Schema Markup vs Plain HTML: What Actually Moves AI Citation Rates

Schema markup alone does not reliably increase AI citation rates, a 2026 Ahrefs study of 1,885 pages found no statistically significant lift on ChatGPT, Google AI Mode, or Perplexity. Clean semantic HTML, answer-first structure, and topical authority drive citations. Schema's real job is machine-readability and rich results, not direct citation uplift.

The internet is full of posts telling you to "add FAQ schema and watch your AI citations explode." The data doesn't fully support that narrative. That doesn't mean schema is worthless, it means the *mechanism* is different from what most GEO content claims, and if you're deploying engineering hours based on hype, you're likely solving the wrong problem.

This piece breaks down what the 2025-2026 research actually shows, where schema genuinely earns its keep, and how plain HTML beats structured data in several citation-driving scenarios.

What the Research Actually Says

The most rigorous study to date on this question was published in early 2026: Ahrefs tracked 1,885 web pages that added JSON-LD schema markup between August 2025 and March 2026, compared against 4,000 control pages. The results were sobering. Google AI Overviews showed a 4.6% citation decline (the only result that cleared the significance bar). AI Mode showed +2.4% and ChatGPT +2.2%, both well within noise.

That's the controlled, longitudinal take. On the observational side, the picture is messier:

• An AirOps analysis of 16,851 queries found pages with JSON-LD had a 38.5% citation rate vs. 32.0% without, a 6.5 percentage point advantage.
• A UC Berkeley GEO-16 study (Kumar & Palkhouski, arXiv 2025) analyzing 1,702 citations across three AI engines found structured data associated with a +39% citation impact, ranking behind metadata freshness (+47%) and semantic HTML (+42%) in multivariate analysis, though the researchers note these are association strengths, not clean causal lifts, since technically stronger pages also tend to implement schema.
• OtterlyAI ran a sitewide schema deployment across 2,000+ URLs and reported Google AI Overviews citations jumping 1,500%, but their own post-analysis concluded the December 2025 lift was algorithmic, not schema-driven, because competitors who hadn't implemented schema saw equivalent coverage gains simultaneously.

The honest synthesis: correlation is real but causation is weak. Pages with schema tend to be better-maintained, more authoritative, and more technically sound overall, making it hard to isolate schema as the independent variable.

For a broader look at which signals move citations across platforms, see our analysis of 21 sources AI engines consistently cite.

Why Plain HTML Pulls More Weight Than You'd Expect

The finding that semantic HTML correlates with higher citation lift than schema alone, seen across multiple observational studies, is the most underappreciated result in this debate. Here's why it makes sense.

AI engines don't read your JSON-LD the way a browser reads CSS. Most LLM-based systems were trained on raw HTML and text. They learned to associate heading hierarchies, paragraph structures, and list formatting with answer-worthy content long before structured data was a consideration. When a model encounters a clean <h2> question followed by a crisp two-sentence answer, it can extract that Q&A pair directly. When it encounters a wall of narrative prose, even if decorated with perfect FAQPage schema, extraction is harder.

The implication: structure visible in the HTML is often more directly actionable than metadata attached to it.

Multiple 2025-2026 studies consistently find that proper H2 hierarchy is among the strongest independent citation predictors. Pages with question-phrased headings (containing a ?) are cited roughly 2x more frequently, with the heading text itself often pulled verbatim, a pattern replicated across Semrush's AI Overviews research, the GEO-16 study, and AirOps' citation corpus.

Plain HTML wins on:

• Training-data alignment, most LLMs were trained on HTML, not schema
• Direct answer extraction, clean paragraphs under descriptive headers get pulled verbatim
• Universal compatibility, works across all AI platforms, not just Google's ecosystem
• Speed, no implementation overhead, no risk of schema validation errors

Where Schema Markup Still Earns Its Keep

Schema is not snake oil. It's just not the citation lever most GEO content claims it is. Here's where it legitimately moves the needle:

1. Google's Ecosystem (AI Overviews, AI Mode, Search) Google is the one platform where structured data demonstrably improves discoverability. The 2025 Wellows analysis found pages with proper schema markup had +73% higher selection rates in AI Overviews compared to unstructured pages. FAQPage schema in particular aids AI extraction of your Q&A content, even though FAQPage no longer generates visual rich results in Google Search (removed May 7 2026), the markup still helps Google's systems understand and surface your answers in AI Overviews.

