not present.

Generative engine optimization is the work of becoming one of those sources. It overlaps with search engine optimization but is not the same thing: classical SEO competes for a position in a list of links, while GEO competes to be quoted inside an answer the user may never click through from. The signals overlap, the tactics differ, and the measurement is considerably harder.

How Generative Engines Choose Sources

AI Overviews, ChatGPT with browsing, Perplexity and Copilot do not work identically, but they share a pattern. A question is decomposed into sub-questions, each is used to retrieve candidate passages, and an answer is synthesised from the passages the system judges most relevant and trustworthy. Citations point at the documents those passages came from.

Three consequences follow, and they shape everything we do. First, retrieval happens at passage level rather than page level — a specific, self-contained paragraph can be cited from a page that would never rank first as a whole. Second, the system must be able to tell what your organisation is and what it does, which makes entity clarity more important than keyword density. Third, a claim that is stated plainly and attributably is far more quotable than the same claim buried in promotional prose.

There is no ranking to buy and no submission to make. What you can do is make your content easy to retrieve, unambiguous about who is saying it, and worth quoting.

What Our GEO Work Covers

Our Approach

The Technical Foundation

Crawler Access Is a Decision, Not a Default

AI crawlers use distinct user agents, and a robots.txt that only addresses the wildcard leaves your position on each of them unstated. That is a decision made by omission. We set explicit directives per agent so the choice is deliberate and reviewable — permitting engines you want citations from, and blocking any you would rather not train on your material. Blocking a crawler removes you from that engine's answers entirely, which is a real trade-off worth making consciously.

Rendering and Retrievability

Google renders JavaScript before indexing. Several AI crawlers do not, or do so inconsistently. A page whose content only exists after client-side rendering may be perfectly visible to Google and effectively empty to an AI crawler. We check what each fetcher actually receives rather than assuming parity, and this alone accounts for a large share of the cases where a well-optimised site is never cited.

llms.txt and Machine-Readable Summaries

The llms.txt convention gives assistants a single fetchable file stating what an organisation does and where the authoritative page for each topic is. Support is still emerging and no engine guarantees it will be read, so we treat it as cheap insurance rather than a primary lever — it costs little to publish and removes ambiguity for any system that does consult it.

Structured Data as an Entity Graph

Isolated blocks of schema on individual pages describe pages. A connected graph, where every page references one Organization and one WebSite by identifier, describes an organisation. The difference matters when an engine is deciding whether the Atisolve on your services page, your blog and your contact page are the same entity. We implement schema as a single graph for exactly that reason.

Content That Gets Quoted

Generative engines quote passages that answer something specific. That has practical consequences for how pages are written. A direct answer stated in the first sentence under a heading is far more extractable than the same point reached in the fourth paragraph. Concrete detail — a number, a constraint, a named trade-off — survives summarisation where general claims do not.

Question-formatted content earns citations disproportionately, because it matches the shape of what users ask. This is why we treat FAQ sections as substantive content rather than decoration: ten well-written answers on a service page are ten pre-formed passages an engine can lift, each addressing a real objection.

Marketing language works against you here. Superlatives and unsupported claims are exactly what a system tuned to avoid promotional content will discard. Writing that reads as informative rather than persuasive is more likely to be quoted — which is an uncomfortable but consistent finding, and one we design content around.

Measurement, Honestly

GEO measurement is immature and anyone claiming otherwise is overselling. There is no rank tracker for AI answers, results vary between users and sessions, and engines change their retrieval behaviour without notice. Two identical questions asked minutes apart can produce different sources.

What we can do is measure systematically rather than anecdotally: a fixed question set, run at intervals, against the same engines, recording whether you were cited, what was said, and whether it was accurate. Over months that produces a trend you can act on. We also watch referral traffic from AI platforms, which is small but growing and unusually well-qualified, since a visitor arriving after an assistant recommended you is already part-persuaded.

We will not promise a citation for a given query, because nobody controls that. What we will do is tell you where you currently stand, what is preventing you being cited, and whether the gap is technical, editorial or simply that a competitor has published something better.

Who This Suits

What You Receive

• Citation baseline — how the major engines currently answer your buying questions, whether you appear, and what they say about you.
• Technical remediation — entity schema, crawler directives, rendering fixes and llms.txt, implemented or specified for your team.
• Content restructuring plan — page by page, which question each should answer and how the answer needs to be stated.
• New content — produced where a genuine gap exists, written to be quotable rather than merely comprehensive.
• Ongoing measurement — the question set re-run on a schedule, with movement and accuracy tracked over time.
• Plain reporting — what changed, what it appears to have produced, and what remains uncertain.

Getting Started

The useful first step is a citation baseline. It is bounded, and it answers the question that actually matters: when someone asks an AI assistant the questions your buyers ask, what happens today? That result usually determines whether the work ahead is technical, editorial or competitive. Get in touch and we will run it against your buying questions.