Last updated July 2026.
Schema markup sits at the loudest intersection in GEO research right now. Three studies looked at the same question, used different methods, and reached three different conclusions. Before you invest time in a schema sprint or dismiss structured data entirely, you need to understand what each study actually measured.
This piece covers the JSON-LD types most observed in AI-cited pages, the mechanism behind each, and the three conflicting data points you will encounter in every schema conversation in 2026.
The 7 JSON-LD types most associated with AI citations
The table below maps each schema type to the AI-citation mechanism it supports and the engines where practitioners and researchers observe it most. This is a reference view, not a ranked list of effectiveness.
| Schema type | What it signals to AI engines | Engines where it appears most | Primary use case |
|---|---|---|---|
| FAQPage | Structured Q&A pairs the engine can extract as a direct answer without paraphrasing | Google AI Overviews, Google AI Mode | Pages answering multiple related questions on one topic |
| HowTo | Step-by-step instruction format the engine can reproduce as a numbered answer | Google AI Overviews, Perplexity | Tutorials, process guides, how-to content |
| Article / NewsArticle | Author identity, publish date, topic, and editorial context in a parseable block | ChatGPT (browsing), Perplexity, Google AI Overviews | Blog posts, editorial features, industry explainers |
| Product | Name, description, price, availability, and review aggregate for a specific item | Google AI Overviews (commercial intent), Microsoft Copilot | Product detail pages, comparison pages |
| Review / AggregateRating | Third-party credibility signals via star ratings and review counts | Google AI Overviews, Microsoft Copilot | Product reviews, service reviews, software comparison pages |
| BreadcrumbList | Site hierarchy context that helps engines understand where a page sits relative to the domain | Google AI Overviews, Gemini | Every page type; supports topical authority signals |
| SpeakableSpecification | Marks specific passages as optimised for voice and AI reading, signalling which sentences carry the core answer | Google AI Overviews, voice assistants | Feature-deep-dive pages, answer-first content |
Why the mechanism matters more than the markup type
AI engines do not cite pages because they have schema. They cite pages because the engine decided the content best answers the user’s query. Schema helps an engine confirm what a page is about and extract the relevant passage faster. That is a parsing assist, not a ranking boost.
The practical implication: schema helps most when the content underneath it is already the best available answer to a specific question. Schema applied to weak, generic, or poorly structured content does not transform that content into a citable source.
FAQPage schema works because it pre-formats the content as a list of questions and direct answers. Google’s AI systems can extract individual Q&A pairs from a FAQPage block and surface them in AI Overviews without rewriting anything. That extraction efficiency is the mechanism, not a special preference for FAQPage content.
HowTo schema works by the same principle: a numbered step-by-step format lets the engine reproduce the instruction sequence in the generated response. The more precisely the HowTo markup matches the prose structure of the page, the easier the extraction.
Article and NewsArticle schema matter in contexts where the engine needs to confirm the source’s credibility: who wrote it, when it was published, and what organisation produced it. These signals matter more for ChatGPT in browsing mode and for Perplexity, both of which weight freshness and author attribution more heavily than Google AI Overviews does.
Product and Review schema dominate in commercial-intent AI Overviews because Google’s AI systems treat those pages as candidates for direct product recommendations. The AggregateRating block (star rating plus review count) provides a credibility shortcut the engine uses when generating comparison answers.
What the studies actually found
Three data points define the current debate. Each is real. Each has important scope limitations. Here they are, side by side.
1. The Ahrefs controlled study: no meaningful uplift (2026)
Ahrefs tracked 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matching them against 4,000 control pages. The result: adding schema produced no statistically meaningful uplift in AI citations across Google AI Overviews, Google AI Mode, or ChatGPT.
The specific numbers: Google AI Overviews showed a small decline (around -4.6%). Google AI Mode showed +2.4%. ChatGPT showed +2.2%. Both positive figures were statistically indistinguishable from zero.
The critical scope limitation: every page in the Ahrefs dataset already had 100 or more AI Overview citations before the experiment began. The study cannot tell you whether schema helps pages with zero citations. It tells you that adding schema to already-cited pages does not reliably increase citation volume.
Source: Ahrefs, “We Tracked 1885 Pages Adding Schema. AI Citations Didn’t Move.” Louise Linehan and Xibeijia Guan, May 2026 (ahrefs.com/blog/schema-ai-citations/).
2. The Kumar and Palkhouski arXiv preprint: +39% lift (2025)
A UC Berkeley preprint by Arlen Kumar and Leanid Palkhouski (arXiv:2509.10762, September 2025) audited 1,100 URLs across 1,702 citations drawn from Brave, Google AI Overviews, and Perplexity. The study ranked structured data as the third-strongest predictor of AI citation likelihood, associated with a +39% lift.
The two stronger predictors in that preprint: metadata and freshness (+47%) and semantic HTML structure (+42%).
The critical scope limitation: this is a correlation study, not a controlled experiment. It identifies structured data as a feature present more often on cited pages, not a feature that causes citation gains when added. An observational study and a controlled addition study measure different things.
