Intro
The advice has been repeated everywhere: “Add an FAQ block, reformat your content as questions and answers, and AI engines will start citing you.”
A 2026 peer-reviewed study covering 602 prompts and 23,745 citation-level features says the opposite. Q&A formatting, taken in isolation, does not improve citation absorption. The pages AI engines actually reuse are longer, more modular, semantically aligned with the answer, and dense with extractable evidence. Surface formatting is not the driver.
This page sets the record straight, summarizes the empirical finding, and explains what to optimize instead.
Why the FAQ-only approach falls short
The intuition behind the FAQ tactic is reasonable. AI engines often return answers shaped like questions, so mirroring that shape on the page sounds aligned. Several SEO blogs have generalized that intuition into a rule.
The rule does not hold up under measurement.
Three reasons explain the gap between the tactic and the result.
First, AI engines do not select sources by shape — they select by content. A page that wraps thin claims inside questions and short paragraphs adds visual structure without adding extractable evidence. Engines parse the underlying claim, not the punctuation around it.
Second, FAQ blocks often dilute the page. Pages that already had a strong answer in the body sometimes lose absorption when the same point is restated, in shorter form, inside an FAQ. The model picks the weaker formulation because it is closer to the question.
Third, FAQ schema is a hint, not a ranking signal for citation. It helps search features render rich snippets. It does not, on its own, change which passage a generative engine extracts.
The FAQ tactic is not wrong — it is just incomplete. It is one of the cheapest interventions and one of the smallest in effect.
What the empirical data actually shows
The reference study (Zhang, He & Yao, 2026, “From Citation Selection to Citation Absorption”, arXiv:2604.25707v2) analyzed 602 prompts across multiple generative engines and extracted 23,745 citation-level features from the answers produced.
The authors split the features into two groups: page-level attributes (length, structure, authority, recognizability, language, domain context) and passage-level attributes (semantic alignment with the prompt, structural legibility of the passage, density of extractable evidence such as definitions, numerical facts, comparisons, procedural steps).
Their headline finding for practitioners: Q&A formatting alone does not improve absorption. Pages that performed best were:
- Longer than the median, with more anchorable sections
- Modular — content broken into independent, self-contained units
- Semantically aligned with the answer the engine was constructing
- Dense with concrete evidence: numbers, definitions, comparisons, named entities, dated facts
The paper reframes generative engine optimization as evidence-container design. The container has page-level properties (authority, recognizability, language, domain context) and passage-level properties (semantic alignment, structural legibility, evidence density). FAQ formatting touches one slice of structural legibility. Everything else is left untouched if a brand only ships FAQs.
This is the core of the AI answer evidence layer we maintain at Capston: every page is scored on the seven container properties, not on a single tactic.
Evidence density vs surface formatting
The distinction worth internalizing is between surface formatting and evidence density.
Surface formatting changes how a page looks: H2 questions, bulleted lists, table-of-contents blocks, FAQ schema. It is fast to ship. It signals organization. It does not, by itself, add anything the engine can quote.
Evidence density changes what the page actually says: a specific number, a dated source, a clear definition, a side-by-side comparison, a named procedure, an attributed quote. Every one of those is an extractable unit. A model lifting a sentence from the page gets a complete, defensible claim.
A useful diagnostic: read any paragraph on the page and ask “if a generative engine had to cite this without paraphrasing, what would it quote?” If the answer is “nothing specific”, evidence density is low — regardless of how many FAQ blocks the page contains.
This is exactly the lens applied in our Capston Core methodology during the content audit stage. Pages are graded on extractable units per 500 words, not on the presence of FAQ markup.
What to do instead (4 actions)
A practical sequence for teams currently over-investing in FAQ blocks.
1. Audit evidence density before audit formatting.
For each priority page, count the extractable units (numbers, definitions, comparisons, dated facts, procedural steps). Below 4 per 500 words is thin. Above 10 is strong. Fix density before touching FAQ markup.
2. Tighten semantic alignment with the actual prompts.
Run the prompts that matter. Read the answers the engine constructs. Rewrite page passages so they match the angle, the vocabulary, and the level of specificity the engine expects. A page semantically aligned with five real prompts will outperform a page padded with twenty generic FAQs.
3. Modularize, then add structure.
Break long paragraphs into self-contained units. One claim per paragraph. One definition per block. Engines lift modular content cleanly; they struggle with multi-claim paragraphs. Once modular, structural cues (headings, lists, schema) start to compound.
4. Use FAQ blocks for prompts they actually answer.
FAQ blocks still have a place — for genuine questions that recur across the prompt set, for which a short, evidence-dense answer exists. Three real FAQs with hard data beat twelve generic FAQs with rephrased filler. The FAQ block then carries evidence density, not just shape.
These four actions are the practical translation of the paper’s finding. They are also how we sequence work for clients on the AI visibility scoring program.
How this fits into Capston Core
The Q&A formatting myth is a useful example of why Capston Core treats GEO as evidence-container design rather than a checklist of tactics. The Capston Core methodology audits seven container properties, the AI answer evidence layer tracks them prompt by prompt, and the AI visibility scoring system measures whether absorption is actually moving.
A brand that only ships FAQ blocks will see the score barely move. A brand that closes the evidence-density gap will see it move on the dimensions that matter for commercial outcomes.
→ Back to Capston Core
FAQ
So FAQ schema is useless?
No. It still helps render rich snippets in classical search and provides a clean structural hint. It is just not a driver of generative citation absorption on its own.
What is “evidence density” in practical terms?
The count of extractable units per 500 words — numbers, definitions, dated facts, named comparisons, procedural steps. We treat 4–10 as a working range, with priority pages above 10.
Does length matter on its own?
Longer pages correlate with higher absorption in the study, but only when the length adds modular, evidence-dense sections. Padding hurts.
Which paper is this based on?
Zhang Kai, He Xinyue & Yao Jingang, “From Citation Selection to Citation Absorption” (arXiv:2604.25707v2, 2026). 602 prompts, 23,745 citation-level features.
Reference
Zhang, K., He, X., & Yao, J. (2026). From Citation Selection to Citation Absorption: An Empirical Study of Generative Engine Citations. arXiv:2604.25707v2.
Final CTA block
Stop optimizing for FAQ shape. Start measuring evidence density.
Audit your evidence density
Read the methodology