2. Rich Results and SERP Real Estate Schema's most reliable payoff remains traditional: star ratings, product price snippets, breadcrumbs. (Note: FAQPage and HowTo rich results no longer generate visual SERP enhancements, HowTo was removed in 2023, FAQPage fully removed May 7 2026, but these schema types still aid AI engine extraction and are worth maintaining.) Rich results drive organic traffic and engagement signals that make a page *citation-worthy* in the first place.

3. Entity Disambiguation Organization schema, Person schema, and SameAs properties help knowledge graphs correctly identify your brand entity. This matters for AI systems that use entity graphs to resolve brand references, a consistent entity profile reduces the chance an AI confuses you with a competitor or misattributes a quote.

4. FAQPage on High-Intent Pages The AirOps data showed FAQPage schema at a 45.6% citation rate, second only to BreadcrumbList (46.2%). On pages targeting informational queries where a searcher's question is precisely answered by your FAQ block, the schema reinforces what the content already does well.

For a platform-by-platform breakdown of how AI engines treat different content signals, see how AI engines cite brands differently.

The Citation-Rate Comparison Table

*Citation lift by technical factor type, controlled vs. observational studies show different schema effects. Freshness + semantic HTML outperform schema as isolated variables.*

Head-to-Head: Schema Markup vs. Clean HTML

The verdict: For AI citation rates *specifically*, clean semantic HTML with answer-first structure outperforms schema as a direct lever. Schema remains essential for Google SEO infrastructure, rich results, and entity clarity, but treating it as a GEO silver bullet sets unrealistic expectations.

The Platform Divergence Problem

One reason schema's citation story is so muddled: different AI platforms have fundamentally different relationships with structured data.

Google AI Overviews runs on a system deeply integrated with Google's Knowledge Graph and indexing pipeline. Schema feeds that pipeline. Citations correlate with proper markup because Google *built* its understanding of the web partly *around* Schema.org.

ChatGPT and Perplexity retrieve content via web fetch at query time. They're pulling live HTML and running it through their own extraction layers. JSON-LD in a <script> tag is often parsed separately from the visible content, and some studies show it's partially ignored. What those engines *do* prioritize is whether the fetched page answers the question clearly in the first 150 words of visible text.

Gemini sits between these poles, stronger schema awareness than ChatGPT but weaker than Google Search's native stack.

The practical implication: if your goal is cross-platform citation coverage, HTML structure is the more universal investment. Schema is a Google-first optimization that delivers diminishing returns elsewhere.

For a platform-by-platform breakdown of citation signals in Q1 2026, see our State of LLM Search Q1 2026 report.

The Right Framework: Both, But in the Right Order

The question isn't *which* to implement, it's *which to prioritize when resources are constrained* and *why*.

Do HTML structure first:

1. Write H1 → H2 → H3 hierarchy that matches how a human would ask questions about the topic
2. Lead every section with a direct answer in the first 1-3 sentences (inverted pyramid)
3. Use bullet lists and short paragraphs, these get extracted verbatim by AI engines
4. Make the title tag and meta description precise, query-matching, and under 60/160 characters respectively

Then layer schema:

1. Article / BlogPosting on all content (minimal lift, no downside)
2. FAQPage on pages with explicit Q&A blocks
3. Organization + SameAs on your homepage and About page (entity graph anchoring)
4. BreadcrumbList sitewide (strong observational citation correlation)
5. Product schema for e-commerce (enables rich results that drive the traffic that drives citation)

Skipping to schema before fixing heading structure and answer density is a common mistake. Schema describes content. It doesn't improve content. An AI engine won't cite a poorly-structured page more often because you added JSON-LD, it will cite a well-structured page with *or without* schema.

Use our free GEO audit tool to see your current citation structure score before choosing where to invest.

