Getting found by AI assistants: structure, sources and citations

Every conversation about generative engine optimisation should start with a boring question, because the answer is often no. Can the assistant get the page?

For Google, the bar is the one that already exists. Its guidance on generative AI features says a page "must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements," and that "there are no additional technical requirements." The same documentation adds that you do not need to create new machine-readable files, AI text files, or markup to appear in these features. That single sentence disposes of most of what is currently being sold as GEO tooling.

For the assistants that crawl separately, access is a specific decision you have probably made by accident. OpenAI documents its crawlers individually: OAI-SearchBot is "used to surface websites in search results in ChatGPT's search features", GPTBot is "used to make our generative AI foundation models more useful and safe" through training crawls, and ChatGPT-User handles "certain user actions in ChatGPT and Custom GPTs" — a live fetch when someone asks for your page, not automatic crawling. Three different jobs behind three different tokens. A robots.txt rule that blocks the training crawler and leaves the search crawler alone is a coherent position; blocking all of them and then asking why ChatGPT does not mention you is not.

The third failure is rendering. Snippet controls apply here too — Google lists nosnippet, data-nosnippet, max-snippet and noindex as the levers that limit what can appear — and a nosnippet inherited from an old template is an unintentional opt-out. More common on Shopify: the fact a buyer needs lives inside a script-injected app widget, a tab that renders on click, or an image of a specification table. Retrieval sees HTML. If the answer is not in the HTML, the answer does not exist.

This is the part no amount of structure fixes. A summariser handed six pages that say the same thing does not weigh them on quality; it cites what it retrieved. If your page is the seventh restatement of an industry statistic, there is no version of it good enough to be chosen reliably, because nothing about it is yours.

Google's own advice points the same way: "creating content that people find unique, compelling, and useful will likely influence your website's presence in generative AI search," with the emphasis on firsthand work rather than recycled summaries. For a retailer, firsthand is narrower and more available than it sounds.

You own your policies — delivery cut-offs, returns windows, warranty terms, what happens when something arrives damaged. You own your specifications, including the awkward ones: what this fits, what it does not fit, what the actual measured weight is rather than the manufacturer's. You own comparisons between the things you sell, including the ones that talk someone into the cheaper option. And you own whatever your team knows from doing the work — the sizing quirk, the failure mode, the reason the expensive one is only worth it in one specific case. That last category is the one that gets cited, because nobody else can write it.

What you do not own is the general question. A page called "what is merino wool" competes with an encyclopedia. A page called "which merino weight for UK winter commuting, and when cotton is the better buy" has one plausible source, and it is you.

The pages that earn citations are rarely product pages. They are selection guides, compatibility pages, comparison pages, policy pages written as answers rather than legal text, and the occasional genuinely opinionated piece. A catalogue does not generate any of them.

Which turns an editorial strategy into a throughput problem, and this is where most programmes stall. If publishing a comparison page means a developer ticket, a two-week queue and a deploy, the plan that called for forty of them produces six. The fix is the same one that fixes campaign landing pages: a section library — a set of composable, tested blocks a merchandiser arranges without touching code. Comparison table, spec block, FAQ, step list, citation-friendly answer block. Build the components once, then the constraint is writing, which is the constraint you actually wanted.

It also keeps the structure consistent. A hand-built page gets its FAQ markup right; the fortieth one does not. A block that emits correct markup every time removes an entire class of quiet failure — and while structured data buys you nothing directly in AI features, Google is explicit that it should still match the visible text on the page, which is exactly the rule hand-rolled pages break.

Google publishes a Generative AI performance report in Search Console for content appearing in its generative AI features on Search and Discover. That is a first-party number, and it is the one to start from.

Everything else is estimation. The tools that promise to tell you how often ChatGPT mentions your brand work by running prompts repeatedly and counting the results, which measures a sample of a non-deterministic system — useful as a directional signal, worthless as a metric with a target attached. Assistants answer differently to different users, in different sessions, with different retrieval. Reporting a percentage from that to a board is inventing a number, and inventing numbers is how this discipline earns its current reputation.

The measurable proxies are the ordinary ones. Are the pages indexed. Are the crawlers allowed. Do the answer passages appear in featured snippets, which draw on the same extraction behaviour. Is referral traffic from assistant domains rising. Do the questions arriving at customer service look like the ones your pages answer.

Three things get pitched constantly and none of them survives contact with the documentation.

A file that declares your content to AI systems. Google states you do not need to create new machine-readable files or AI text files to appear in its AI features. Adding one costs an hour and does nothing; the harm is that it feels like the work, and the actual work is harder.

Schema added specifically for AI. Google says structured data is not required for generative AI search and there is no special markup to add. Keep your Product and FAQ markup because it earns rich results in ordinary search — that is a good enough reason — but no property in it is an AI lever.

Content produced at volume to blanket a topic. The failure is mechanical: more restatements of what is already on the web make you more replaceable, not less. Forty pages nobody would cite is worse than four somebody would, because the four are now harder to find inside your own site.

The honest version of generative engine optimisation is unglamorous. Let the crawlers in, put the facts in the HTML, write the passages so they survive being lifted, and spend the budget on the handful of pages where you are genuinely the source. There is no shortcut in the documentation because the vendors have said, repeatedly and in writing, that there is not one.

• Shopify section library
• Answer engine optimisation for a product catalogue
• What AI assistants say about your brand, and why