How to Write a Page an AI Agent Can Actually Cite

“How do we get AI to cite us” is the most-asked question in this field, and most of the answers in circulation are guesswork. I went through the documentation from Google, Microsoft, OpenAI and Perplexity and extracted only what is actually documented. The headline finding: nosnippet and max-snippet directly affect AI Overviews and AI Mode — Google states this explicitly. The second: FAQPage and HowTo are no longer in Google’s gallery of supported features. And the third: no vendor documents how a passage gets selected for quoting. Every number you have seen about “the length of a quoted passage” comes from tool marketing, not research.

What is actually within your control

The first thing to separate is officially documented controls from style advice somebody invented. This table draws that line explicitly: the middle column is what the vendor writes, not what someone inferred.

nosnippet and max-snippet: the only documented levers

This is the part that surprises people most, because most of us know these two directives from classic SEO and assume they only govern the grey text under a search result. On Google’s robots meta tag page, updated 24 March 2026, it is stated explicitly that nosnippet applies to all forms of search results — web search, Images, Discover, AI Overviews and AI Mode — and that it prevents the content from being used as a direct input for the generative features.

max-snippet carries parallel wording: it limits how much of the content may be used as a direct input for AI Overviews and AI Mode. Which means that if you once set max-snippet:50 across a set of pages to control how they display in results, you also capped, without knowing it, how much Google can draw from them for a generative answer. Worth checking what is currently set on your site.

Two caveats from the same document are worth knowing. First: data-nosnippet works only on the three tags span, div and section, and Google explicitly warns against toggling it via JavaScript. The second, and the more interesting one, is in the next section.

The structured data paradox

In Google’s optimization guide for generative search, on a page updated 10 July 2026, it says structured data is not required for generative AI search and that there is no special schema.org markup you need to add. That sounds like the end of the discussion, but a different document adds a layer.

On the robots meta tag page there is a carve-out for max-snippet: the limit does not apply in cases where a publisher has separately granted permission — and one of the routes named is in-page structured data. The same document also states that structured data declared inside a data-nosnippet element remains usable for search results.

FAQPage and HowTo have been deleted from the gallery

If the guide you are reading recommends adding FAQPage schema to “get into AI answers”, it is years out of date. In 2023 Google reduced the display of FAQ and HowTo results, and most of the industry still treats that as a “restriction”. The situation today is more final than that: the FAQPage documentation page now redirects to the documentation updates page, and the changelog states the documentation was removed after the feature stopped appearing in results. The same happened to HowTo.

In the gallery of supported features, updated 15 June 2026, neither appears at all. What does appear and is relevant for content: Article, Q&A, Breadcrumb, Organization, Profile page and Review snippet.

As for speakable, which shows up in plenty of advice as “the schema for the AI era”: its documentation, updated 10 December 2025, defines it as a beta feature intended for audio playback on Google Assistant-enabled devices, returning up to three articles. Nothing in that documentation connects it to AI Overviews or AI Mode.

Google and Microsoft disagree on the premise itself

Here is the most interesting gap between vendors, and it explains why so much contradictory advice is in circulation. Google, in that same July 2026 guide, writes that there is no requirement to break your content into tiny pieces for AI to understand it better, and that there is no ideal page length. That is an explicit rejection of the entire “write in short blocks” school.

Microsoft says the opposite. In a post from 8 October 2025 written by a principal product manager at Bing, it is described that assistants like Copilot break content down into smaller structured pieces in a process called parsing, and that those pieces are then assembled into answers. The post notes that with question-and-answer structures assistants can often lift the pair word for word, and recommends self-contained phrasing — sentences that make sense even when pulled out of context.

One caveat I want to state clearly: that post appears on the Microsoft Advertising blog, which is marketing content rather than technical documentation, even though it was written by a senior product person. The practical conclusion is not to transplant Microsoft’s model onto Google. If you write in clear, self-contained paragraphs, that is documented as helpful at Microsoft, neutral at Google, and good for human readers either way.

OpenAI and Perplexity: two controls and no control

OpenAI has two documented controls that most people do not know about. First: OAI-SearchBot is entirely separate from GPTBot, and you can allow one while disallowing the other. The documentation states that a site that has blocked OAI-SearchBot will not be shown in ChatGPT search answers, though it may still appear as a navigational link. A robots.txt change takes effect within roughly 24 hours. I broke down the distinction between the bot families in the 2026 bot map.

The second control: the publisher FAQ states that even for disallowed pages, OpenAI may surface the link and page title in ChatGPT Atlas, and that the way to prevent this is a noindex tag. As for how sources are selected in the first place, all that is documented is a single sentence: ranking uses multiple factors, and placement is not guaranteed. The factors are not listed.

At Perplexity the picture is different again. Its source labels document states that it rates the whole website a source comes from, not each individual page, and lists publication-level criteria: whether the site corrects its mistakes, whether it says who wrote each piece, and whether it keeps news separate from advertising and opinion. Its publishers program, announced 30 July 2024, is a commercial arrangement of revenue sharing and API access — it contains no technical signal a site can send.

What you can measure today

Two new reporting systems have entered the picture, and both are more limited than they sound.

1. Google Search Console — a generative AI performance report covering AI Overviews and AI Mode. It reports impressions only: no clicks, no CTR, no position. The documentation notes that if two results from the same site are shown, that counts as a single impression. There is also now a permission toggle in Settings, and excluding your site prevents it from being linked to within the generative features. Both capabilities are documented as rolling out gradually to a subset of sites — not to everyone.
2. Bing Webmaster Tools — an AI performance report that entered public preview on 10 February 2026. It exposes total citations, average cited pages, page-level citation activity, and grounding queries: the phrases the AI used when retrieving content. The documentation notes those queries are a sample. Coverage is Copilot, AI-generated summaries in Bing, and select partner integrations.
3. ChatGPT referrals — identifiable in analytics, because OpenAI appends a UTM parameter naming chatgpt.com as the source to the URLs it refers to.

The numbers you should not believe

I looked for a primary source for each of the following numbers and found none. They circulate in countless guides, and they all come from one place: marketing by tools selling “answer engine optimization”.

There is no source for the typical length of a quoted passage — not 25 tokens and not 19 words. There is no source for the share of verbatim quotes versus paraphrase. There is no source for the average number of citations in a single answer. The only thing I found with published methodology is a research paper on news source citing patterns in AI search systems, analysing more than 24,000 conversations and 65,000 responses containing over 366,000 citations, of which 9% reference news sources. Even from that I would not derive an “average citations per answer” and publish it as fact.

My conclusion is not defeatism but prioritisation: everything documented and within your control sits in the first three articles of this cluster — correct permissions, content that exists in the initial HTML, and a discovery layer built on what is actually read. Writing clearly in self-contained paragraphs is good advice on its own merits. It just is not a documented mechanism, and I would rather say that than invent a number.

Frequently asked questions