# Does Schema Help AI Citations? The Evidence Is Mixed

> Ahrefs found no lift from adding schema; OtterlyAI saw Google gains on one brand. What each test shows, and when schema is worth your time.

URL: https://llmagnet.com/blog/schema-markup-ai-citations  
Author: Ben Cohen, Founder & CEO at LLMagnet  
Published: 2026-06-16  
Updated: 2026-11-10  
Category: Research

Schema has not been shown to lift AI citations as a general rule. In a controlled test by Ahrefs, there was no major uplift on any platform after 1,885 pages added JSON-LD, while OtterlyAI’s test on its own site saw gains on Google and losses on three other engines. Add schema that matches real page content, then measure it on your own pages before expecting citations.

[Google’s documentation](https://developers.google.com/search/docs/appearance/ai-features) says you don’t need any special schema.org markup to appear in AI Overviews or AI Mode. It still recommends that existing structured data match the visible text on the page.

## Contents

1. What the Ahrefs Test Found
2. What OtterlyAI’s Test Found
3. What Correlation Studies Show
4. Where Schema Could Matter
5. Why the Results Differ
6. Which Sites Should Spend Time on This
7. What to Do This Week
8. What Populated Schema Looks Like
9. Common Mistakes
10. How to Measure Whether Schema Is Working
11. FAQ

## What the Ahrefs Test Found

Ahrefs, an SEO software vendor, published its test on May 11, 2026. It tracked 1,885 pages that added JSON-LD between August 2025 and March 2026, against 4,000 control pages that never did, and compared citations in the 30 days before and after each change.

- **Google AI Overviews:** -4.6%, small but statistically significant. Both groups were already declining and the cause is unknown.
- **Google AI Mode:** +2.4%, statistically indistinguishable from zero.
- **ChatGPT:** +2.2%, statistically indistinguishable from zero.

![Change against control pages over 30 days for 1,885 pages that added JSON-LD: AI Mode plus 2.4 percent and ChatGPT plus 2.2 percent, neither statistically significant, and AI Overviews minus 4.6 percent.](https://llmagnet.com/images/blog/charts/schema-ahrefs-test.png)

*Source: Ahrefs, May 2026. Vendor data. The tested pages were already heavily cited.*

The method is worth understanding, because it’s how to read any schema claim. Ahrefs used a difference-in-differences comparison: how much the treated pages changed, minus how much similar pages that never added schema changed. AI Mode shows why the second half matters. Treated pages rose 43% in raw terms, but the controls rose almost as much, which leaves the 2.4%. A plain before-and-after chart would have looked like a 43% win.

In an initial scan of about 6 million URLs, cited pages were nearly three times as likely as non-cited pages to have JSON-LD. That is the kind of gap correlation studies report, and it didn’t turn into a lift when pages actually added the markup.

Ahrefs pooled Article, FAQ, Product, HowTo and Organization schema, so its test can’t say whether one type works better than another. Every page already had 100 or more AI Overview citations, so it can’t say whether schema helps pages engines aren’t yet retrieving. The window was 30 days, and many pages changed other things at the same time.

The test also covered JSON-LD in static HTML only, so it says nothing about markup injected by JavaScript.

Ahrefs also reports a searchVIU experiment in which ChatGPT, Claude, Perplexity, Gemini and Google AI Mode all ignored schema when fetching a page live and extracted only visible HTML.

## What OtterlyAI’s Test Found

[OtterlyAI’s experiment](https://otterly.ai/blog/schema-markup-real-impact-ai-search/) (vendor data) covered one SaaS brand: its own. It added FAQ, Organization, SoftwareApplication, Review Snippets and Article schema on December 7, 2025, and tracked 319 US prompts until March 7, 2026. OtterlyAI says the results may not apply to other verticals.

On the three studied pages, brand coverage changed as follows: AI Overviews +611%, AI Mode +42%, ChatGPT -71%, Copilot -64%, Gemini -35%, and Perplexity 0%. These are large percentages on a small sample with no control pages.

