Research

The Schema Myth: What AI Platforms Actually Use to Decide What to Cite

Add schema markup. Add FAQ structured data. Tell AI what your page is about. The advice is everywhere. The evidence is not. Two controlled experiments now suggest it is largely wrong.

By

TL;DR

  • LLMs tokenize JSON-LD as plain text. The structured relationships encoded in schema are not visible to the model — it reads characters, not hierarchy.
  • In a controlled experiment, a fictional address embedded in invalid, made-up schema was returned accurately by ChatGPT and Perplexity — because they read the script block as raw text, not structured data.
  • Search Atlas studied schema coverage vs. citation rates across three AI platforms and found no meaningful difference in performance.
  • Schema accounts for roughly 10% of citation factor weighting. Content quality and domain authority account for ~70%.
  • One genuine exception: Microsoft’s Fabrice Canel confirmed on record that schema helps Microsoft’s LLMs understand content (SMX Munich, March 2025).
Side-by-side showing JSON-LD schema code on the left with syntax highlighting versus the same code as flat plain text on the right, illustrating how LLMs tokenize schema as raw text not structured data

Schema markup has become the default technical recommendation for anyone trying to improve their visibility in ChatGPT, Perplexity, and Google AI Overviews. Implement FAQPage schema and AI will understand your content better. It sounds credible. It travels fast. And two separate research studies now suggest it is largely wrong.

What the Experiments Show

In early 2026, SEO consultant Mark Williams-Cook ran a controlled test to establish whether LLMs actually parse JSON-LD as structured data, or just read it as text. He created a fictional company and embedded its address inside invalid, made-up schema markup, with no visible text equivalent on the page.

Both ChatGPT and Perplexity returned the address. Not because they understood the schema, but because they read the script block as raw text, the same way they read everything else. The @type declarations, the property relationships, the structured hierarchy — none of it registered. The models saw a block of text that happened to contain an address.

LLMs tokenise JSON-LD as plain text. The structured relationships you build into schema are not relationships the model sees.

Search Atlas reached a similar conclusion from a different angle. Their December 2024 study compared schema coverage against citation rates across OpenAI, Gemini, and Perplexity. Domains with comprehensive schema coverage performed no better than domains with minimal or no schema across all three platforms.

Where the Correlation Comes From

Some studies do show a positive correlation between schema presence and AI visibility. Ahrefs offered the most coherent explanation: sites that implement schema tend to be better-maintained and higher-quality overall. The schema is not driving citations. Site quality is driving both the schema adoption and the citations independently.

One genuine exception exists. Microsoft’s Fabrice Canel stated at SMX Munich in March 2025 that schema helps Microsoft’s LLMs understand content. That is an on-the-record platform statement and worth acknowledging. It is also the strongest direct evidence in favour of schema for AI, which says something about how thin the supporting case is elsewhere.

What Actually Matters

The research across platforms points consistently to a different set of signals. Content quality and domain authority account for an estimated 70% of citation factor weighting. Schema accounts for roughly 10%. The practical priorities, in order of impact:

  • Content quality and topical authority. The single highest-leverage factor. A well-written, credible page on an authoritative domain outperforms technically optimised content on a weak one every time. AI systems have processed enough of the web to distinguish credible from thin, regardless of how the thin content is marked up.
  • Recency. ChatGPT cited content published within 30 days at an 82% rate in 2026 analysis. Perplexity’s window is even tighter. AI tools that retrieve live web content are especially sensitive to publication and update dates. Keeping content current is a higher-leverage activity than adding schema to static pages.
  • Question-structured headings. H2 and H3 headings written as questions help AI systems extract your content cleanly. This is the kernel of truth inside the FAQ advice. It comes from the visible heading structure, not from schema layered on top. Write the heading — you do not need the markup to get the benefit.
  • Primary data and named sources. Citing research, naming authors, and linking to original data signals credibility to the models deciding what to surface. Original data that does not exist elsewhere is particularly high-value, because the AI has no alternative source to prefer.

The Bottom Line

Schema is not useless. It still has value for traditional search rich results and maintaining a technically sound site. But if your AI visibility strategy is anchored to schema implementation, you are optimising heavily for something that accounts for roughly 10% of what matters, while the factors that actually move the needle get less attention.

The schema advice is a good example of a pattern that repeats itself in SEO: a technically plausible claim, a few correlational data points, and an industry that repeats it before the controlled experiments catch up. They have now started to.

Sources

Frequently asked questions

Common questions about schema markup and AI citations

Schema markup has a measurable but limited effect on AI visibility. Research suggests it accounts for roughly 10% of the factors that influence whether an AI platform cites your content. The dominant factors — content quality, domain authority, and topical relevance — account for an estimated 70%. Schema still has value for traditional search rich results, but it should not be the anchor of an AI visibility strategy.

LLMs tokenize JSON-LD as plain text. They read the contents of the script block the same way they read any other text on the page — as a sequence of characters. The structured relationships encoded in schema (the @type declarations, property hierarchies, linked entities) are not parsed as structured data by the model. A controlled experiment confirmed this: a fictional address embedded in invalid, made-up schema was returned accurately by ChatGPT and Perplexity, not because they understood the structure, but because they read the text.

The research points consistently to four high-impact factors. Content quality and domain authority account for an estimated 70% of citation factor weighting. Recency matters significantly — ChatGPT cited content published within 30 days at an 82% rate in 2026 analysis. Question-structured headings (H2 and H3 written as questions) help AI systems extract content cleanly. Primary data and named sources signal credibility. Schema markup accounts for roughly 10% of the picture.

No. Schema still provides real value for traditional search rich results — FAQ snippets, product markup, local business data, event listings — and for maintaining a technically sound site. One notable exception: Microsoft’s Fabrice Canel stated at SMX Munich in March 2025 that schema helps Microsoft’s LLMs understand content. The argument is not to abandon schema but to stop treating it as a primary AI visibility lever when it accounts for a small fraction of what actually moves the needle.

Topical authority is the degree to which a website is recognized as a credible, comprehensive source on a specific subject. AI platforms learn to associate domains with areas of expertise by processing large volumes of text across the web. A site that consistently publishes high-quality, accurate content within a defined niche builds a signal that its content is reliable in that area. This is one of the highest-leverage factors in earning AI citations — more so than technical optimizations like schema.

Content recency is one of the most consistently cited factors in AI citation research. ChatGPT cited content published within the past 30 days at an 82% rate in 2026 analysis, and Perplexity’s recency window is even tighter. AI tools that perform live web retrieval are especially sensitive to publication and update dates. Keeping content fresh — updating statistics, adding new data, and refreshing high-value pages — is a higher-leverage activity than adding schema markup to static content.

No credit card required

Find out where you stand.

We test the prompts your local customers are actually typing into AI right now.

Get your free audit