Structured data for AI: what is required and what is not

This is one of the few questions in AI search with a documented answer, which makes the amount of money changing hands over it remarkable.

The pitch usually arrives as a technical audit. Your site is missing LocalBusiness markup, FAQ markup, Service markup, so AI systems cannot understand you, so here is a quote. The premise is stated as fact. The premise is wrong.

What does Google actually require for its AI features?

Two conditions, both boring.

The page must be indexed. The page must be eligible to be shown with a snippet, meaning you have not blocked snippets with a meta robots directive.

That is the list. Google's guidance on optimizing for generative AI features states there are no additional technical requirements. On schema specifically it says it twice, in almost the same words: "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add."

If your page can appear in a normal Google result with a snippet under it, it is technically eligible for the generative features. Nothing further is needed.

This follows from how the features are built. They retrieve from the Search index and use query fan-out over core Search ranking. There is no second index with a stricter entry requirement, which is the assumption most schema pitches quietly rely on.

Is FAQ schema still worth adding for AI answers?

No, and this one has a date on it.

Google published the change on 8 August 2023, in a Search Central post authored by John Mueller. On FAQPage, verbatim: "Going forward, FAQ (from FAQPage structured data) rich results will only be shown for well-known, authoritative government and health websites. For all other sites, this rich result will no longer be shown regularly." On HowTo, per the update appended to the same post: "As of September 13, Google Search no longer shows How-to rich results on desktop, which means this result type is now deprecated."

Check the current state yourself, because it is faster than arguing. Open Google's live gallery of search result types. There is no FAQ guide and no HowTo guide in it. The list runs Article, Book actions, Breadcrumb, Carousel, Course list, Dataset, Discussion forum, Education Q&A, Employer aggregate rating, Fact check, Event, Image metadata, Job posting, Local business, Math solvers, Movie carousel, Organization, Shopping types, Profile page, Q&A, Recipe, Review snippet, Software app, Speakable, Subscription and paywalled content, Vacation rental, Video.

FAQ rich results have been dead for non-government and non-health sites since August 2023, and they were dead before AI Overviews existed. Anyone selling FAQ schema as an AI Overviews play is selling a rich result that stopped appearing three years ago, relabelled.

The FAQ content is a different matter and it is still valuable. Real questions, answered plainly in the first sentence, are exactly the units a generative answer can lift. It is the markup that stopped doing anything, not the writing.

What is an AI schema package actually worth?

Sort it component by component and price each one. The arithmetic takes twenty minutes and you can run it on any quote.

The quote. A $3,200 one-time "AI schema package" with five components: FAQPage, HowTo, AggregateRating on the homepage pulling in your Google reviews, LocalBusiness, and Organization with sameAs links. Billed evenly, that is $640 per component (derived).

Component one, FAQPage. No rich result for your site type since August 2023, and no FAQ guide in Google's current gallery. Google says no schema is required for generative features. Value: zero.

Component two, HowTo. Deprecated on 13 September 2023. Value: zero.

Component three, self-serving AggregateRating. Google's structured data policies prohibit marking up reviews about your own business on your own site. This is the single most common schema violation on local business websites, usually installed by a plugin that scrapes the client's own Google reviews. Value: below zero. It carries manual action risk with no upside.

Component four, LocalBusiness with the most specific subtype. Real. Local business is in the current gallery, and accurate machine-readable facts help entity resolution. Value: real.

Component five, Organization with sameAs pointing at every verified profile. Real, and useful for exactly the same reason. Value: real.

Add it up. Three of five components are worth nothing or less, which is $1,920 of a $3,200 quote, or 60%derived. The two real components are perhaps an hour each, and on a modern platform some of it is generated for you already.

The honest note on my own arithmetic. Real quotes are not billed evenly, so ask for the hours behind each line rather than assuming. The sorting is what matters, not my split. Run it on your quote and the shape usually holds.

