Generative engine optimization, and the claims we will not make

What generative engine optimization actually is

Generative engine optimization is the practice of making a source more likely to be retrieved, quoted and correctly attributed when an AI system composes an answer. The term comes from a 2023 research paper by Aggarwal and colleagues, later presented at KDD 2024, which built a benchmark of generative-engine answers and tested which changes to a source document improved its visibility in the response. Their headline result — that certain content changes, particularly adding citations, quotations and statistics, improved visibility by up to around forty per cent in their benchmark — is genuinely interesting and routinely misquoted. It was measured on a research harness, it varied substantially by domain, and it is not a claim about what ChatGPT will say about your company next Tuesday.

That gap between an interesting finding and a sales promise is most of what is wrong with the current GEO market. So it is worth being precise about the mechanism.

The retrieval layer is still a search index

Answer engines do not consult a private ranking of the web that vendors can influence. They run retrieval — often several searches — against an index, pull passages from the documents that come back, and generate prose over them. Google is unusually direct about the consequence: there are no additional requirements to appear in AI Overviews or AI Mode beyond being indexed and eligible to be shown with a snippet, and no special files, markup or markdown are needed. ChatGPT’s search results are served by its own search crawler working over pages it has been permitted to fetch. Perplexity’s bot exists to surface and link websites.

If your page cannot be crawled, rendered, indexed and snippeted, it is not a candidate for any of this. That is why the first phase of any engagement here is technical SEO rather than anything AI-specific — a fact that makes for a poor pitch deck and a much better outcome.

GEO, AEO and SEO: what is genuinely new

Three acronyms are competing to describe roughly the same work, and the distinction matters less than the vendors selling it would like. Answer engine optimization generally means being the source of a direct answer; generative engine optimization means being retrieved and cited inside a synthesised response. Both sit on top of search.

Stripped of the marketing, exactly three things are new:

Crawler access is now a business decision. Historically you allowed search crawlers because there was no reason not to. Now training access, search access and user-triggered fetching are separate permissions with different commercial consequences, and the default is whatever your platform shipped.

Extraction has replaced the click as the unit of value. A model lifts a passage. Whether that passage makes sense without the paragraph above it is now a content design question.

Measurement broke. Rank tracking does not describe a non-deterministic answer that varies by user and by run.

Everything else on the typical GEO deliverables list — good content, clear structure, technical health, real authority — is the organic search programme you should already have been running.

Establishing a baseline before touching anything

The most common failure in this work is the absence of a control. These systems return different answers to the same prompt on different days, so “we did GEO and now ChatGPT mentions us” is not evidence of anything unless somebody recorded what it said before.

A baseline means an agreed prompt panel — the questions your buyers actually ask, including the unflattering comparison ones — run repeatedly across the engines that matter to you, with results recorded rather than remembered. What comes out of it is usually more useful than the visibility score attached to it: you find out which competitors are consistently named, which third-party sources the models keep quoting when describing your category, and which specific factual errors about your business are circulating. That last category is often the highest-value finding in the whole engagement, because it is fixable and nobody knew about it.

Entity foundations: being an unambiguous thing

Answer engines are describing entities, not documents. If a model cannot resolve which organisation you are, it will either omit you or blend you with something else — and the second failure is worse, because it produces confident, wrong statements about your business.

Resolving the entity is unglamorous and largely within your control. One consistent legal and trading name. One consistent category description used in the same words on your own site, in your structured data and on the third-party profiles you control. Organization markup with a stable identifier and sameAs links pointing at profiles that genuinely exist and genuinely belong to you. Consistent location and contact details everywhere they appear, which is where this overlaps directly with local search work.

A word on Wikidata, which is regularly sold as a GEO deliverable. Its notability policy admits an item if it has a valid sitelink to a Wikimedia project, or refers to a clearly identifiable entity that can be described using serious and publicly available references, or fills a structural need. That second condition is a real bar, not a formality: an entry created for a business with no independent published coverage is a candidate for deletion, and creating one about yourself is exactly the pattern that attracts scrutiny. We will tell you whether you plausibly meet the criterion. We will not manufacture the sources that would make you appear to.

Content structured so a passage survives being lifted

Content written to be extracted looks different from content written to be read top to bottom, and the difference is mostly about self-containment.

The practical rules are dull and effective. Answer the question in the first two sentences of the section, before the context and the caveats. Make each section stand alone, so a passage pulled out of the middle of the page still says what it means — pronouns pointing at a heading two screens up do not survive retrieval. State definitions plainly and once. Put comparisons in an actual table rather than in prose, because a table is trivially parseable and a paragraph of hedged comparison is not. Attribute claims to a source the reader and the model can follow, which is why the citations block at the foot of this page exists.

There is also a quality floor underneath all of this that no amount of formatting substitutes for. Google’s guidance on generative AI features says the same thing its content guidance has said for years: commodity material restating what is already everywhere has nothing to be retrieved for. The pages that get cited tend to carry something first-hand — a method, a constraint, a number somebody actually measured, an opinion with a reason attached. That is a content problem more than a technical one.

