Does schema help AI citations? What the data shows

Almost every AI search checklist includes “add schema markup”. It sounds logical: structured data should be easier for machines to understand. But logic and evidence are different things, and now there’s evidence.

Here’s what the best study so far found, what Google says, what schema is still useful for, and where the time is better spent if your goal is to be named and cited in AI answers.

The study: 1,885 pages that added schema

Ahrefs looked at 1,885 pages that added JSON-LD schema between August 2025 and March 2026, and matched them with 4,000 control pages that never added it. They compared AI citations in the 30 days before and after the schema went live, using a matched difference-in-differences test (which compares the change in the schema group with the change in the control group over the same period) and three further checks.

Ahrefs’ conclusion: “Adding schema produced no major uplift in citations on any platform.”

Schema myths and what the evidence says

The pattern behind these claims is reasonable-sounding logic without measurement. When someone recommends schema for AI, ask what evidence they’re relying on. Most of the time it’s an assumption rather than a test.

What the study can and can’t tell you

• It studied pages already cited. The sample was pages that already had 100+ AI citations. Schema might behave differently for pages with none, but there’s no evidence yet that it does.
• It pooled schema types. Article, FAQ, Product, HowTo and Organization schema were analysed together.
• It measured 30 days. A long-term effect isn’t ruled out, but nothing in the data suggests one.
• It’s one study. But it’s a controlled one, which is more than most “schema helps AI” claims rest on.

What Google says

Google’s guidance on its AI features is direct: “There’s also no special schema.org structured data that you need to add”, and you “don’t need to create new machine readable files, AI text files, or markup to appear in these features.” The requirement is simply that a page is indexed and eligible to show a snippet in Search.

Why schema doesn’t move AI citations much

The most likely reason is that AI assistants read the page itself. They retrieve it, read the text and quote passages. Schema describes the page in a separate block of code, but the passage that gets quoted is the visible text. If the text is clear, schema adds little; if the text is unclear, schema doesn’t fix it.

What does show up in the data is about the text and its match to the question. In Ahrefs’ study of 1.4 million ChatGPT prompts, cited pages had titles much closer in meaning to the prompt, and pages with readable URLs were cited more often. In the GEO study, adding statistics, quotations and source citations to the text improved visibility in AI answers.

What schema is still good for

Don’t rip out schema you already have. It still has real jobs in normal search:

If you run an ecommerce store, Product schema is still worth keeping accurate. Prices and availability in search results matter to shoppers whether or not AI cites the page. Just make sure the same price, stock status and key specs also appear as plain text on the product page, because that visible text is what an AI assistant reads when it describes your product.

Where to spend the time instead

If you were about to spend a week on schema for AI, here’s what the evidence suggests will do more:

1. Retitle key pages around the exact buyer question they answer. Title-to-question match is one of the clearest signals in ChatGPT citation data.
2. Rewrite the first sentence of each section so it answers the heading directly.
3. Add evidence: a sourced statistic and a named quote on each important page.
4. Put comparisons and prices in HTML tables, in the visible text.
5. Check crawler access and rendering, so AI crawlers can read the text at all.
6. Get listed on the roundups and review sites AI cites for your category.

If your page isn’t cited at all yet

The study looked at pages AI already cited, so it’s fair to ask whether schema helps a page that’s invisible. There’s no evidence that it does, and there are better explanations for invisibility. Work through these first:

1. Can AI search crawlers reach the page? Check robots.txt and your CDN.
2. Is the text in the HTML, or loaded by JavaScript after the page opens?
3. Is the page indexed in both Google and Bing?
4. Does the title say, in buyers’ words, the question the page answers?
5. Does each section answer its heading in the first sentence?
6. Is your brand mentioned on the third-party pages AI cites for this topic?

Each of these has a clear mechanism: a crawler that can’t reach or read a page can’t cite it, and a page that doesn’t match the question is less likely to be chosen. Schema changes none of them. Our 10-point diagnosis covers the full list.

Before and after: the change that matters

Take a pricing page with perfect Product and Offer schema, but visible text that says only “Simple pricing for growing teams” above three cards with prices rendered as images. To an AI assistant reading the page, there’s nothing to quote about price.

Now the same page with no schema changes, but rewritten: the first line says “Plans start at $15 per user per month, billed annually”, the plans are in an HTML table with what each includes, and a short FAQ answers “Is there a free trial?” and “Can I pay monthly?” in plain text. That page has several passages an AI answer can use. The schema was never the problem; the text was.

What about FAQ schema?

FAQ schema is often recommended specifically for AI, because AI answers look like FAQ answers. But in the study, FAQ schema was among the types pooled together, and the overall effect was nil. What helps is having real questions and answers in the visible page: a heading phrased as the question, and a direct answer below it. Whether you also mark it up as FAQ schema doesn’t appear to matter for AI citations.

A two-hour schema check for normal search

Since schema still matters for rich results, a short check is worthwhile, just for the right reasons:

1. Run your key page types (product, article, homepage) through Google’s Rich Results Test.
2. Fix errors that stop eligible rich results, such as missing price or availability on product pages.
3. Make sure marked-up values match what’s visible on the page: same price, same rating, same date.
4. Remove schema for content that no longer exists, like expired offers or events.
5. Check Search Console’s enhancement reports for warnings once a quarter.

That’s maintenance, not an AI project. Two hours a quarter is plenty for most sites.

What to tell your team or agency

We'll keep our existing schema for rich results and fix any errors,
but we won't add schema as an AI citation tactic. A controlled study
of 1,885 pages found no meaningful lift in ChatGPT or AI Mode
citations, and Google says no special schema is needed for AI features.

Instead, for the next month:
1. Retitle our top 10 pages around the buyer question each answers
2. Rewrite the first sentence of every section to answer it directly
3. Add a sourced statistic and a named quote to each page
4. Pitch the 5 roundups AI cites most for our category

How to test it yourself

If you want to check on your own site, run a simple before-and-after test:

1. Pick 10 similar pages. Add schema to five; leave five alone.
2. Make no other changes to either group for 30 days.
3. Track citations for both groups: Bing’s AI Performance report for Copilot, plus a fixed set of prompts in ChatGPT and Perplexity.
4. Compare the change in the schema group with the change in the control group, not just the schema group on its own.

The control group matters. AI citations move around month to month, so without one, normal variation looks like a result. Keep the test running for the full 30 days even if early numbers look promising, and write down the result either way, so the question doesn’t come back every quarter.

A common scenario

An agency proposes an “AI readiness” package: FAQ schema on every page, an llms.txt file, Organization schema and a “structured data audit”. The work takes weeks and produces a nice report. Citations don’t change, because none of it changes the text AI reads or what other sites say about you.

A better use of the same weeks: rewrite the ten pages that should answer your top buyer questions, add evidence and tables, fix any crawler blocks, and pitch the five roundups AI cites most. That’s the work the citation data actually rewards. Our llms.txt guide covers the other half of that package.

Common mistakes

• Treating schema as the AI fix. AI reads the visible text.
• Adding FAQ schema to pages with no real FAQs. Write the questions and answers into the page instead.
• Skipping access checks. Schema on a page crawlers can’t reach does nothing.
• Over-trusting checklists. Ask for evidence behind each recommendation.
• Neglecting off-site work. Branded mentions on other sites correlate far more strongly with AI visibility than anything you can add in code.