**TL;DR** — Across 22 client sites in April and the first two weeks of May 2026 we audited whether Schema.org structured data — specifically Article, HowTo, FAQPage and Organization markup — measurably changes the probability of a page being cited in Google AI Overview answer cards. Across 4,300 tracked queries and 1,140 captured citations the headline was narrower than the SEO consensus suggests: Article schema produced no detectable lift on citation probability when content quality was held constant, and FAQPage schema mildly hurt extraction frequency in answer cards. Two schema-related signals did move the needle: clean Organization markup with `sameAs` and `description` fields measurably increased the rate at which the answer card cited the brand directly when the answer named it, and HowTo schema increased extraction by 4.3 percentage points — but only on pages whose steps were also rendered as semantic HTML `<ol>` lists. Schema in 2026 is a disambiguation lever, not a citation lever, and the editorial budget many teams spend on it would pay better against the lead paragraph.
Why we ran this audit
For three years running our intake conversations have included some version of the same question: "should we add more schema?" Through 2023 and 2024 the answer was usually a careful yes — schema was inexpensive, it was suspected to help with rich-result eligibility, and there was no observable downside. By Q1 2026 that intuition was costing real editorial hours. Several clients were spending two days a sprint on schema work — adding HowTo to every tutorial, FAQPage to every commercial page, Article to every blog post — and treating the markup as a lever on AI search citation specifically. We had no audited evidence the lever existed, and the time was coming out of editorial work that we did have audited evidence for. The audit's purpose was to settle the question: does any of this schema actually change what the composer cites?
The second motivation was definitional. "Schema helps AI search" is a sentence that gets repeated at every SEO conference in 2026, and it is doing two different jobs at once — sometimes meaning "schema increases the probability your page is selected as a source" and sometimes meaning "schema influences how your brand is described in the answer." Those are different claims and they need different audits. Folding them together produces the worst kind of consensus: the kind everyone agrees with because everyone is hearing the half they prefer. We wanted to test the harder, more useful version: does schema markup change selection, separately from how it changes presentation?
How we ran the measurement
22 client sites — 8 SaaS, 6 publisher, 5 DTC, 3 B2B services — across April and the first ten days of May 2026. For each site we paired pages: a treated page with Article, HowTo or FAQPage markup, and an untreated control page on the same site with closely matched content quality, length, lead structure, and historical citation rate. Where no clean control existed, we created one by removing schema from a previously treated page and waiting 14 days for the index to settle before measuring. Each pair was tracked across a fixed 80-query basket of queries the treated page was eligible to be cited on, twice daily for the audit window. We logged citation events, the cited paragraph (from our existing extraction reconstruction), and the answer-card prose that accompanied each citation.
Two normalisation moves matter for reading the numbers below. We classified pages by content shape in advance — tutorial, comparison, definitional, opinion — and we report results within each shape because schema effects looked different in different shapes; aggregating would have hidden the HowTo signal. We also excluded any treated page where the schema was added or removed inside a 14-day window of the measurement, because the post-change interval is a transient we did not want polluting the steady-state comparison. The reported numbers are for the steady-state population, which is what an editorial team actually controls and what a "should we ship schema" decision actually faces.
The shape of the result
The flat headline first. Across the full cohort, pages with Article schema were cited at 23.4% of eligible queries; matched pages without Article schema were cited at 24.1%. The difference is inside the noise floor of an audit of this size, and we report it as zero. The same comparison for HowTo across tutorial pages — and only tutorial pages, because HowTo on a comparison page is misapplied — came in at a 4.3 percentage-point lift, marginally significant. FAQPage schema produced a 2.1 percentage-point drop in citation probability versus matched pages without it. None of these effects were large enough to justify the editorial budget being spent on them in isolation; the HowTo lift came with a caveat that erased it on most pages.
Two non-flat findings sat inside the noise. The HowTo lift was concentrated entirely on pages where the steps were also rendered as semantic HTML `<ol>` lists with one step per `<li>`; on pages where HowTo was present but the steps were in body prose or in styled `<div>` blocks, the lift collapsed to zero. Reading this carefully, the schema is not doing the work — the schema and the markup together are doing the work, and the schema alone does nothing. Separately, on queries that named the brand directly in the answer prose, pages on sites with clean Organization markup carrying sameAs and description fields received the citation 11 percentage points more often than pages on sites without it. That number is reliably large and the mechanism is plausible: the composer is more confident which entity it is naming when the entity has machine-readable identity attached.
Driver one: Article and FAQPage schema do not move citation probability
Article schema is the most-shipped structured data type on the editorial pages we audit, and its measured contribution to citation probability was indistinguishable from zero. We expected this for a specific reason: Article schema describes a publication artefact — author, publishDate, articleBody — but it does not describe the answer the page provides to any specific query, and the composer's selection decision is keyed to whether the page answers the implicit query well, not to whether the page is well-described as a publication. Article schema helps Google understand that a page is an article. By 2026 the composer can already tell.
