# Schema for AI Search: The Structured Data Evidence Guide

> Most advice on schema for AI search is outdated or unproven. See which schema.org types still earn anything today, backed by Google and a matched study.

- URL: https://missiongrowth.io/blog/schema-for-ai-search
- Published: 2026-08-27 · Updated: 2026-09-23
- Author: Ömer Furkan Aktaş, Founder, Mission Growth
- Publisher: Mission Growth. Company facts: https://missiongrowth.io/llms.txt

Schema markup, also called structured data, is a code block, usually JSON-LD, added to a page to describe its content in a format search engines and AI systems can read directly, instead of inferring it from the visible text.

Google's own guidance says structured data isn't required for generative AI search, and the one quasi-experimental test of adding it anyway found no major citation lift on any platform. The useful question about schema for AI search is which types still do anything at all.

Google's optimization guide states it plainly: "Structured data isn't required for generative AI search, and there's no special schema.org markup you need to add. However, it's a good idea to continue using it as part of your overall SEO strategy, as it helps with being eligible for rich results on Google Search."

That's a rich-result claim, not a ranking one: Google's own guidance ties the benefit to eligibility for rich results, never to where a page ranks.

Ahrefs tested the opposite bet directly, matching 1,885 pages that added JSON-LD between August 2025 and March 2026 against 4,000 control pages that didn't, and landed on the same conclusion from the other direction: no major uplift in citations on any platform.

The full three-source comparison behind that finding, Google's position, Ahrefs' test design, and the correlational numbers vendors lean on instead, is [the full evidence breakdown](https://missiongrowth.io/blog/ai-seo-services) already worked through elsewhere on this site. What's left here is narrower: which schema markup for AI search types still earn anything real, and which two the market keeps recommending long after Google stopped rewarding them.

## FAQPage and HowTo: the two types everyone still recommends that no longer do anything

Google narrowed FAQPage's rich result to a small list of sites and killed HowTo outright on the same day, September 14 2023.

Most schema advice still treats both as top picks; search "faq schema ai search" today and most of what comes back repeats it.

That changelog entry updated the FAQ structured-data documentation "to state that the feature is only shown for well-known, authoritative government and health websites." Unless the site is a government domain or a major health authority, FAQPage stopped earning a rich result that day, nearly three years before most people noticed.

The same entry removed the HowTo case study and its documentation outright, "as this rich result is no longer shown in search results, on both desktop and mobile devices." Neither correction is niche trivia.

One retired a type completely. The other fenced the remaining one off from every commercial and SaaS site in a single afternoon, three years before Google made the FAQ removal official for everyone else.

FAQPage's decline didn't stop there. The rich result disappeared entirely starting May 7 2026, "this feature will no longer appear in Google Search," and Google pulled the documentation the following month, confirming "the FAQ rich result feature is no longer shown in Google Search results."

For a commercial or SaaS reader, though, the 2026 date closes a door that had already been shut for years. The 2023 narrowing is the one that actually changed anything, since it excluded them regardless of what happened three years later.

::figure{src="/blog/figures/schema-for-ai-search-2.svg" alt="Timeline: Google narrows FAQPage to gov/health sites and deprecates HowTo on Sep 14, 2023; removes the FAQ rich result entirely on May 7, 2026." caption="Google narrowed FAQPage and killed HowTo the same day in 2023; FAQPage's rich result didn't fully disappear until May 2026." width="720" height="198"}

None of this makes FAQPage or HowTo markup harmful to add. Using either format for plain, machine-readable Q&A or step content is harmless. It earns no rich result and no special AI-citation credit.

That distinction decides whether schema work is [worth paying a vendor for](https://missiongrowth.io/blog/ai-seo-services): billing for FAQPage and HowTo as top AI-search recommendations means billing for advice that stopped being current in 2023.

## Schema for AI search: which types still earn something today

Organization, Article, Product/Offer and genuine Review/AggregateRating markup still earn real Google Search eligibility today.

None of them earn a proven AI-citation lift, and none of it counts as schema markup for AI visibility in the way vendor blogs use that phrase, a guaranteed citation lever.

It counts as something narrower and more useful: real Google Search eligibility, type by type. The wider list Google still rewards includes structural types like BreadcrumbList and image markup, but the four below carry most of the confusion, and most of the correction, current as of September 2026.