And the comparison. The same $3,200 against Business Profile category work has an actual large-sample finding behind it. Darren Shaw's study of 1.8 million Google Business Profiles across 4,209 categories found the exact match category outranks every adjacent one, with the pattern repeating for almost every category in the dataset. Your competitors' categories will show you where to start, and benchmarking their review cadence will show you the pace to match.

Where does the "schema helps AI" claim come from, then?

From a real statement about a different mechanism, and the conflation is where the money is made.

Fabrice Canel, a principal product manager at Microsoft Bing, said on stage at SMX Munich in March 2025 that schema markup helps Microsoft's large language models understand web content. He also pointed publishers at IndexNow for pushing fresh content. That is a first-party statement from a platform operator and it is not nothing.

Here is the distinction that vendors collapse. There are two possible paths by which markup could reach an AI answer.

The indirect path: search engines ingest structured data at index time, use it for entity resolution, and the already-resolved understanding of who you are feeds the grounding layer behind the generative answer. This is well supported, it is what Canel's statement is consistent with, and it is a good reason to implement accurate markup.

The direct-parse path: a language model reads your JSON-LD block at answer time and treats it as a preferred source. There is no credible published test demonstrating that this drives citations at scale. Not a debunk, an absence. If someone shows you one, I want to see it.

The commercial version of the pitch always implies the second path, because the second path justifies a package rather than an hour of cleanup. Ask which mechanism the quote is priced against. It is a fair question and the answer tells you everything.

Both paths, incidentally, depend on a rule Google states directly in its AI features guidance: make sure your structured data matches the visible text on the page. Markup asserting hours your page contradicts is worse than no markup.

Which structured data is still worth implementing?

Plenty, for reasons that have nothing to do with AI.

For a local business: LocalBusiness plus the most specific subtype available, meaning Restaurant, Dentist, HVACBusiness, Plumber, Electrician, DaySpa and so on rather than the generic parent. Organization with sameAs pointing at every verified profile you hold. And Service for what you actually sell.

That is a short list and it is deliberately short. Not for AI visibility. For entity disambiguation and for the rich results that still exist.

Skip FAQPage. Skip HowTo. Skip any plugin that marks up your own reviews on your own site.

This is the same pattern that runs through the whole category, which is why so much AEO work turns out to be SEO with a new name. Sometimes that is fine, because the underlying work is real. It stops being fine when the renaming is used to charge twice.

How do you tell whether your markup is doing anything?

Test it, then look at what actually happens.

Run the page through Google's Rich Results Test. It tells you whether the markup is valid and which features the page is eligible for. If the answer is that it is eligible for nothing, the markup is decoration.

Then check Search Console for the corresponding result types. If you have had valid Product markup for six months and zero rich result impressions, the markup is not the constraint.

For the AI side, stop guessing and read your logs. AI crawler activity is visible in your server logs, identified by IP range rather than user agent string, and it is the only fully deterministic first-party measurement available in this whole subject. If crawlers are fetching your pages, they are getting what they need. If they are not, more markup will not summon them.

What to do this week

1. Run your three most important pages through the Rich Results Test. Note what each is eligible for. 2. Search your site source for `FAQPage` and `HowTo`. Delete both. Keep the questions on the page. 3. Check whether a plugin is marking up your own Google reviews as AggregateRating on your homepage. If so, turn it off today. 4. Fix any validation errors that remain. Invalid markup is worth cleaning up regardless of this debate. 5. Verify your LocalBusiness details match your Business Profile exactly. Hours are the field that drifts. 6. Run the component-by-component arithmetic above on any schema quote in front of you.

Got a schema quote in front of you?

Send me one page URL from your site at eric@seod.com and I will tell you what structured data is on it right now, whether it is valid, what it makes you eligible for, and whether the quote you were given is buying anything real. If you paste the line items, I will sort them the way this article does. No charge, no pitch.

I do this because it takes me about five minutes with public tools, and the alternative is an owner paying four figures for markup that qualifies them for nothing.

Otherwise, keep reading the AI search library.