Structured data and machine-readable facts

Worth stating clearly, because it is oversold: there is no schema type that makes a model cite you, and Google has said explicitly that no special structured data is needed for its AI features.

What structured data does is reduce ambiguity. Organisation markup that states your name, description, identifiers and verified profiles gives a consistent machine-readable version of facts that are otherwise scattered across a footer, an about page and a directory listing from four years ago. Breadcrumb markup states where a page sits. Article and service markup state what a page is. None of that is a ranking lever, and all of it makes you cheaper to describe correctly.

The corollary is that inconsistent facts are actively harmful. If your site says one thing, your profiles say another and your markup says a third, you have given a retrieval system three candidate answers and no way to choose. Pick one version of the truth and repeat it everywhere.

Crawler access: the lever most sites get backwards

This is the part of GEO that is genuinely new, genuinely consequential, and genuinely misunderstood — and it is a policy decision, not a technical one.

The agents are not interchangeable. OpenAI documents GPTBot as the crawler that gathers content for model training, and OAI-SearchBot as the one that powers search results inside ChatGPT; its documentation states that sites opted out of OAI-SearchBot will not appear in ChatGPT search answers, and that blocking GPTBot does not have that effect. Google-Extended controls whether crawled content is used to train and ground Gemini, and Google states it does not affect inclusion in Google Search or act as a ranking signal. Anthropic publishes an equivalent split between its training crawler, its search crawler and the fetches Claude makes when a user asks for a page. Perplexity distinguishes its indexing bot from user-initiated retrieval, and notes that user-initiated fetches generally do not follow robots.txt because a person asked for that specific page.

Two conclusions follow. First, the sentence “we blocked the AI bots” is not a description of a policy; it needs a list. We have seen sites remove themselves from assistant search results while intending only to opt out of model training, which is an expensive way to protect an asset nobody was going to pay for anyway. Second, blocking training crawlers does not remove your content from models already trained, nor from pages a user pastes in, so it should be chosen for the reasons it genuinely serves — licensing position, competitive concern, principle — and not on the assumption that it prevents your material being used.

We write this up as a decision record with a date and a rationale, and we verify against server logs that the agents you allowed are the agents actually arriving.

Third-party corroboration, and the tactics we refuse

Models describing a category tend to lean on sources that are not the vendor’s own website: industry publications, comparison sites, forums, standards bodies, established directories. Being present and accurately described in those places matters more here than it does in classic search, because a self-description carries less weight when several independent sources are available to synthesise.

That creates an obvious temptation, and a whole cottage industry servicing it: paid Wikipedia editing, listicles placed on freshly registered “best agencies” domains, reciprocal mention schemes, review-site seeding. These work briefly. They are also trivially recognisable, they violate the platforms’ own policies, and when they unravel the reputational cost lands on your brand and not on the agency that ran them. We do not do them, and we will say so early rather than let it become an awkward conversation later.

The legitimate version is slower: earning coverage by having something worth covering, appearing in the registers and directories your industry genuinely uses, publishing data other people want to cite, and being the source that a journalist or an analyst quotes. It is the same work that has always built authority, which is why it fits inside a wider marketing programme rather than sitting off to one side as an AI project.

Measuring AI visibility honestly

Official data exists and it is thin. Google’s Search Console now includes a generative AI performance report covering AI Overviews and AI Mode, but it reports impressions only — no clicks, no position — and it covers Web search results only, with rollout still incomplete. That is genuinely useful as a trend line and useless as an attribution model.

Everything else is inference. Referral traffic from assistants shows up in analytics where those platforms pass a referrer, and it is typically small in volume and disproportionately high in intent. Server logs show which AI agents are fetching which pages, which tells you about access rather than about citation. Commercial visibility trackers run prompt panels at scale and produce a share-of-citation number, which is directionally useful provided everyone remembers it is a sample of prompts chosen by a vendor, not a measurement of the market.

So we report three things separately and do not blend them into a single score: what the official impression data says, what the repeated prompt panel says, and what happened to branded search, direct traffic and qualified enquiries. That third group is the one that pays for the work. Building it properly is analytics work, and it is worth doing before the GEO budget rather than after it.

What we will not promise

No guarantee of citation, because nobody controls a model’s output. No submission to an AI index, because none exists for organic answers. No claim that a specific file, tag or format causes inclusion, because the platforms have said the opposite in writing. No before-and-after screenshot presented as proof, because a single prompt run proves nothing about a non-deterministic system. And no recommendation dressed up as established practice when it is really a plausible hypothesis — we would rather label it as one and let you decide whether to fund an experiment.

What we will say is that the foundations are unusually stable for a new field: be crawlable, be indexed, be an unambiguous entity, be genuinely useful, be quotable, be corroborated, and decide your crawler policy on purpose. If you want a candid read on whether that work is worth doing for your business right now, start a conversation and we will tell you what we would do in your position, including when the answer is nothing yet.