FAQ with Schema.org for AI visibility: what works, what backfires, and why the answer depends on the platform

Structured data is one of the oldest, most consistent SEO recommendations. When generative AIs started to matter, the obvious conclusion was: if Schema.org helps Google understand your content, it should help the AIs too. The experimental evidence from 2025–2026 tells a more complicated — and more interesting — story than that assumption suggested.

The experiment that flipped the expectation

OtterlyAI ran a controlled experiment between December 2025 and March 2026, running 319 prompts and measuring the impact of Schema markup on citations across seven different AI platforms. The results broke the expectation clearly:

In Google AI Overviews, citations grew 611% after Schema was implemented. In ChatGPT, they fell 71% over the same period. Six of the seven platforms tested showed no positive effect attributable to Schema.

The explanation is technical but accessible: large language models don't "read" JSON-LD like a structured database. They tokenize Schema as plain text — the same way they tokenize any other content on the page. To the model, "@type": "FAQPage" is a text string like any other, not a semantic instruction that changes how it processes the information. For Google AI Overviews, on the other hand — which operates in a layer closer to the traditional search engine, with direct access to Google's structured index — Schema works the way it was designed to.

What this means in practice

It doesn't mean "don't implement Schema." It means Schema solves different problems depending on where you look. For visibility in Google AI Overviews — which shows up in around 48% of all tracked searches in 2026 — implementing FAQPage and other Schema types has a real, documented effect. For citation by ChatGPT or Perplexity, the strategy has to be something else.

What actually helps in ChatGPT and in AIs that operate by text retrieval is the underlying content — the quality and specificity of the questions and answers as written, whether or not they're marked up with Schema. A well-written FAQ block, with questions that mirror exactly what someone would ask in a real conversation and answers that lead with the point in the first sentence, will get cited more with or without the Schema markup.

How to write FAQ questions that AIs actually cite

The difference between an FAQ that exists and an FAQ that gets cited is in how specific the questions are. Questions written for traditional SEO tend to be phrased as expanded keywords: "What is HR management?" or "How much does a time clock cost?" They're generic, they answer search volume, but they don't mirror the conversational tone of someone using an AI today.

Questions written for AI citation have more situational specificity: "Is a company with 15 employees required to keep time-clock records?" or "What's the difference between an electronic time clock and paper timesheets from a labor-audit standpoint?" These questions map directly onto what someone would type into ChatGPT, and the answer — if direct and informative — has a high chance of being the basis of the generated response.

(illustrative) Compare two FAQ blocks from a fictional HR software company:

Keyword-optimized version:

• "What is payroll?"
• "How do you manage HR?"
• "What are the advantages of our software?"

AI-citation-optimized version:

• "Is a company with 12 full-time employees required to keep time records?"
• "What's the penalty for filing payroll taxes late?"
• "Can payroll be integrated with an external accounting system?"

The second set answers questions a real business owner would ask an AI before signing up for a service. The first set doesn't resolve any specific question for anyone.

Where Schema still makes sense even without changing ChatGPT citations

Beyond Google AI Overviews, there are two more arguments for keeping Schema on the site even in the face of mixed evidence. The first is longevity: the AI ecosystem is changing fast, and there's no guarantee the models will keep ignoring Schema — future versions may implement structured JSON-LD reading explicitly.

The second is interoperability with non-LLM systems: voice assistants (Google Assistant, Alexa), knowledge panels (Google Knowledge Panel), rich snippets in traditional search — all of these systems benefit from Schema markup and stay relevant for most businesses.

The combined strategy

The practical result of this analysis is a strategy on two independent axes: implement Schema FAQPage to capture the real benefit in Google AI Overviews, and at the same time rewrite the content of your questions and answers into the specific conversational format that AIs like ChatGPT actually cite — regardless of the markup. The two efforts don't contradict each other; they solve different problems on different platforms.