A [brightonSEO deck](https://speakerdeck.com/thomaspeham/geo-experiments-2026-what-we-tested-what-failed-and-what-actually-works) from April 2026 says a sitewide rollout “correlated with” +1,500% in AI Overviews and +377% in AI Mode citations. OtterlyAI’s article repeats the 1,500% but labels the 377% as SERP features, and neither gives a baseline, so we can’t reconcile those figures with the per-page table.

![Change in brand coverage on three OtterlyAI pages, December 2025 to March 2026: AI Overviews plus 611 percent, AI Mode plus 42 percent, Perplexity 0 percent, Gemini minus 35 percent, Copilot minus 64 percent, ChatGPT minus 71 percent.](https://llmagnet.com/images/blog/charts/schema-otterly-engines.png)

*Source: OtterlyAI, March 2026. Vendor data. One brand, no control pages.*

Two details weaken a schema explanation. Site-wide coverage rose from roughly 27% to the low 30s, but competitors with no schema change moved as much or more, and OtterlyAI attributes that lift to an algorithm change. A fact placed only in FAQ schema was not surfaced by any engine, although that observation window was only several days.

Google’s two surfaces also stayed below their August 2025 levels, and OtterlyAI credits the AI Overviews gain to traditional SEO and search-feature eligibility, not to a GEO effect. Its own conclusion is to keep schema as an SEO practice, not as a way to grow AI visibility.

## What Correlation Studies Show

These studies compare pages that have schema with pages that don’t. They can’t show that schema caused a citation.

- **[AirOps with Kevin Indig](https://www.airops.com/report/the-fan-out-effect-what-happens-between-a-query-and-a-citation)** (vendor data, 16,851 queries in ChatGPT, published April 2026): pages with JSON-LD were cited 38.5% of the time against 32.0% without. The two groups were similar on word count, headings, domain authority and query match.
- **[Kurt Fischman’s paper](https://machinelike.ai/research/does-schema-markup-predict-ai-citation)** (730 citations from ChatGPT and Gemini across 75 commercial queries, published February 2026): having schema at all showed no significant effect (p = .296). Google rank position was the dominant predictor. Within schema-bearing pages, Product or Review schema with populated fields was cited at 61.7% against 41.6% for generic types (p = .012), most strongly for domains rated DR 60 or below.
- **[SE Ranking’s AI Mode study](https://seranking.com/blog/how-to-optimize-for-ai-mode/)** (vendor data, 2,328,533 pages): FAQ schema showed no meaningful effect, while real FAQ sections in the content averaged 4.9 citations against 4.4.
- **[SE Ranking’s ChatGPT study](https://seranking.com/blog/topical-authority-study)** (vendor data, 216,524 pages): pages with FAQ schema averaged 3.6 citations against 4.2 without.
- **[Search Atlas](https://searchatlas.com/research/the-limits-of-schema-markup-for-ai-search/)** (vendor data, published December 2025, sample size not reported): domain visibility showed no consistent trend as schema coverage increased.

![Citation rate within schema-bearing pages in Kurt Fischman’s paper: Product or Review schema with populated fields 61.7 percent, generic schema types 41.6 percent.](https://llmagnet.com/images/blog/charts/schema-fischman-types.png)

*Source: Kurt Fischman, February 2026, a preprint. 730 citations across 75 commercial queries. p = .012.*

Fischman’s paper shows how easily this kind of analysis can get the sign wrong. His first pooled model suggested schema cut the odds of citation (odds ratio 0.546, p < .001). He traced that to a sampling artifact: Google’s top 10 already over-represents pages with schema. Within Google’s results, schema appeared on 43.1% of cited pages and 44.8% of non-cited pages, which is effectively the same. Rank did the work, with position 1 pages cited in 43% of the queries where they appeared and position 7 pages in 5%.

AirOps’ figures need similar care. Only 1.0% of the pages it retrieved had JSON-LD, so its 13,341 schema rows are compared with 735,542 without. It reports type-level rates such as FAQPage at 45.6% and Organization at 44.3%, but they aren’t controlled for retrieval rank, which it names as the strongest predictor of citation. Don’t choose schema types from that list.