Source: Kumar & Palkhouski, “AI Answer Engine Citation Behavior: An Empirical Analysis of the GEO16 Framework,” arXiv:2509.10762, September 2025.
3. The Otterly.AI sitewide experiment: +1,500% appearances, then a retraction (2025-2026)
After rolling out five schema types across their own website between December 2025 and March 2026, Otterly.AI recorded a +1,500% increase in Google AI Overviews appearances over three months.
Their own conclusion: the lift was not caused by schema. Competitors who made no schema changes saw equivalent movement over the same period. Otterly.AI attributed the spike to a broader algorithmic shift at Google, not to the schema rollout. The increase was also Google-specific: citations fell on ChatGPT, Gemini, and Copilot during the same window.
The metric tracked here was AI Overviews SERP appearances, not citations in the traditional sense. That distinction matters when comparing this result to the Ahrefs citation study.
Source: Otterly.AI, “Schema Markup’s Real Impact on AI Search” (otterly.ai/blog/schema-markup-real-impact-ai-search/). Thomas Peham also presented these findings at BrightonSEO, April 2026.
What this means for your schema strategy
The research does not support abandoning schema. It supports being precise about what you expect schema to do.
Schema as a baseline signal. Every page that could plausibly appear in an AI Overview should have correctly implemented schema for its content type. Missing schema is a friction point AI engines must work around. That friction is unnecessary.
Schema as a parsing assist, not a citation lever. Do not expect that adding FAQPage markup to a page that currently receives zero AI citations will suddenly earn you citations. The content itself must be the best available answer first.
The mechanism that actually works. The Ahrefs data and the Kumar and Palkhouski correlation both point to the same underlying principle: schema helps when the content structure it describes is genuinely clear, direct, and question-answering in format. A FAQPage block wrapped around vague prose answers does not help. A FAQPage block that accurately describes four crisp Q&A pairs on a specific topic gives the engine exactly what it needs to extract an answer.
The front-of-page signal. A separate Kevin Indig analysis of 1.2 million ChatGPT responses found that 44.2% of ChatGPT citations were drawn from the first 30% of a page’s content. That finding is specific to ChatGPT, not to AI citations broadly, and should be attributed as such. But it reinforces a consistent pattern across multiple studies: the answer-first format, where the direct response to the query appears near the top of the page, outperforms formats that build to a conclusion. Schema alone does not fix a page structure problem.
Tools that handle schema for AEO
If your bottleneck is generating and validating schema at scale, these are the tools practitioners in this category use.
Otto SEO audits existing pages for schema coverage gaps and generates JSON-LD blocks across site sections. It is built for teams running schema sprints across large page inventories.
Surfer includes schema validation in its Content Editor workflow. If you are writing and optimising content in the same session, Surfer lets you confirm schema is in place before you publish.
HubSpot AEO Grader scores pages for AEO readiness, including structured data. It is a useful diagnostic if you want a fast audit of where schema is missing or malformed across a domain.
Otterly.AI is worth mentioning precisely because it ran the most detailed public schema experiment in this category and published the results including the part where schema did not explain the outcome. The Otterly.AI platform tracks AI Overviews appearances over time, which makes it a practical tool for measuring any schema experiment you run on your own site.
Temso covers the full AEO loop from schema generation through citation monitoring. For teams that want to connect schema implementation to citation tracking without switching tools, it is one of the few platforms that handles both ends of the workflow at a single flat price.
A comparison of these and other AEO platforms is at /rankings/aeo-tools. Schema-specific scoring criteria are explained at /methodology.
The practical checklist
Before you run a schema sprint expecting citation gains, confirm:
- Your content already answers the target query more directly than the current top-cited pages. Schema applied to weaker content does not outperform stronger content without schema.
- Your schema markup accurately describes the content structure on the page. A FAQPage block with questions that do not match the prose Q&A below it is a mismatch, not a signal.
- You have a measurement baseline. The Otterly.AI experiment is a useful model: track AI Overviews appearances before, during, and after the rollout, across at least two competing domains that are not making schema changes, so you can distinguish a schema effect from an algorithmic one.
- You are validating with Google’s Rich Results Test before publishing. Malformed JSON-LD is common enough that validation is not optional.
- You are not treating schema as a substitute for answer-first content structure. The front-of-page placement signal from the Indig ChatGPT analysis, the FAQPage parsing mechanism, and the Ahrefs finding on already-cited pages all point in the same direction: the content underneath the markup is the variable that moves AI citation rates.
The /glossary has definitions for each schema type mentioned here if you need precise reference language for a team brief or stakeholder document.
Schema markup is necessary infrastructure for AEO. It is not sufficient on its own, and the research does not currently support a clean causal story between schema addition and citation gains. Build the markup correctly, build the content correctly, and measure before and after with a comparison group. That is the honest version of a schema strategy in 2026.
Want to see which AI Overviews you are losing right now? Run a citation audit at /rankings/aeo-tools to find tools that track your schema coverage and AI citation rate across Google AI Overviews, ChatGPT, Perplexity, and other engines simultaneously.