::dataset{key="schema-for-ai-search-types" name="Schema.org types and their Google Search status, September 2026"}

| Schema type | Google Search feature today | What to actually do |
|---|---|---|
| Organization / Person | Entity clarity, `sameAs` disambiguation; not a documented ranking or citation signal | Cheap and low-risk; add it, don't expect a citation lift from it alone |
| Article / BlogPosting | Standard eligibility for article-type search features | Keep author, date and publisher current |
| Product / Offer | Product rich result, a separate mechanism from a Shopping feed | Keep price, availability and currency current; a feed doesn't replace the on-page markup |
| Review / AggregateRating | Review rich result; fake, undisclosed-incentivized or cross-site-aggregated reviews are prohibited | Never aggregate ratings from another site; never include fake or undisclosed incentivized reviews |

A Product/Offer block and a Google Shopping feed answer two different eligibility questions. The feed drives Shopping placement; the on-page markup drives the Search rich result.

An ecommerce site running a Shopping feed doesn't get to skip Product/Offer schema. It's a separate submission, done independently of the feed.

SoftwareApplication is where two corrections collide. Google's required-properties table lists three items together: `name`, `offers.price`, and a rating or review, either `aggregateRating` or `review`.

Skip any one of the three and the rich result disappears completely. `applicationCategory` and `operatingSystem` are only recommended.

That's a real bind for a listicle-style tool entry with no genuine reviews yet. Google's review-snippet policy prohibits exactly the shortcuts that would fill the gap fast: "don't include fake or undisclosed incentivized reviews on your page or in your structured data markup," and "don't aggregate reviews or ratings from other websites."

A page with no first-party reviews collected has no legitimate route to the SoftwareApplication rich result until it collects some. There's no way to shortcut that wait.

None of this changes the causal picture from the intro. Adding any of these types still earns what it always earned, rich-result eligibility and entity clarity. Schema is one deliverable among several in the six-part breakdown of [ai search optimization services](https://missiongrowth.io/blog/ai-seo-services), not the AI-citation lever vendor blogs sell it as.

## Is your schema actually doing anything? How to tell

A schema markup block can pass every validator and still never influence an AI citation.

Validity and citation impact are two separate, separately measured claims. Trusting structured data for AI search means knowing which claim you're looking at.

A validator checks syntax. Impact is a separate question it can't answer. Google's own Rich Results Test and Search Console's Enhancements reports confirm a markup block is technically correct and eligible for a rich result.

Neither tool can tell you whether that markup changed how often an AI answer engine cites the page. Eligibility and impact are different claims, and only one of them is measurable today.

::figure{src="/blog/figures/schema-for-ai-search-1.svg" alt="Dot chart: schema's effect on AI citations, AI Overviews down 4.6%, AI Mode up 2.4%, ChatGPT up 2.2%, from Ahrefs' matched study of 1,885 pages." caption="Adding schema for AI search moved AI Overview citations -4.6% and left AI Mode and ChatGPT statistically unchanged, across 1,885 pages Ahrefs tracked." width="720" height="269"}

The question "does schema markup help AI Overviews" has exactly one test built to answer it directly. Ahrefs matched 1,885 pages that added JSON-LD against 4,000 pages that didn't. Only one of the three platforms moved by a statistically significant amount.

Google AI Overviews fell 4.6%, an average of roughly 12 fewer daily citations per page on that platform. Ahrefs' own chart flags the result as outlier-driven.

A handful of pages losing as many as 400 citations a day, or gaining 200, dragged the treated group's average down. Strip those out and the treated and control groups look similar.

AI Mode moved 2.4% and ChatGPT moved 2.2%, both statistically indistinguishable from zero.

That's part of why Ahrefs called it a null result: "we can't tell whether the schema did a tiny bit of good or nothing at all."

That magnitude is the actual lesson here. A matched test spanning 1,885 pages and 4,000 controls could barely separate a real effect from noise.

Tracking one site's own before-and-after AI-citation count after adding schema means measuring a signal smaller than the test built specifically to detect it. Treat any "we added schema and citations went up" claim, including your own, as an anecdote rather than evidence.

Whether any specific AI engine, ChatGPT, Perplexity, Gemini, actually parses JSON-LD as structured data rather than reading the surrounding text is still an open question.

State it that way rather than repeat either claim in circulation: that tokenization strips schema markup out, or that these systems specially fetch and parse it. None of the platforms has confirmed either one.

What a schema check can actually confirm:

- Run the markup through Google's Rich Results Test or validator.schema.org to confirm it's technically correct.
- Check Search Console's Enhancements reports to confirm eligibility for the rich result, a separate claim from validity.
- Don't attribute a single site's AI-citation count to one schema change. The study built to measure that effect across thousands of pages could barely detect it.

For the raw-HTML-versus-JS-injected implementation rules and mobile-parity checks behind that first bullet, see the [technical SEO checklist](https://missiongrowth.io/blog/technical-seo-checklist-2026).

