How to Know If Schema and Structured Data Are Helping Your AI Citations (2026 Guide)

Key takeaways

• "AI-cited pages are 3x more likely to have JSON-LD" is a correlation. In Ahrefs' matched test of 1,885 pages, adding schema did not measurably raise citations in ChatGPT or AI Mode.
• To tell whether schema helps your site, you need a baseline, a treated group, a matched control group, and at least 4 to 8 weeks of data.
• Check crawl access and visible page content first. Schema can't help a page AI crawlers can't reach or parse.
• Schema still has jobs outside raw citation counts: rich results, entity clarity, and accurate brand representation. Measure those separately.
• Product and Review markup with filled-in attributes (price, rating, specs) is the one area where early evidence hints at a small edge, mostly for lower-authority domains.

Why this question is harder than it looks

Every few weeks someone posts a chart showing that cited pages have schema. Then someone else replies that schema does nothing. Both can be technically right, which is what makes this topic annoying.

The core problem is that sites with schema are usually better sites. They have developers who care, stronger content, more links and more authority. If those sites get cited more, you can't tell whether the markup did the work or just came along for the ride.

So the goal of this guide isn't to tell you whether schema "works" in general. It is to show you how to find out whether it works on your pages, with your content, in the AI systems that matter to you.

What the evidence says so far

Before you design a test, it helps to know what others have already found.

The correlation everyone quotes

Ahrefs looked at about 6 million URLs and found that AI-cited pages were almost 3x more likely to have JSON-LD than non-cited pages (Ahrefs study). Vendor posts then turned this into claims like "pages with schema are 36% more likely to appear in AI summaries" (WPRiders). That number comes with no primary source in the text, so treat it as unverified correlation.

The controlled test

The same Ahrefs team ran the experiment properly. They took 1,885 pages that added JSON-LD between August 2025 and March 2026, matched them against roughly 4,000 control pages, and used a difference-in-differences method.

Treated and control pages were both already declining in AI Overviews before schema was added. That detail matters. If you had only looked at your own pages, you might have blamed schema for a platform-wide swing.

There are caveats. Every page in the sample already had 100+ AI Overview citations, so the study says nothing about pages AI hasn't noticed yet. All schema types were pooled together, and 30 days may be too short to see entity-level effects.

What AI systems read at retrieval time

searchVIU built a test page with prices placed in visible text, JSON-LD, and hidden microdata or RDFa, then asked ChatGPT, Claude, Perplexity, Gemini and Google AI Mode for the prices. As summarized in the research, none of the systems used the hidden structured data during direct retrieval. Gemini did best, and it got there by rendering the page.

OtterlyAI ran a smaller, single-site experiment in March 2026 with wildly uneven results (ChatGPT -71%, AI Overviews +611%, Perplexity 0%). The authors themselves attribute the Google gains to traditional SEO mechanisms. It's one site, and the post quotes inconsistent figures, so I'd file it under anecdote.

A hint of an exception

A preprint by Kurt Fischman (730 citations from ChatGPT and Gemini, 75 commercial queries) found that Google rank position was the dominant predictor of citation. One exception: pages with Product or Review schema with populated attributes like price, aggregateRating and specs were cited more often than pages with generic schema (61.7% vs. 41.6%). The effect was strongest for lower-authority domains. It's a preprint, so hold it loosely, but it suggests the factual payload in the markup may matter more than the markup's presence.

What Google says

Google's guide to optimizing for generative AI features lists "overfocusing on structured data" as something to stop worrying about, says schema isn't required for AI search, and says there's no special markup for it. It still recommends structured data for rich results. Microsoft's Fabrice Canel said in 2025 that schema helps Microsoft's LLMs understand content, so Copilot and Bing-backed systems are the place where the question is genuinely open.

One more caveat on the data side: Promptwatch Data doesn't have a report on schema, so this is not something we've measured. What it does show is which page types get cited. In ChatGPT's citation share by content type (Sep 28 to Oct 4, 2026), product pages made up 22.4% of classified citations, landing pages 16.8% and listicles 13.6%. Page type and content quality are measured levers. Schema is not.

Step 0: Make sure AI can reach and read the page

Before you test anything, rule out the boring failures. They are more common than you'd expect.

1. Check robots.txt, CDN and WAF rules for every major AI crawler. Promptwatch's AI crawler traffic data (June 8 to Sep 6, 2026) shows OpenAI at 79.8% of verified crawler requests in the final week. If a provider that's large in that mix is near zero in your logs, it's probably blocked.
2. Confirm the answer is in the visible HTML. Markdown made up only 0.05% of AI search citations and HTML 99.94%, which tells you that rendered HTML is what gets cited. Heavy client-side JavaScript is a bigger risk than missing markup.
3. Make sure the schema matches the page. Markup that says something the visible content doesn't is a liability.

Tools like Promptwatch log IP-verified AI crawler visits per page, which is the quickest way to answer "are they even reading it?"

For a general crawl of your own markup, Screaming Frog, Sitebulb and Lumar can all extract and validate structured data at scale.

Step 1: Audit what schema you actually have

You can't evaluate an effect if you don't know where the markup is. Crawl the site and export every page with JSON-LD, the types used, and which properties are populated.

Then sort pages into groups:

• No schema at all
• Generic schema (Article, Organization, BreadcrumbList) with few properties
• Attribute-rich schema (Product with price, availability, aggregateRating, identifiers)
• Schema that doesn't match visible content (fix these first)

That last group will often explain weird results later. Validate with Google's Rich Results Test and the Schema.org validator.

Step 2: Build a baseline before you change anything

This is the step people skip, and it ruins the experiment.

