Google retired FAQ rich results in June 2026, but answer engines still quote support content constantly. What actually earns a citation, and the structural changes that make your docs quotable.
Also available in Dutch: Zo wordt je kennisbank geciteerd door ChatGPT en AI-zoekmachines
Support content has quietly become one of the most-quoted categories on the web. When someone asks an assistant whether a product supports SSO, or how long a return window runs, the answer is assembled from help-center articles — usually without the person ever loading the page. That changes what a good help center is optimised for, and most of the advice still circulating was written for a search feature that no longer exists.
Two things worth knowing before you rewrite anything. First, Google removed its FAQ rich result documentation in June 2026 and the feature no longer appears in search results; in its final months it was already restricted to well-known government and health sites. If your content plan was built around winning FAQ rich snippets, that plan expired. Second, the llms.txt proposal reached a version 2 in August 2026. It is a community proposal rather than an official standard — worth adding because it is cheap, not because it is decisive.
A model assembling an answer is looking for a passage it can lift with minimal risk: unambiguous, attributable, and complete enough to stand alone. That is a different target from a page engineered to hold a reader's attention. The marketing instinct — build tension, contextualise, reveal the answer at the end — is precisely the structure that gets a page skipped.
| Pattern | Quotable version | Why the difference matters |
|---|---|---|
| Opening line | “The return window is 30 days from delivery for unopened items.” instead of “Returns are something we take seriously at Acme.” | The first passage answers the query; the second requires the model to keep reading and risk assembling the answer from elsewhere. |
| Specifics | “Supported on plans from €49/month upward” instead of a checkmark in a feature grid. | Icons and grids carry meaning visually and lose it in extracted text. State the limit in a sentence somewhere on the page. |
| Scope | “This applies to EU orders only; US returns run 14 days.” instead of leaving the exception on a separate page. | An answer quoted without its caveat becomes wrong. Put the qualifier in the same paragraph as the claim. |
| Pronouns | “Acme's return window” instead of “our return window”. | Quoted out of context, “our” has no referent. Naming the subject makes the passage attributable. |
Markup no longer buys a Google FAQ result, but it is still how a machine tells a heading from a navigation label. Article or TechArticle with a genuine dateModified, breadcrumbs that mirror the visible hierarchy, and organisation markup that ties the content to a named entity all remain worth shipping. What has never worked is markup that disagrees with the page — describing a question in JSON-LD that no visible text answers is the fastest route to being ignored.
Referral analytics understate assistant traffic badly, because most of these answers end without a click. Three measurements that behave better: ask the assistants your top twenty support questions on a schedule and record whether you are cited and whether the answer is right; watch for the branded queries that follow an assistant answer, which is where the demand actually surfaces; and track deflection on the articles themselves. We wrote separately about why deflection and resolution are not the same number, which matters here because a cited article and a helpful article are not automatically the same article.
The same properties that make an article quotable by an external assistant make it answerable by your own. A support AI reading your help center hits exactly the problems above — buried answers, missing qualifiers, contradictory duplicates — which is why we treat this as one job rather than two. The mechanics of that are in how to write help-center articles an AI can answer from, and the knowledge-base side of the product is on the features page.
Google removed its FAQ rich result documentation in June 2026 and the feature no longer appears in Google Search results, so FAQPage markup no longer earns a rich result there. It remains useful for other search engines and for assistants that parse structured data, and existing markup can be left in place, but it should no longer be treated as a Google ranking tactic.
It is a community proposal, currently at version 2 as of August 2026, for a markdown file at the root of a site that points AI agents to clean summaries of key content. It is not an official standard and no engine guarantees to use it, but it is inexpensive to publish and several documentation platforms now generate one automatically, so it is reasonable to add without expecting it to be decisive on its own.
Publish one self-contained answer per URL, state it within the first hundred words, include the specific numbers and limits as plain sentences rather than icons or grids, keep any qualifying exception in the same paragraph as the claim, and make sure the page renders in HTML without requiring a login or client-side fetch. Citation follows passages that can be lifted and remain accurate without the surrounding page.
Referral analytics undercount it because most assistant answers end without a click. More reliable signals are running your top support questions against the major assistants on a schedule and recording whether you are cited and whether the answer is accurate, watching for branded search queries that spike after an assistant answer, and tracking self-service deflection on the articles themselves.
No, and maintaining two versions tends to produce contradictions that damage both. The structural changes that make a page quotable by an assistant, such as answering early, stating specifics plainly and keeping one canonical page per fact, are the same changes that make it useful to a human who landed from a search result and wants an answer immediately.
2026-08-09 · 9 min read
The biggest driver of your AI resolution rate is not the model — it is your content. A practical guide: one question per article, front-loaded answers, real values instead of screenshots, and how to test with your own tickets.
2026-08-20 · 9 min read
Hallucination is the number-one objection to support AI, and it is largely a retrieval and permissions problem rather than a model problem. The controls that actually reduce invented answers.
2026-08-10 · 10 min read
A proactive message is one the customer did not ask for. Which signals are worth acting on, what a message has to contain to be support rather than noise, and how to measure win-back without fooling yourself.
See how honestly-measured AI resolutions cut your support load — start on the Free plan, no credit card, no sales call to get started.