Guide 07 · Evidence review
Schema markup for AI: does it increase citations?
No large effect has been shown. The best test so far is Ahrefs’ May 2026 study: 1,885 pages that added schema, compared with 4,000 similar pages that did not. Citations rose 2.4% in AI Mode and 2.2% in ChatGPT, both within noise, and fell 4.6% in AI Overviews. That drop was statistically significant but small, and Ahrefs says it cannot pin it on schema.
One study is not settled science. It is also one more study than the people recommending schema for AI citations have produced.
Google’s guide to optimizing for generative AI features now says it directly: structured data is not required for generative AI search, and there is no special schema.org markup to add. Schema still has uses in ordinary results, which is why this page does not tell you to remove it.
Why is schema so often recommended for this?
Because the reasoning sounds airtight. Structured data describes your page in a machine-readable form; language models consume machine-readable data; therefore structured data should help a model understand and cite you. Every step of that reads as obvious.
The gap is in the middle. A retrieval crawler fetches your page and the answer is generated from its text. Nothing in that path requires the JSON-LD block, and no engine has said it weights one. The argument establishes that schema could be used, not that it is.
It also inherits a decade of SEO habit, where schema demonstrably does something — it produces rich results in classic search. Carrying a working technique across to a new system and assuming it still works is the single most common error in GEO advice, and it is not specific to schema.
It means one experiment looked for a lift and did not find one. It does not mean schema is worthless, and it does not mean a larger or better-designed study would find nothing. It means nobody has yet shown the effect that is being sold to you.
Should you remove your schema?
No. Schema has real uses and none of them are in dispute here.
| What schema is for | Evidence | Verdict |
|---|---|---|
| Rich results in classic search | Well established | Keep |
| Unambiguous entity identification | Mechanically true — it says what you are | Keep |
| Consistent facts across your own pages | A discipline, not a ranking factor | Keep |
| Lifting AI citations | One test, no effect | Do not expect it |
The second row is worth dwelling on for a plugin company. Accurate Organization or SoftwareApplication markup states in a fixed, parseable place who publishes the product, what it is called and what it costs. Whether or not a retrieval crawler reads it, that removes ambiguity for anything that does — and getting your own product facts stated consistently is worth doing on its own terms.
What has been observed is that engines hold stale product facts for a long time, and hold different stale facts from each other. Perplexity named a sunset membership plugin in all three of its membership answers, twice in third position, once calling it the best lightweight option; Google AI Mode did not name it at all. Keeping your own machine-readable description correct is cheap insurance against being the next example.
What should the effort go into instead?
Schema is close to free, so this is not really a trade-off — it is about where the next hour goes once the markup is correct. On the measured evidence, in this order:
Being retrievable at all. Valid certificate, server-rendered HTML, retrieval crawlers not blocked at robots.txt or at the firewall. Binary, and broken more often than anyone expects.
Original data nobody else has. A number that exists only because you measured it is the one thing a model cannot get from a competitor’s page.
Structure that matches the question. Question-form headings are cited at about 18% against 8.9% for statement headings, and 44.2% of citations come from the first 30% of a page. Answer first, then explain. That is the closest thing to “structured data” that the evidence actually supports — structure in the prose, where the answer is generated from.
What would change my mind?
A single controlled experiment is thin evidence for a general claim. It tested one markup type on one site against one engine’s AI Overviews. A different schema type, a different engine, or a larger sample could all produce a different result.
A replication finding a lift would move this page immediately, and I would rather publish that than defend a position. So would an engine stating that it reads structured data when assembling an answer — that would be an announcement, not an inference, and it is the kind of thing that gets announced.
Until then the honest framing is not “schema does not work”. It is: schema has one measured test behind it and that test found nothing, so do not budget for a lift you have no evidence of. Keep the markup, keep it accurate, and spend the expectations somewhere they are earned.
How this was made: figures come from our own runs in the engines’ public interfaces or from the studies named above. Drafted with an AI assistant (Claude) and reviewed by Alauddin Chowdhury. Method and AI use.
Are the engines describing your plugin correctly?
Send me your wordpress.org slug. I will ask both engines about your plugin by name and send back what they say — including any version, price or ownership detail that is out of date, which has to be corrected per engine rather than once. Free, no call, usually within a day.