Does schema markup still matter for AI search?

Somewhere in your backlog there is a ticket that says add schema markup for AI search. It came from a GEO audit, an agency proposal, or a conference slide. Nobody on the thread has asked what the evidence for it is, because it sounds too obviously correct to question — of course a machine prefers machine-readable data.

The honest answer is not no. It is that schema markup is worth implementing for reasons that have almost nothing to do with AI, and that the specific reason you keep being given — that it buys you citations in ChatGPT or AI Overviews — is unproven. Those are two separate claims and they get collapsed constantly. Keeping them apart is most of what this post is for.

Start with the news a lot of sites missed: FAQ rich results are gone

On May 7, 2026, FAQ rich results stopped appearing in Google Search. Not restricted, not reduced — removed, for everyone. Google announced it through a notice in its structured data documentation rather than a blog post, and gave no public explanation.

This was the second half of a longer retreat. HowTo rich results were deprecated in August 2023, and at the same moment FAQ rich results were restricted to well-known, authoritative government and health websites. The May 2026 change removed eligibility even for those. The tooling followed: in June 2026 Google dropped the FAQ search-appearance filter, the rich result report, and Rich Results Test support, and Search Console API support for FAQ rich results ends in August 2026.

Be precise about the failure mode here, because it is the same one that makes the AI question worth handling carefully. Nobody was wrong to build FAQ schema in 2021. It worked, it took SERP real estate, and you could measure it. The mistake is not noticing when a tactic's payout gets switched off — and then continuing to sell it.

What the public evidence on schema and AI actually says

Here is everything a careful person could find as of mid-2026, laid out in full rather than filtered for whichever direction you already lean.

• Google, 2026. Its guide to optimizing for generative AI features lists “Structured data isn't required for generative AI search” among the myths it debunks, while noting structured data remains useful for rich results. Google's separate AI features documentation goes further: “You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data.”
• Google, April 2025. An earlier statement said structured data “offers advantages in AI search” — with no specifics about which advantages, which surfaces, or how anyone measured it.
• Microsoft, March 2025. Bing said schema markup helps its LLMs understand content for Copilot. This is the only first-party vendor statement anyone has produced that actually supports the practice for LLMs, and Copilot is a small share of the market.
• Search Atlas, 2026. One study compared domain-level schema coverage against citation rates across OpenAI, Gemini, and Perplexity outputs and found no correlation — domains with extensive markup were cited no more often than domains with little or none. A single study, reported by the vendor that ran it, and flagged by its own authors as observational rather than causal.
• Everyone else. As of March 2026 there were zero peer-reviewed studies on schema's effect on LLM visibility, and no public technical statements from OpenAI, Anthropic, or Perplexity about whether their systems parse schema at all.

Notice what is absent from that list. No controlled experiment. No operator confirming its retrieval pipeline reads JSON-LD. No before-and-after with a disclosed sample size. The confident advice you have been given rests on one vendor statement, one vague Google line, and a great deal of inference.

The Princeton study did not test schema

The evidence most often offered for schema-drives-citations is the 2024 GEO paper out of Princeton and IIT Delhi (arXiv:2311.09735, published at KDD '24). It is a real paper with real findings, and it turns up in schema pitches constantly. It did not test schema markup.

What it tested was content strategies applied to source pages, measured against generative-engine responses. Adding quotations produced roughly a 41% improvement in visibility. Adding statistics and citing sources produced gains in the 30-40% range. Fluency optimization landed around 15-30%. Keyword stuffing showed minimal or negative effect. Structured data does not appear among the tested strategies at all.

The paper needs caveating in its own right, too. It is roughly two years old, it evaluated GPT-3.5-turbo plus one live Perplexity system, and the authors explicitly warned that their results “may need to adapt over time.” Two years is several model generations. Treat its numbers as directionally interesting rather than current.

The case for implementing schema anyway

None of the above says skip it. It says implement it for the reasons that hold up, which are unglamorous and real.