Visualizing the Decision Path

*GEO technical priority: HTML structure and content quality are prerequisite investments before schema delivers reliable returns.*

Who Should Do What

Prioritize HTML structure if:

• You're optimizing for ChatGPT, Perplexity, or cross-platform citation coverage
• Your pages are narrative-heavy with poor heading hierarchy
• You have limited dev resources and need the highest ROI per hour
• You're in an early stage of GEO and haven't audited your content structure yet

Prioritize schema if:

• Google Search and Google AI Overviews are your primary citation targets
• You have e-commerce pages that need rich result eligibility
• Your brand entity is ambiguous or confused in AI knowledge graphs
• You already have clean HTML structure and want the incremental Google boost

Do both if:

• You're a publisher or SaaS brand with editorial volume (schema is infrastructure cost worth paying at scale)
• You're running a free GEO audit and seeing both structure and schema gaps
• You want to maximize FAQPage schema for AI extraction *and* cross-platform citation coverage

Understanding how AI engines cite brands differently across platforms is the prerequisite to knowing which investment applies to your distribution.

Glossary Quick-Reference

• JSON-LD: JavaScript Object Notation for Linked Data, the recommended schema injection method (in a <script type="application/ld+json"> tag)
• FAQPage schema: Structured data type that marks up Q&A pairs; previously generated rich results in Google Search (removed May 7 2026), but still aids AI engine extraction of Q&A content
• Semantic HTML: HTML that uses meaningful elements (<article>, <h1>-<h6>, <nav>, <section>) to communicate document structure to machines
• Citation rate: The percentage of AI-generated responses to a monitored prompt set that include a link or reference to a given page or domain
• GEO (Generative Engine Optimization): The practice of optimizing content and technical signals to increase citation frequency and accuracy in AI-generated answers

For definitions of more GEO-specific terms, see the Promptive Glossary.

Frequently Asked Questions

Does adding schema markup directly increase my AI citation rate?

In controlled studies, no. A 2026 Ahrefs experiment tracking 1,885 pages found no statistically significant citation increase on ChatGPT (+2.2%) or Google AI Mode (+2.4%) after adding JSON-LD schema, both results were indistinguishable from noise. Schema supports Google's indexing infrastructure and enables rich results, but does not appear to directly signal citation-worthiness to AI answer engines.

Which schema types have the highest observed correlation with AI citations?

Observational research from AirOps (16,851 queries) found BreadcrumbList at 46.2% citation rate, FAQPage at 45.6%, and Organization at 44.3%, versus 32.0% for pages with no schema. However, these are correlation numbers, not causal lift figures; pages with schema tend to be technically stronger overall.

Is plain HTML better than schema for Perplexity and ChatGPT citations?

Yes, based on current evidence. Both platforms retrieve live HTML at query time and extract visible content. Clear heading hierarchy, answer-first paragraph structure, and short well-labeled sections drive direct extraction. JSON-LD in a <script> tag has minimal documented impact on ChatGPT and zero on Perplexity in the 2026 OtterlyAI study.

Did Google remove any schema rich results recently?

Yes. Google removed HowTo rich results in September 2023, they no longer generate visual SERP enhancements. FAQ rich results were restricted to government and health sites in August 2023, then fully removed from Google Search on May 7 2026. Critically, both FAQPage and HowTo remain valid schema.org types that Google still parses and that help AI engines extract and cite your content, only the visual SERP presentation is gone. Keep FAQPage and Article/BlogPosting schema; they still earn their keep for AI answer-engine extraction.

How do I know if my schema is hurting rather than helping?

Malformed or conflicting schema can confuse crawlers. The clearest signal is a drop in rich result impressions in Google Search Console. Use Google's Rich Results Test to validate. In the Ahrefs controlled study, adding schema was associated with a statistically significant 4.6% *decline* in AI Overviews citations, the mechanism isn't confirmed, but poor implementation quality is a plausible contributor.

How does Promptive measure schema's effect on AI citation rates?

Promptive monitors AI citations at the prompt level across ChatGPT, Claude, Gemini, Perplexity, Google AI Overviews, and ChatGPT Shopping. You can track citation rate before and after a schema deployment as a natural experiment, comparing your brand's share of voice against a control set of monitored competitors over the same window. Run a free GEO audit to establish your baseline.