Does Schema Markup Help AI Citations?

Schema markup does not have a demonstrated effect on whether AI assistants cite a page. Pages that get cited are more likely to carry JSON-LD, but the controlled tests that compare pages that added it with similar pages that did not find no reliable lift. Google's own guidance says no special markup is needed to appear in its AI features. Schema is still worth having for the search features it supports. It is a poor bet as an AEO lever.

Key Takeaways

• The correlation is real and misleading. Cited pages more often have structured data, and so do well-built pages that were likely to be cited anyway.
• Matched-control studies of pages that added schema found changes that were small, mixed in direction or indistinguishable from zero.
• Google says no special markup is required for its AI features. Standard Search requirements apply.
• In one fixture test, consumer assistants on the web did not return values from a valid JSON-LD block, and several fetch layers stripped script elements before the model saw the page.
• Keep schema for what it is documented to do, and do not count it as an AI citation tactic.

Why It Looks Like It Works

Take any large set of cited pages and check how many have JSON-LD. A high share will. Now take a set of pages that were never cited. A lower share will. That gap is what vendors often quote, and it is a correlation. Sites that invest in schema tend to invest in everything else too: clean titles, fast pages, strong inbound links, original content. Any of those could explain the citations.

The fix for a correlation is a comparison group. Take pages that added schema, pick similar pages that did not, and look at both before and after. That design is what separates "pages with schema got cited more" from "pages that added schema got cited more than they otherwise would have".

What the Controlled Tests Found

Three analyses published in 2025 and 2026 looked at this, and none found a positive effect that survived a control.

• A matched-control study of about 1,885 pages that added JSON-LD, compared with roughly 4,000 similar pages that did not, saw citations in Google AI Overviews fall slightly, and ChatGPT and AI Mode move by amounts not different from zero. The raw before-and-after number for AI Mode looked strong until the controls were included, because the controls gained almost as much.
• A cross-sectional study from Search Atlas (December 2025) found visibility was the same across five buckets of schema coverage.
• A single-site test from Otterly (March 2026) saw brand coverage rise about 6 percent after schema went live, while competitors who changed nothing rose as much. Six of seven engines could not return the JSON-LD on the page when asked for it.

Each has limits. They pool all schema types, most cover pages that were already cited, and none is a randomised experiment. The honest summary is that no study shows a lift, several show nothing, and no one has published a clean test for a single schema type on pages that were not yet cited.

What Assistants Can Actually See

Part of the explanation is mechanical. A fixture test from WordLift in July 2026 put values inside a valid JSON-LD block on a test page and asked consumer assistants to read them back. ChatGPT and Claude on the web returned none of the values in three attempts each. Several of the fetch layers in between removed every script element before the text reached the model. JSON-LD lives in a script element.

That does not mean structured data is invisible everywhere. Google's systems parse it for search features, and Google's crawler renders JavaScript. It means that for the assistants that fetch a page and read its text, markup that exists only in a script tag may never arrive. Whatever you need an assistant to know should also be in the visible text of the page.

What Google Says

Google's AI optimisation guide, updated 10 July 2026, states that no special markup is required for its AI features, and that pages need to meet the ordinary Search requirements: indexable, with a snippet-eligible page. Google's Search Central changelog says FAQ rich results stopped appearing in Search on 7 May 2026, so FAQPage markup added for AI purposes no longer even earns the Google feature it was designed for.

When Schema Is Still Worth Adding

• Product, review, event, organisation and article markup for the Google features each one supports. That is a documented benefit with its own measurement, in Search Console.
• Consistent entity information. Organisation markup that matches your visible About page does no harm and helps systems that do read it agree on who you are.
• Paywall markup if you have a paywall, so Google can tell it from cloaking.

The test for any markup is whether you would still add it if AI assistants did not exist. If yes, add it. If it is only there for an AI citation, the evidence does not support the effort.

What to Do Instead

Spend the effort where there is stronger evidence.

1. Put the answer in the page text. Lab studies find that on-topic content, a stated price and a recent date change which source a model cites first. Markup is not on that list.
2. Check that assistants can reach the page at all. A robots.txt rule or firewall that refuses an assistant removes you before any markup matters. See AI crawler blocking.
3. Measure before and after with a control. If you do add schema, pick comparable pages you leave alone and track both. Pages nobody touched can swing widely within two weeks, which is why a before-and-after number alone proves nothing (see does updating old content help AI citations for the figures).
4. Sample enough answers. One answer is an example, not a reading. How to track brand mentions in ChatGPT explains how many.

Where This Leaves the Question

If someone sells schema as the way to get cited by AI, ask for a test with matched untouched pages and a count of the answers behind it. The published evidence points to nulls. If your developer asks whether to add Organisation markup, the answer is yes, for ordinary reasons, and you should not expect your next AI readings to move because of it.

Lectern measures whether assistants name and cite your brand across ChatGPT, Gemini, Perplexity and Google AI Overviews, so a markup change can be read against a baseline instead of a hunch. The free Hobby plan runs a weekly check of 10 buyer questions, and the paid Lead plan, priced per brand on the pricing page, refreshes daily.