1. Rich results still exist. Product, Review, Breadcrumb, Article, LocalBusiness, and Event results all still render in Google Search, and every one of them requires valid markup. That is a visible, measurable payout with nothing to do with AI.
2. It is machine-readable ground truth. Schema states unambiguously that this string is a price, that one is an author, this is the legal name of the business. Any parser — search index, retrieval layer, aggregator, browser extension — has one less thing to guess. That is a reasonable prior, not a demonstrated mechanism, and it should be described that way.
3. It is cheap. For most sites this is a plugin, a template change, and an afternoon of validation. The cost side of the ledger is genuinely small.
4. The downside is asymmetric. Google's structured data guidelines require markup to reflect the page — “Don't mark up content that is not visible to readers of the page” — and prohibit misleading markup such as fake reviews. A mismatch between your JSON-LD and your visible content is a manual-action risk, not a free bet. Wrong schema is worse than no schema.

So the honest position: implement it because it is correct and inexpensive, not because someone promised it buys citations. If it turns out to help with AI visibility, you already have it. If it does not, you have not spent much.

Which types are worth the effort in 2026

One caveat on the first row. The sameAs property — pointing your Organization block at your Wikipedia entry, Wikidata item, LinkedIn page, and other authoritative profiles — is widely recommended as an entity-disambiguation signal. That comes from vendors and practitioners, not from any confirmed platform statement. It is sensible, it costs nothing, and there is no published proof it does anything. Say that out loud when someone sells it to you as a ranking factor.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Organization",
  "@id": "https://example.com/#organization",
  "name": "Example Coffee Roasters",
  "legalName": "Example Coffee Roasters Ltd",
  "url": "https://example.com/",
  "logo": {
    "@type": "ImageObject",
    "url": "https://example.com/logo.png",
    "width": 512,
    "height": 512
  },
  "description": "Small-batch coffee roaster in Portland, Oregon, supplying cafes and offices since 2014.",
  "foundingDate": "2014-03-01",
  "contactPoint": {
    "@type": "ContactPoint",
    "contactType": "customer support",
    "email": "hello@example.com",
    "telephone": "+1-503-555-0142"
  },
  "sameAs": [
    "https://www.linkedin.com/company/example-coffee-roasters",
    "https://en.wikipedia.org/wiki/Example_Coffee_Roasters",
    "https://www.wikidata.org/wiki/Q00000000",
    "https://www.crunchbase.com/organization/example-coffee-roasters",
    "https://www.youtube.com/@examplecoffee"
  ]
}
</script>

Every value in that block should be findable on the site. The description should match your about page. If the sameAs list points at a Wikipedia article that does not exist, delete the line — an invented entity link is precisely the kind of mismatch Google's guidelines are written against, and it is the one place cheap schema turns expensive.

Measure it instead of assuming it

The reason this argument never resolves is that almost nobody instruments it. If you are going to invest in schema, invest another twenty minutes in being able to tell whether it did anything.

1. Validate what you actually ship, not what your plugin claims. Run each key template through Google's Rich Results Test and the Schema Markup Validator, testing the rendered page if you inject markup client-side.
2. Track rich-result impressions in Search Console by search appearance. That is the one place schema's effect is directly observable — and where you will watch the FAQ line go flat.
3. Roll out by template, not all at once. Adding Product schema to one category of pages and not another gives you a comparison you would otherwise never have.
4. Track AI citations separately from rich results, using prompt monitoring or referral parameters like utm_source=chatgpt.com. If schema helps there, that is where it will surface.
5. Give it a full quarter before concluding anything. Recrawl cycles, model updates, and vendor policy changes all move slower and more erratically than a monthly report.

What to do this week

1. Search your codebase and CMS for FAQPage. Decide deliberately whether to keep it — there is no penalty either way — and stop any active work producing more of it for rich results.
2. Check whether you have a single site-wide Organization block with an accurate sameAs list. Most sites do not, and it is the cheapest entity work available.
3. Audit for mismatches: prices, ratings, authors, and dates in your JSON-LD that do not match what a reader sees. Fix those before adding a single new type.
4. Add BreadcrumbList and Article schema to your templates if they are missing. Both still render, and both are template-level work rather than per-page work.
5. Write down what you expect schema to do and how you would know. If the answer is “get cited by ChatGPT” and there is no way to check, you have bought a story rather than a tactic.