## Where Schema Could Matter

An engine handles a page in stages, and schema could help at some and not others. Separating them explains why “does schema work?” has no single answer.

![Four stages of an AI answer and the evidence for schema at each: crawl and index, untested; retrieval, a correlation in ChatGPT from AirOps but no test; live page fetch, searchVIU found engines read visible HTML only; answer, a fact placed only in FAQ schema was not surfaced in OtterlyAI’s test.](https://llmagnet.com/images/blog/charts/schema-pipeline.png)

*Our own summary of the studies linked in this article.*

**Indexing is the open question.** Ahrefs notes that schema might still help pages AI systems haven’t seen yet, during crawling or indexing, and its test can’t see that. AirOps suggests structured data may help retrieval systems categorize a page, but doesn’t test it.

**Live fetching has the weakest case.** OtterlyAI asked seven platforms to fetch and report a page’s schema. Gemini returned the correct JSON-LD. ChatGPT, Claude, Perplexity and Copilot failed. AI Mode listed types but invented a Service schema that wasn’t on the page. OtterlyAI attributes part of this to HTML-to-Markdown pipelines that strip script tags. Treat this as a weak signal, and as a warning: asking a chatbot whether it can see your markup won’t tell you much.

## Why the Results Differ

The studies don’t measure the same thing, so they don’t cancel out.

- **Design.** Ahrefs compared pages over time against controls. The correlation studies compare different pages at one point in time.
- **Pages.** Ahrefs used heavily cited pages, OtterlyAI used a few pages on one site, and the correlation studies use broad samples.
- **Schema type.** Ahrefs pooled types. Fischman’s paper separated attribute-rich from generic types, on a much smaller sample.
- **Platform.** Results vary by engine, and in opposite directions in OtterlyAI’s test.

The fairest summary is that adding schema to pages already doing well didn’t produce a citation lift in Ahrefs’ controlled test. Whether it helps pages not yet cited, or specific types on specific engines, remains open.

## Which Sites Should Spend Time on This

These recommendations are our judgment from the studies above.

| Your site | What to do | Why |
| --- | --- | --- |
| **Ecommerce with real prices and ratings** | Fill in Product, Offer and AggregateRating fully | Best correlational support, and useful for product data anyway |
| **B2B SaaS or services** | Let your CMS handle Organization and Article, and add SoftwareApplication only with real fields | OtterlyAI’s SaaS rollout didn’t beat competitors that changed nothing |
| **Domain rated DR 60 or below** | Start with attribute-rich types on commercial pages | Fischman’s advantage was largest here, on a small sample |
| **Already heavily cited pages** | Don’t expect a lift, and spend time on content first | Ahrefs’ test covered exactly this group |

## What to Do This Week

Google [stopped showing FAQ rich results](https://developers.google.com/search/updates) on May 7, 2026, so the old search-appearance reason for FAQ markup has gone. Given the evidence, this order is our judgment:

1. **Write real Q&A into your pages.** SE Ranking’s data favors visible FAQ sections over FAQ markup. Answer questions your buyers actually ask.
2. **Fill in Product or Review schema on commercial pages** where you have real prices, ratings and specifications. The evidence is correlational, so expect a small and uncertain benefit.
3. **Keep Article schema accurate,** with a named author and a correct `dateModified`. This is basic hygiene, and no study here shows it lifts citations.
4. **Leave generic Organization and WebPage schema to your CMS or SEO plugin.** It isn’t worth manual effort for AI citations.
5. **Make sure markup matches the visible page.** Google asks for this, and the searchVIU result suggests visible text is what several engines read when they fetch a page.

