Generative Engine Optimisation
GEO, described honestly
Being retrieved and cited by generative systems is a genuine and consequential problem. It is also the part of this field where the least is publicly known — so this page is as much about the limits of the claim as the claim itself.
No AI provider publishes a ranking algorithm for generative answers. Anyone who tells you they know how to guarantee a mention in one is guessing, selling, or both.
That does not make the problem unaddressable. It means the honest approach is to work on the conditions that are necessary under any plausible mechanism, measure what can actually be observed, and label the difference between the two. That is what this page describes and what our audit does.
What a generative engine is doing
When a generative system answers a commercial question, several distinct things happen in sequence, and GEO is concerned with all of them. The system interprets the question. It decides whether it needs to retrieve anything, or whether what it already holds from training is sufficient. If it retrieves, it issues its own queries — not your customer’s words — against an index or a search product. It selects a handful of documents from what comes back. It synthesises a short answer from those documents plus its own internal representation of the subject. Then, depending on the product, it may attach citations to some of what it said.
Each step is a separate place you can be excluded. You can be missing from the index. You can be in the index but lose the selection step to a directory page that covers twenty businesses including yours. You can be selected but contribute nothing quotable, so the answer paraphrases a competitor’s page instead. You can be quoted but not cited. And you can be described from the model’s internal representation — which may be months out of date — without any retrieval happening at all.
This is why GEO cannot be reduced to a checklist of on-page changes. A page that is perfectly structured still has to survive a retrieval step it cannot observe, against competitors it did not choose, in a product whose behaviour changes between releases.
What we can say with confidence
A short list, and deliberately short. Generative products demonstrably read structured data, follow crawl directives, and retrieve from indexes that respect conventional discoverability. They demonstrably favour sources that state facts plainly over sources that imply them. They demonstrably rely on independent corroboration when resolving who an entity is — which is why a business with three authoritative profiles is described more confidently than one with none. And they demonstrably cite at different rates depending on the product, with some citing almost everything and others citing nothing.
What nobody can currently say
Which of those factors matters most, and by how much. Whether a given change will produce a mention. How long a change takes to propagate into a model’s behaviour, as opposed to its retrieval. Whether being cited more often in one product has any effect in another. Anyone presenting numbers against those questions is presenting a hypothesis dressed as a finding.
Boundaries
SEO, AEO and GEO — what each one owns
The three overlap enough that they are often conflated. They answer different questions and they fail in different ways.
| SEO | AEO | GEO | |
|---|---|---|---|
| Goal | Rank in a list of results | Be quotable as an answer | Be retrieved and cited by a generative system |
| Scope | Your site plus inbound links | Almost entirely your own pages | Your site plus the whole information environment about you |
| Under your control | Largely | Almost completely | Partially at best |
| Measurable today | Yes — positions and impressions | Yes — every signal is directly testable | Only by repeated observation, not by mechanism |
| Typical failure | You rank on page three | Your page has no extractable answer | A directory or competitor is retrieved instead of you |
| Feedback speed | Weeks | Immediate — re-run the audit | Unpredictable, and different per product |
Where to spend effort
Five things that plausibly influence retrieval
Chosen because they are necessary under every plausible mechanism, measurable, and useful even if generative search stopped growing tomorrow.
- 1
Entity resolvability
A system must be able to decide, without ambiguity, that the business on your site is the business the question is about. Organization schema, a consistent legal name, a stable address and sameAs links to profiles that say the same things are what make that decision cheap. Ambiguity is resolved by skipping you.
- 2
Independent corroboration
Claims that appear only on your own site are weak evidence. Claims repeated by directories, registries, industry bodies, publications and review platforms are corroborated. This is the substance behind the Citations & Authority component, and it is the single largest thing most businesses neglect.
- 3
Specificity over breadth
Generative retrieval is triggered by narrow questions. A page that answers one specific question thoroughly is more retrievable than a page that mentions twelve services in passing. This is the opposite of the instinct to consolidate everything onto a services page.
- 4
Extractable structure
Answer-first paragraphs, question-shaped headings, FAQ markup, short self-contained statements. If the useful sentence in your page depends on three paragraphs of context, it will not survive extraction.
- 5
Freshness that is visible
Dated content, updated pages and declared modification times help a system prefer your current position over its own stale internal representation of you. An out-of-date description is often an artefact of nothing newer being available.
What we deliberately do not do
- Promise a mention, a placement or a citation.
- Report a “GEO rank” as though a ranked list existed.
- Attribute an observed change to a specific action without evidence.
- Present an estimate as an observation.
In our assessment
Where GEO shows up in the score
There is no single GEO number, because there is no single GEO mechanism. It is distributed across the components that can actually be evidenced.
Citations & Authority
15% of the scoreWhether answers cite your site and whether credible independent sources exist about you. The closest thing to a direct GEO measurement, and observed rather than estimated wherever a citing platform is connected.
AEO / Website Readiness
10% of the scoreThe retrievability and extractability conditions on your own side. Fully measurable, which is why it is audited rather than inferred.
Brand Accuracy
15% of the scoreWhether what is generated about you is true. A proxy for whether the sources being drawn on are yours and current, or third-party and stale.
Competitive Share
10% of the scoreWho is retrieved instead of you. Marked as not estimable — it requires observed answers, and is reported as no data rather than guessed.
Estimated component scores are derived from measured signals on your own website and public entity signals — not from live AI platform answers. Connect AI platform integrations to replace estimates with observed data.
Questions
Questions about GEO
Both, at the moment. The underlying problem is real: generative systems retrieve and cite a small number of sources, and which sources they choose has commercial consequences. The term itself is young and is frequently used to sell certainty that does not exist. Treat any GEO service that promises placement in AI answers with the same scepticism you would apply to a guaranteed number-one ranking.
Not reliably, and not in a way that stays true. The major providers publish no ranking algorithm for generative answers, retrieval behaviour differs between products, and the same product changes between model versions. What can be done honestly is observation — running a defined set of questions repeatedly, recording which sources are cited, and looking for patterns — while being explicit that patterns are not mechanisms.
Work on the things that are necessary under any plausible mechanism: be retrievable, be unambiguous about who you are, state your facts in machine-readable form, be corroborated by independent sources, and write content that answers specific questions directly. None of these depend on a particular model's selection logic. All of them are measurable, which is why they form the basis of our audit.
AEO is almost entirely about your own pages: can a machine read this page and lift a clean answer out of it. GEO is about the wider information environment: does the system reach for your page at all, does it cite you, and what do other sources say about you that shapes the answer. You can complete an AEO programme alone. You cannot complete a GEO programme alone, because it depends on sources you do not control.
No. Several generative products retrieve from a conventional search index before they generate, so conventional discoverability remains an input rather than a legacy concern. The more accurate framing is that SEO became one of several inputs to a process that also weighs entity clarity, corroboration and answerability.
Published retrieval documentation from the platforms themselves, or a large enough body of repeated, controlled observations to distinguish a causal factor from a correlated one. Until one of those exists, the responsible position is to measure what we can verify, label estimates as estimates, and avoid attaching confident mechanisms to observed outcomes.
Start with what can be measured
The free check covers every signal on your side of the problem, with the evidence shown.