FAQ Content and AI Answers

A format that lost its reward and gained a rumour

FAQ content occupies an unusual position in 2026. The visible commercial reason for publishing it in a marked-up form was withdrawn by Google on 7 May 2026, when FAQ rich results stopped appearing in Google Search. In the same period, question-and-answer formatting became one of the most commonly recommended practices for improving how content appears in AI assistants. Those two movements point in opposite directions, and the second is far less evidenced than the first.

This entry documents what can be checked. The retirement of the rich result is a matter of platform record. The proposition that AI answer systems preferentially lift text arranged as questions and answers is not a matter of record, and it is stated here as an open question rather than resolved in either direction. AIOFacts treats the difference between a documented platform behaviour and a widely repeated practitioner belief as the most important distinction in this topic.

What actually changed at Google, and what did not

The change was gradual and is fully traceable in Google's own documentation. In August 2023, Google announced changes to HowTo and FAQ rich results, and from September 2023 the FAQ rich result was shown only for what Google described as well-known, authoritative government and health websites. In May 2026 Google added a deprecation notice stating the feature would no longer appear in Google Search from 7 May 2026, and in June 2026 the documentation for the feature was removed, along with the Search Console appearance filter, the rich result report, and Rich Results Test support. Reporting from Search Engine Journal covered the removal at the time.

Three things did not change, and they are frequently conflated with the things that did. FAQPage remains a valid type in the Schema.org vocabulary, defined there as a web page presenting one or more frequently asked questions. Google has previously stated that unused structured data does not cause problems for Search, so existing markup does not need to be stripped out. And the separate QAPage type, intended for pages where a single question receives multiple user-submitted answers, was not part of this deprecation. Losing a search result presentation is not the same as losing a data format.

How retrieval reaches a passage

The intuition behind FAQ optimization is that answer systems retrieve at a granularity smaller than the page, so text already cut into answer-sized units should be easier to lift. The first half of that intuition is publicly observable. Google's featured snippets documentation describes systems that determine whether a page would make a good featured snippet and elevate it, and it states that publishers cannot mark content up to opt in. The only documented controls run the other way: nosnippet, data-nosnippet, and max-snippet limit what can be shown.

Microsoft's AI Performance report in Bing Webmaster Tools, announced on 10 February 2026, exposes a mechanism that complicates the second half of the intuition. Among its metrics is Grounding Queries, described as showing the key phrases the AI used when retrieving content that was referenced in AI-generated answers. That is a meaningful detail: the retrieval query is generated by the system, not supplied verbatim by the person asking. Google's guidance separately describes query fan-out techniques used to develop responses across multiple subtopics.

The implication is narrower than it first appears. If the phrase used to retrieve is written by the system rather than typed by the user, then matching a literal question string on the page is a weaker lever than it seems, because the string being matched against is not the one you predicted. What survives is the more ordinary requirement that the passage be self-contained and unambiguous.

The platforms do not agree, and the disagreement is the finding

Google's guide to optimizing for generative AI features is direct: structured data is not required for generative AI search, there is no special Schema.org markup that needs to be added, and there is no requirement to break content into tiny pieces for AI to better understand it. It goes further, stating that publishers do not need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search. Its companion page on AI features repeats that there are no additional requirements to appear in AI Overviews or AI Mode.

Microsoft has said something different in emphasis. Search Engine Land reported in March 2025 that Fabrice Canel, a principal product manager at Bing, confirmed that schema markup helps the company's language models comprehend content, and advised using IndexNow to push updates so generative models have current references. Both statements come from people with access to their own systems and no access to each other's.

AIOFacts records this as an unresolved conflict rather than picking a winner. Platform behaviour varies, statements from platform staff are not always precise about which subsystem they describe, and neither party publishes the retrieval internals that would settle it. A practice recommended on the strength of one vendor's remark should be described that way.