## What Populated Schema Looks Like

In Fischman’s paper, “populated” meant fields such as pricing, aggregate rating and specifications. The product and figures below are made up for illustration.

```
{
  "@context": "https://schema.org",
  "@type": "Product",
  "name": "Quillon Payroll Plus",
  "offers": {
    "@type": "Offer",
    "price": "49.00",
    "priceCurrency": "USD"
  },
  "aggregateRating": {
    "@type": "AggregateRating",
    "ratingValue": "4.6",
    "reviewCount": "212"
  }
}
```

Every one of those values should also be visible on the page. A block that only says `"@type": "Organization"` with a name and a URL is what the paper counts as generic.

## Common Mistakes

**Rolling out sitewide with no control pages.** OtterlyAI did this and could only conclude that competitors moved as much as it did. Keep some similar pages unchanged.

**Reading a raw before-and-after.** AI Mode’s +43% in Ahrefs’ raw data is the example. Engines’ overall citation behavior shifts constantly, and your schema change gets the credit by accident.

**Hiding facts in markup.** A fact placed only in FAQ schema wasn’t surfaced by any engine in OtterlyAI’s short test, and Google asks that markup match visible text. Put the fact on the page.

**Relying on JavaScript to add JSON-LD.** Ahrefs only tested static HTML, so there’s no evidence that injected markup behaves the same way. Check the page source, not just the rendered DOM.

## How to Measure Whether Schema Is Working

Search Console’s [generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports), announced June 3, 2026 and [available to all sites](https://support.google.com/webmasters/answer/16984139?hl=en) since August 31, 2026, show impressions in AI Overviews and AI Mode by page, country, device and date. They report impressions only.

1. Choose 10 to 15 pages to change, and a similar set to leave alone as a control.
2. Record both groups’ AI impressions for 30 days.
3. Add the schema to the first group only, and change nothing else on those pages.
4. Compare both groups after 60 to 90 days. Ahrefs noted that its 30-day window might miss slower effects.

Ahrefs suggests a smaller version: 5 to 10 treated pages, 5 to 10 similar controls and a 30-day comparison. If both groups move together, the platform trend is the likely driver. The arithmetic is simple. The pages and figures below are made up for illustration.

| Group | AI impressions before | After 60 days | Change |
| --- | --- | --- | --- |
| 12 pages with schema added | 40,000 | 47,200 | +18% |
| 12 similar pages unchanged | 38,000 | 43,700 | +15% |

The estimated effect is 18% minus 15%, or 3 points. If these pages normally swing by 5 points from month to month, 3 points is noise, and you shouldn’t roll the change out on this evidence.

This covers Google only. For other engines you need to run your prompts directly. LLMagnet measures AI visibility across seven engines and has an agent that prepares fixes for approval. You can start with the [free AI visibility audit](https://ai-visibility.llmagnet.com).

## FAQ

### Does schema markup increase AI citations?

It hasn’t been shown to as a general rule. Ahrefs found no major uplift after 1,885 already-cited pages added JSON-LD. Smaller or correlational studies show mixed results that depend on schema type and platform.

### What did the Ahrefs schema study find?

Against matched controls over 30 days, AI Overview citations fell 4.6%, while AI Mode rose 2.4% and ChatGPT 2.2%, both indistinguishable from zero. It pooled schema types and covered only pages already heavily cited.

### What did OtterlyAI’s schema test find?

On three pages of its own site, OtterlyAI saw AI Overviews rise 611% and AI Mode 42%, while ChatGPT fell 71% and Copilot 64%. It is one brand with no control pages, and a separate deck reports larger figures without a baseline.

### Which schema types are worth adding?

Product or Review schema with real, populated fields has the best correlational support, in a 730-citation paper. Visible FAQ sections beat FAQ markup in SE Ranking’s data. None of these studies proves that a type lifts citations.

### Does schema help on ChatGPT and Perplexity?

Ahrefs found no significant effect on ChatGPT and didn’t test Perplexity. AirOps found a correlation in ChatGPT. In OtterlyAI’s single-site test, brand coverage fell on ChatGPT and didn’t change on Perplexity.

### How can I test whether schema works on my site?

In Search Console, compare AI impressions for 10 to 15 pages that get schema against a similar set that doesn’t, over 60 to 90 days. Change nothing else on those pages. This measures Google only.