Playbook · 16 min read

The AI Search Schema Markup Playbook

Google retired FAQ rich results from Search on 7 May 2026, and a matched pairs test published the same month found that adding JSON-LD barely moved AI citations at all. Here is what the evidence actually supports, which schema types are still worth shipping, and the validation discipline we run on our own site before anything publishes.

What does schema markup actually do for AI search citation?

Schema markup is a machine readable label on your content that helps a crawler interpret the page. It is not, alone, a reason an engine cites you.

Schema markup, the JSON-LD blocks you add to a page under the schema.org vocabulary, tells a crawler what a piece of content actually is: a question and answer pair, a sequential process, an article with a named author and a publish date.

Our own AI Search Visibility Field Manual lists structure and schema among the six levers that decide AI citation, alongside crawlability, correct delivery, freshness and original data, and it says plainly that structure and schema matter less than the hype around them suggests.

This playbook is the deep dive on that one lever: what the evidence says it actually does, which types are worth the afternoon, and the mistakes that turn correct markup into decoration. For the broader discipline it sits inside, start with what GEO actually is.

Does adding schema increase AI citations? What the May 2026 Ahrefs test found

A May 2026 Ahrefs test added schema to 1,885 pages against matched controls and found the shift was close to statistical noise on two engines, and a small decline on a third.

Vendor guides routinely point out that pages cited by AI engines carry schema markup far more often than pages that are not cited, and use that gap to argue schema drives citations. Ahrefs tested that claim directly, in a study published 11 May 2026.

Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, each matched against three control pages on different domains with similar citation volume beforehand that never added schema. The difference in differences: Google AI Overviews citations fell 4.6% relative to the controls, a small but statistically real decline. Google AI Mode moved 2.4% and ChatGPT moved 2.2%, both too small to separate from noise.

Ahrefs also found cited pages were almost three times more likely to carry schema than uncited ones, the exact correlation the vendor guides lean on. Its own reading is that this reflects site quality: the engineering discipline to ship clean JSON-LD tends to sit on the same pages doing everything else right, not that the markup itself is pulling weight.

One limit is worth stating plainly, because it is the kind of detail a search snippet drops. Every page in the Ahrefs sample was already earning 100 or more AI Overview citations before schema was added, so the test says nothing about whether markup helps a page that has not been discovered yet. If you are starting from zero, our own step by step GEO audit is the place to start, not a schema sprint.

None of this changes the stakes of being cited at all. AI referral traffic already converts 23 times better than Google's, by our own measurement. It just means schema is not the lever that gets you there.

Google retired FAQ rich results in May 2026. Does FAQPage schema still matter?

Yes. Google stopped showing the FAQ dropdown in Search from 7 May 2026, but FAQPage markup is still valid, and other engines, including Bing's Copilot, still read it on their own terms.

Google's Search Central documentation confirmed that FAQ rich results stopped appearing in Search as of 7 May 2026, as Search Engine Roundtable reported on 11 May 2026. Google is withdrawing the related reporting tools in stages: the FAQ search appearance filter and Rich Results Test support go in June 2026, and Search Console API support for the feature follows in August 2026.

The visible dropdown in Google's results page is gone. FAQPage remains a valid schema.org type, and Google's own guidance says leaving the markup in place causes no harm, it is simply ignored by that one feature now.

What a Google feature retiring does not do is remove the markup from every other engine's view of a page. Microsoft's Fabrice Canel, then a principal product manager on Bing, said at SMX Munich in March 2025, as Search Engine Land reported, that schema markup helps Microsoft's own LLMs understand content, a use with nothing to do with Google's SERP dropdown.

It is the same lesson we drew from why llms.txt does not earn citations: whether a mechanism is visible, trendy or recently retired is not evidence about whether it earns a citation. Removing FAQ markup because the dropdown disappeared also removes a hint for the recommendation layer described in the new front page is the name an AI recommends, where a model is working from a handful of parsed signals, not a rendered results page.

Which schema types are actually worth implementing?

Five types cover almost every B2B page: FAQPage, HowTo, Article with dateModified, Organization and Person. Each signals something different, and no AI engine needs more than these to place you correctly.

None of this matters if the page is not reaching the crawler correctly in the first place. We have seen flawless schema sit on a page an AI crawler could not read at all, because of a single wrong header: the content type mistake that makes perfect HTML invisible. Fix delivery before you touch markup, confirming access with the AI Crawler Access Playbook first.

Why must every FAQ schema answer match the visible page word for word?

A mismatched answer leaves an engine checking both copies with no way to know which is correct. We refuse to publish anything from our own engine where the two do not match exactly.

FAQPage schema asks you to repeat, inside a script tag, the exact answer a visitor already sees on the page. The temptation is to tidy the schema version, shortening it or rewording it for the dropdown that, as the previous section covered, Google no longer renders anyway.