## The one implementation trap specific to schema

A JSON-LD block added through Google Tag Manager is invisible to GPTBot, ClaudeBot and Perplexity's crawler.

That's the shortcut teams reach for specifically to skip a code deploy for markup, and it fails even when the rest of the page passes a View Source check.

Google Tag Manager is the fastest way to ship a JSON-LD block without a code deploy, and it's exactly the wrong tool for the job.

Vercel and MERJ measured actual crawler behavior across Vercel's own network in December 2024 and found that none of the major [AI crawlers](https://missiongrowth.io/blog/gptbot-ai-crawlers) execute JavaScript: "The results consistently show that none of the major AI crawlers currently render JavaScript. This includes: OpenAI (OAI-SearchBot, ChatGPT-User, GPTBot), Anthropic (ClaudeBot), Meta (Meta-ExternalAgent), ByteDance (Bytespider), Perplexity (PerplexityBot)."

A JSON-LD block that GTM writes into the page at runtime never exists for any of those crawlers. Not delayed, not eventually indexed, simply never seen, because a GTM tag is JavaScript that runs client-side after the page has already loaded.

Two crawlers are the exception, and naming them matters because most advice treats "AI crawlers" as one undifferentiated group. Gemini renders JavaScript because it runs on Googlebot's infrastructure, and AppleBot renders through a browser-based crawler; both would see GTM-injected schema. The other five wouldn't.

This is easy to miss because it survives the check most teams actually run. Viewing a page's source, or even confirming that its visible content renders correctly, says nothing about where the schema block itself lives.

A page can pass every content-rendering check and still ship its structured data exclusively through Tag Manager. The surrounding text sits in the initial HTML response; the markup itself is written in afterward, by a script the non-rendering crawlers never run.

We ran into a version of the same rendering problem migrating our own site. We moved our own React single-page app to prerendered static HTML for 20 marketing pages because AI crawlers don't execute JavaScript.

The Tag Manager shortcut fails for the identical reason, one script tag at a time instead of a whole page.

The measurement dates from December 2024, and crawler behavior changes over time, so treat it as the current best evidence rather than a permanent state.

The fix doesn't depend on the date. Confirm the JSON-LD block itself is present in the raw HTML response, a separate check from confirming the surrounding content renders.

A View Source check that only eyeballs the visible text won't catch a tag-manager-injected block; the markup has to be in the document that arrives before any script runs.

None of this makes schema markup pointless. It makes "add more schema for AI search" the wrong instruction.

Google's own position and Ahrefs' test both point the same way: no requirement, no proven citation lift. The type-by-type reality above, not a blanket build-out, is what the work should target: rich-result eligibility and entity clarity, with FAQPage and HowTo dropped from the list entirely.

Pull up the site's current structured-data inventory and check it against two things: does each type still hold Search Console eligibility, not merely valid markup, and is any of it shipping through Google Tag Manager instead of the raw HTML response.

Those two checks catch the actual gap between what a site's schema is supposed to do and what it's currently doing.

## FAQ

### Does schema markup guarantee my products will be cited by ChatGPT or Gemini?

No. Google's own guidance says no special schema.org markup is required for generative AI search, and the one quasi-experimental test of adding it, matching 1,885 pages against 4,000 controls, found no major citation uplift on any platform. Schema still earns Google Search rich-result eligibility; it isn't a citation guarantee.

### Is schema markup still worth doing if I already invest in SEO?

Yes, for two concrete Google Search benefits: rich-result eligibility and clearer entity signals, regardless of any AI-citation effect. Organization, Article, Product/Offer and genuine Review/AggregateRating markup still earn real Google Search features, and SoftwareApplication needs `name`, `offers.price` and a rating or review together, or none of it shows.

### Does schema markup help AI Overviews?

No proven uplift, and the evidence is the same as the general case. The one statistically significant result in the Ahrefs test, a small AI Overviews decline, was outlier-driven; Ahrefs never treats it as proof of harm.

### Which schema type should I add first?

Organization, plus whichever type matches the page's primary content, Article for a blog post, Product/Offer for a commercial page, validated against what's actually visible on the page. Add Review/AggregateRating only once genuine, first-party reviews exist to back it; the SoftwareApplication rich result requires a real rating or review before it can show at all.

### How often should I update schema markup?

Whenever the facts it describes change: price, stock status, `dateModified`. Schema that contradicts the visible page is a validation failure regardless of AI search. Run it back through a validator and Search Console's Enhancements report after any update that touches those fields.