Pick a set of prompts that matter to your business and track, per page, how often it's cited. Do this for at least four weeks. You want to see how noisy the numbers are before any change. AI citations swing a lot from week to week, and platforms change behavior without warning. Promptwatch's data on the citation drop around the GPT-5.3 rollout on March 4, 2026 is a good reminder: average citations per response fell across several models at once. If you'd added schema that week, you'd have blamed the wrong thing.

Track these separately:

• Citations per page, per platform (ChatGPT, AI Overviews, AI Mode, Perplexity, Copilot)
• Whether your domain is cited, versus merely mentioned
• Organic rankings and clicks, since rank is the strongest predictor of citation in the Fischman data
• AI crawler hits per page

Any AI visibility tracker can supply the citation side. Options in the catalog include Profound, Peec AI, Otterly.AI and Evertune. The thing to look for is page-level citation data, not just brand-level mention counts, because you need to tie changes to specific URLs.

If you want more options, the AI rank tracking directory and the GEO software directory list plenty more.

Step 3: Design a treated and control split

This is the part that turns an opinion into evidence.

1. Choose pages of similar type, similar traffic and similar current citation levels.
2. Split them in two. Add schema to one group (treated). Leave the other alone (control).
3. Keep everything else constant. No rewrites, no new internal links, no promotion on the treated group during the test.
4. Test one schema type at a time if you can. Pooling Article, FAQ and Product together is exactly what makes the Ahrefs result hard to interpret.

You don't need thousands of pages. Even 20 to 30 pages per group beats a before-and-after on a single URL. With small samples you'll only detect big effects, so be honest about that in your write-up.

What to use as control

Good control pages are the same template, similar age, similar topic. If you sell products, use matching product pages in the same category. For blog content, use posts from the same quarter on similar topics.

Step 4: Compare the change, not the level

The math here is simple. Don't compare treated pages to control pages directly. Compare how much each group changed.

In this example the treated pages lost citations, but fewer than the control pages. Schema looks like it helped by one citation per page. Without the control group you would have concluded schema hurt. This is exactly the trap in the AI Overviews data from the Ahrefs study.

A few reading rules:

• Run at least 4 to 8 weeks after the change. A 30-day window may be too short for entity effects, and ChatGPT results in OtterlyAI's test spiked and then corrected.
• Look per platform. An effect in AI Overviews and none in ChatGPT is plausible, since Google's features sit on top of Google's index.
• Ask whether the difference is bigger than week-to-week noise from your baseline period. If it isn't, call it no effect.
• Don't stop at the first bump. Early gains that fade are common.

Step 5: Measure the things schema is actually good at

Even if citations don't move, schema may still be worth having. Measure it on the dimensions where it has a plausible mechanism.

Rich results and SERP features

Google still recommends structured data for features that show up in results. Check Search Console's enhancement reports for eligibility and impressions on pages with rich results. Since rank drives AI citations, anything that helps a page win in Google can feed through indirectly. That's a traditional SEO win, and OtterlyAI itself attributes its Google gains to this.

Entity clarity and accuracy

This one is harder to quantify but easy to spot-check. Ask ChatGPT, Gemini, Claude and Perplexity who you are, what you sell, where you're based, and who runs the company. Log the answers before and after adding Organization, Person and sameAs markup. Look for wrong facts becoming right ones. Schema's value here is upstream, in indexing and entity resolution, which is a different question from whether a fetch tool reads JSON-LD on the fly.

Copilot and Bing

Because Microsoft has said schema helps its models and Copilot draws on Bing's index, test this platform separately. Pair it with Bing Webmaster Tools data.

Which schema types are worth testing first

Evidence is thin, so prioritize by plausibility and by cost.

One rule from the practitioner write-ups holds up: keep FAQ schema only where it matches real, visible questions and answers.

Tools that help with the measurement

You don't need all of these. Pick based on which part of the process is your bottleneck.

If you want the whole loop in one place, Promptwatch combines citation tracking, crawler logs, and Google and Bing search data per page, which makes the treated-versus-control comparison much easier to set up. Once you've found the pages that rank but don't get cited, its Action Board and content tools are where you'd go next.

Common mistakes

• Reading the "3x more likely" stat as proof. It's not.
• Testing a single page. One page tells you nothing.
• Changing content and schema at the same time. You won't know which one did it.
• Ignoring platform-wide swings. If your control pages dropped too, schema didn't cause your drop.
• Marking up things that aren't on the page.
• Testing on pages that are already cited heavily. The Ahrefs sample had this problem, so try your test on both well-cited and barely-cited pages.

What I'd do in practice

If your pages rank well and are already cited, adding generic schema probably won't change much, and the current data supports that. Keep the markup you need for rich results and move on to content and authority.

If you run an ecommerce or lower-authority site, test attribute-rich Product markup first. It's the one place where a small effect is plausible.

If your brand is misdescribed by AI assistants, test Organization, Person and sameAs, and judge success by whether the answers get more accurate, not by citation counts.

And if you're paying anyone to sell schema as a short-term GEO tactic, ask to see their treated-versus-control data. If you need help setting up measurement like this, an agency such as 1001 SEO Media can run technical SEO and AI visibility audits. For the DIY route, the table and steps above are enough to get a defensible answer in two months.

FAQ

Does schema directly make ChatGPT cite me?

The best controlled test so far found no measurable change in ChatGPT citations after adding JSON-LD (+2.2%, indistinguishable from zero). That test covered pages already well cited, so it can't speak to pages AI hasn't discovered.

Should I remove schema I already have?

No. Google still recommends structured data for rich results, and the downside of keeping accurate markup is negligible.

How long should I wait before judging a test?

Four to eight weeks after the change, against a baseline of at least four weeks before it. Longer if you're testing entity-level effects.

What matters more than schema?

In the Fischman preprint, Google rank position was the dominant predictor of AI citation. Relevance, authority and crawlable, clearly written visible content come first.