Do not do this. An engine that parses both the visible answer and the structured one and finds them different has no reliable way to know which is the real one, and some will simply discount the markup rather than guess.

This is not a theoretical risk for us. Every FAQ item on this page, and on every page we publish, is written into the schema from the same string as the visible answer by the build step itself, the same discipline covered in how to get cited by ChatGPT, Perplexity and Google AI. If the two ever drifted, it would mean the pipeline broke, not that an editor chose a shorter version for the schema.

The same build step refuses to publish a page at all if the JSON-LD it generates does not parse as valid JSON. That is a low bar, every page should clear it, and it is still worth stating plainly: the pipeline that wrote this guide will not ship broken schema, because nothing enforces that except the system that writes it.

How do you implement schema for AI citation without breaking anything?

Audit what already exists, map each page type to the schema types above, write markup from the same copy as the visible page, validate that it parses, then monitor citations rather than rich result features.

Schema will not make up for a crawlability or freshness problem, so treat this as a correction and consolidation pass, not a sprint to add markup everywhere at once.

1. Audit what is already on the page. Pull the current JSON-LD from every template, including the ones you do not remember adding it to. Templates drift out of sync with content changes more often than anyone checks.
2. Map page types to schema types. Decide once, per template, which of FAQPage, HowTo, Article, Organization and Person genuinely apply, using the table above. Do not add a type a page does not have.
3. Write the schema from the same source as the visible copy. Generate structured data from the same field or string that renders on the page, so an edit cannot update one copy and miss the other.
4. Validate that it parses, confirming more than that it exists. Confirm every script block is valid JSON before it ships. A tag that is present but broken is worse than no tag at all.
5. Confirm crawler access before judging the result. Rule out a delivery or access problem, using the AI Crawler Access Playbook's checks, before deciding schema made no difference.
6. Monitor citations, not rich result features. Track AI referral sessions the way the AI Search Attribution Handbook sets out. Google's own FAQ rich result reporting is being withdrawn through August 2026, so it was never a metric worth building a habit around.

What mistakes turn correct schema into a wasted afternoon?

Treating schema as a shortcut around SEO fundamentals, copying another page's markup without updating every field, letting answers drift from the page, and expecting markup to fix a delivery problem.

• Treating schema as a replacement for the fundamentals. Markup sits on top of the basics, not instead of them. SEO is not dead, it is the floor GEO stands on, and schema is decoration on that floor, not a substitute for it.
• Copying another page's JSON-LD and forgetting to change every field. A duplicated Organization block or a stale dateModified is worse than nothing, because it tells a crawler something false about the page.
• Letting the schema answer drift from the visible one after an edit. This is the exact failure the matching rule above exists to prevent, and it happens quietly, months after the page first shipped.
• Expecting schema to fix a crawl or delivery problem. If the page is not reaching the crawler correctly, no amount of markup on top changes that.
• Chasing llms.txt instead of the levers schema and structure actually belong to. We covered that specific substitution in the evidence that llms.txt does not earn citations.

Our take: schema is real, low cost and worth doing correctly. It is not the lever that moves the needle on its own, the Ahrefs evidence above says so plainly, and a team that spends a quarter on markup instead of on freshness and original data has spent it on the wrong one of the six levers.

Frequently asked questions

Does adding schema markup increase AI citations?

Not on its own. Ahrefs' matched pairs test of 1,885 pages, published 11 May 2026, found the change was close to statistical noise on Google AI Mode and ChatGPT, and a small decline on Google AI Overviews. Schema is one of several levers, not a standalone driver.

Is FAQPage schema still worth using after Google retired FAQ rich results?

Yes. Google stopped showing the FAQ dropdown in Search from 7 May 2026, but FAQPage markup is still a valid schema.org type, and other engines, including Bing's Copilot according to Microsoft's Fabrice Canel, continue to read it independently of that one Google feature.

Which schema types matter most for a B2B site?

FAQPage, HowTo, Article with dateModified, Organization and Person cover almost every page. Add a type only where it genuinely applies, not as a blanket default across every template.

Why does a FAQ schema answer have to match the visible page exactly?

Because an engine checking both copies has no reliable way to know which is correct if they differ, and some will discount the markup rather than guess. Generate both from the same source so they cannot drift apart.

What should be fixed before touching schema at all?

Crawler access and the content type header the server sends. Flawless schema on a page an AI crawler cannot read or parse correctly earns nothing, a failure mode that shows up even on pages with otherwise perfect HTML.

How often should schema markup be re-audited?

At least once a quarter, and immediately after any template change. Markup drifts out of sync with content silently, and the failure stays invisible until something checks it.