How LLMs evaluate FAQ schema for citations — and why most implementations get it wrong

Does your FAQ section help or hurt your chances of being cited by ChatGPT and Perplexity? The answer depends less on what you say and more on how you structure it.

A growing body of AI citation research shows that FAQPage schema carries disproportionate weight in LLM answer generation — because FAQ content is pre-segmented into question-answer pairs that require minimal processing. An AI agent can extract, verify, and cite an FAQ answer in a fraction of the time it takes to parse a long-form article.

But most ecommerce FAQ implementations actively work against this advantage. Here is what Visora's scan data reveals about what works.

When an AI shopping agent encounters FAQPage structured data, it gets three things standard content does not provide:

1. Explicit question-answer segmentation. The schema marks exactly where one question ends and the next begins. No parsing ambiguity. 2. Authoritative source attribution. The @id and mainEntityOfPage fields tie each answer back to a specific URL, making it easy to cite. 3. Confidence-weighted extraction. LLMs assign higher confidence to content wrapped in schema types they were trained on extensively. FAQPage is one of the most commonly seen schema types in training data.

Visora's internal citation tracking data shows that product pages with correctly implemented FAQPage schema are cited by AI agents at 2.3x the rate of pages with FAQ content in plain HTML — and at 3.1x the rate of pages with no FAQ at all.

Visora's free audit at geovisora.com/audit checks for 18 citation-readiness signals. Of those, the FAQPage-specific checks reveal a cluster of recurring issues across merchant sites:

1. One giant FAQPage instead of multiple. Many themes output a single FAQPage wrapper for the entire FAQ section. If you have 20 questions, it should be 20 FAQPage entries — each with its own question-answer pair and URL anchor. LLMs prefer granular entries because they can cite a single answer without linking the entire section.

2. Answer text hidden behind "Read More" clicks. If the full answer text is not present in the initial HTML (loaded by JavaScript on click), an LLM's crawler sees only the question. The answer field is then empty or contains placeholder text, which degrades the citation signal to near zero.

3. Missing acceptedAnswer textContent. The acceptedAnswer property requires actual text content. Some implementations put the answer in an image or a nested div with no text representation. Schema validators may pass, but LLM citation systems reject non-text answers.

4. No anchor links (@id with URL fragment). Each FAQPage entry should have a unique @id like "#faq-shipping-time" that corresponds to an anchor on the page. Without it, the LLM cannot confidently map the answer to a visible page position.

5. Generic questions that LLMs already know. "How long does shipping take?" will trigger citation less often than "How long does shipping take to Germany from your US warehouse?" — because the specific version is not answerable from general knowledge. LLMs cite content that provides novel, specific information.

6. Mixed FAQ and HowTo in the same schema block. While combining multiple schema types is valid, doing so without clear separation confuses citation extraction. FAQPage entries interspersed with HowTo steps reduce extraction confidence scores.

7. FAQ after the fold, below the footer or hidden in a modal. LLM crawlers prioritize above-fold content. FAQ sections loaded at the bottom of the page or inside interactive modals may be deprioritized or skipped entirely if the crawl budget is limited.

If your product pages already have FAQ sections, here is a step-by-step checklist to optimize them for LLM citation:

Run your top 10 product pages through Visora's free audit at geovisora.com/audit. Check your FAQPage scores. If they fall below 70 on the Schema Completeness metric, apply the fixes above and rescan. Most merchants see measurable improvement in AI citation frequency within two weeks of correcting their FAQPage implementation.