What the published evidence supports

The most substantial peer-reviewed work in this area is GEO: Generative Engine Optimization, presented at KDD 2024 by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande. It built a benchmark of roughly 10,000 queries and tested nine content modifications against a simulated pipeline, measuring position-adjusted word count and a subjective impression score. The strongest results came from adding quotations, reported at 41 percent and 28 percent on the two metrics, followed by adding statistics at 33 percent and 22 percent, and citing sources at 28 percent and 14 percent. Keyword stuffing measured at minus 9 percent on the first metric, below baseline.

Notice what is absent from that list. The study did not find that formatting content as questions and answers improved visibility, because it did not test that. What it measured was the presence of attributable evidence inside a passage: a quotation, a figure, a citation. Those are properties of what a passage contains, not of the shape it is arranged in.

The study's limits are as important as its findings. It evaluated a constructed pipeline that retrieved sources and synthesized answers with a model generation now several cycles old, and its authors reported that efficacy varied across domains. It is evidence about how one class of system behaved under test conditions, not a description of how any deployed assistant works today.

Why the format was discounted once already

The 2023 restriction and the 2026 removal are worth reading as a single episode with a lesson attached. FAQ markup was cheap to produce, applied indiscriminately, and rewarded with additional space in search results regardless of whether the questions were ones anyone had asked. Reporting at the time of the 2023 change attributed Google's decision to the volume of low-value implementations rather than to any defect in the vocabulary itself.

The generalizable point is structural rather than specific to FAQs. A signal that is inexpensive to fabricate and visibly rewarded attracts fabrication until the reward is withdrawn. Anyone now bolting question-and-answer blocks onto pages in the hope that AI assistants will favour them is repeating the pattern that ended the previous version of this practice, and doing so on weaker evidence than existed the first time, because there was at least a documented rich result then.

What defensible FAQ content looks like

The case for writing genuine FAQ content does not depend on any claim about model internals. It rests on two ordinary observations: that some information is genuinely shaped as a question a real person asks, and that a passage which answers such a question without depending on the paragraphs around it is easier for a human, a search index, or a retrieval system to use in isolation. That is a modest argument, and it is the only one currently supportable.

The practices below follow from documented platform behaviour and from the evidence above. None of them are claimed to influence any specific system.

• Publish questions people actually ask, drawn from support tickets, sales calls, and site search logs, rather than questions reverse-engineered from keyword tools.
• Make each answer self-contained: resolve pronouns, name the entity rather than saying it, and state the scope so the passage remains accurate when read alone.
• Put the evidence inside the answer. The GEO study associated quotations, statistics, and cited sources with higher measured visibility, and each of these is checkable by a reader regardless of any system effect.
• Match the vocabulary type to the content. FAQPage is for questions the publisher answers editorially with one accepted answer; QAPage is for a single question with user-submitted alternative answers, and Google's documentation says not to use it for pages with multiple questions.
• Only mark up question and answer text that is visible on the page to the person reading it.
• Do not duplicate body content as an FAQ block. A question answered twice on one page adds no information and repeats the pattern that led to the format being discounted.
• If a question has no substantive answer, leave it out. A thin answer occupies the position a real one could have held.

The limits of the evidence

Almost everything asserted about FAQ content and AI answers describes an interior no external party can inspect. Retrieval, ranking, passage selection, and citation are performed by systems whose operators publish behaviour and controls, not mechanisms. Google documents what a publisher may prevent through nosnippet and max-snippet, and OpenAI documents that OAI-SearchBot surfaces sites in ChatGPT's search features while ChatGPT-User is not used to determine whether content may appear in search. Neither documents why one passage is selected over another.

This means correlation is the ceiling of the available evidence. Bing's AI Performance report gives publishers per-URL citation counts and the grounding phrases involved, which is the closest thing to a measurable feedback signal that any platform currently offers, and it covers Microsoft surfaces only. A page that is cited more after being restructured may have been cited more for any number of reasons, including changes in the query mix and changes to the systems themselves between measurements.

AIOFacts records FAQ formatting as a reasonable editorial practice with a plausible retrieval rationale and no confirmed platform endorsement outside Microsoft's general remark about schema. Treating it as an established requirement overstates what anyone outside these systems is in a position to know.