FAQPage schema is the more interesting result, because the small negative effect — a 2.1 percentage-point drop — was not what the consensus predicted. Reading the matched pairs case by case, the most plausible explanation is structural: pages with FAQPage schema tended to format more of their answer content as Q-and-A blocks, and the composer extracts from Q-and-A blocks at well below the rate it extracts from opening prose or post-H2 paragraphs (the 1.8% extraction share we measured in the cited-paragraph audit). The schema is not directly hurting citation; the editorial pattern that comes with the schema — chunking answers into Q-and-A pairs instead of writing tight claim-and-evidence prose — is shrinking the extraction surface. The lesson is uncomfortable but operationally clean: shipping FAQPage schema does not cost you citations; shipping the FAQ-shaped content the schema usually accompanies might.
Driver two: HowTo schema lifts extraction only when the steps are semantic HTML
On tutorial pages with HowTo schema, citation rate rose by 4.3 percentage points — a real but modest effect. When we split the HowTo population by step markup, the entire lift was concentrated on pages where the steps were rendered as semantic `<ol>` lists with one step per `<li>`. Pages with HowTo schema but with steps in body prose or in styled `<div>` blocks performed identically to pages with no schema at all. The composer is not reading HowTo schema for the procedure; it is reading the HTML `<ol>` for the procedure, and the HowTo schema appears to act as a small confidence boost on top of the structural signal.
This shifts the schema-budget allocation in an obvious direction. Shipping HowTo schema on a page whose steps are not in an `<ol>` is buying nothing — the schema is decorative. Conversely, the editorial work of converting prose-style instructions into a real `<ol>` picks up roughly the same extraction lift even without the schema. We have not stopped shipping HowTo schema; the marginal cost is small enough that the redundant pair is worth keeping. But when a client team is choosing between adding HowTo to ten more pages and converting the steps on existing tutorial pages to semantic markup, the second move wins on every comparison we ran. The HTML is doing the citation work; the schema is acting as the seatbelt sign.
Driver three: Organization schema disambiguates the brand in the answer
The 11-percentage-point lift on brand-naming queries from clean Organization markup with sameAs and description was the only large effect in the audit. The mechanism is not extraction: the composer is not pulling the description string into the answer prose verbatim. It is selection — the composer is more likely to cite the brand's own domain when the entity is unambiguously identified, because Organization markup with sameAs (Wikipedia, Crunchbase, LinkedIn, GitHub for technical brands) and a description field gives the composer a confident entity resolution it does not have from prose alone. On queries where multiple entities shared a brand name — a frequent problem for short or generic brand names — the lift was bigger; on queries with no name collision the lift was smaller but still present.
The practical version of this is short. For every brand we work with, the homepage carries Organization JSON-LD with name, url, logo, sameAs pointing to verified third-party entity pages, and a one-paragraph description field that matches the prose on the About page. We treat sameAs as the load-bearing field — the more verified third-party identity links, the more confident the composer's entity resolution. Adding a second sameAs entry where one existed gained roughly 4 percentage points on entity-disambiguation queries; adding a third gained another 2. The diminishing returns kicked in around five entries. Six sameAs links is the budget we now ship, and we have not seen a case where shipping more changed anything.
What changed in our content checklist
Three changes. We stopped recommending Article schema as a default editorial line-item — when a client asks whether to add it, we say yes if it is cheap and no if it costs editorial time, because the audited contribution is zero either way. We stopped recommending FAQPage schema entirely on commercial-intent pages, because the editorial pattern that comes with it underperforms tight claim-and-evidence prose on the same questions, and the schema is the smaller half of that effect. We now treat Organization schema as the highest-leverage structured-data investment for any brand whose name is even mildly ambiguous, because the 11-point lift on brand-naming citations is the largest schema effect we have ever measured.
We dropped one habit. Through 2024 and 2025 we built a "schema completeness" score into every monthly client report — a percentage of pages carrying the appropriate schema type. The score moved up steadily for every client and correlated with approximately nothing we cared about. We retired it this quarter. The dashboard line that replaced it counts only Organization markup completeness, which actually predicts citation behaviour, and we treat HowTo and Article as page-level optional decorations rather than portfolio-level metrics. The reporting change was modestly embarrassing to walk back with clients, and the editorial hours it freed up went into lead-paragraph rewrites where the citation work actually happens.
- 01Stop treating Article and FAQPage schema as citation levers. Across 4,300 queries the measured effect on citation probability was zero for Article and a small negative for FAQPage — the editorial budget pays better against the lead paragraph.
- 02Pair HowTo schema with semantic `<ol>`/`<li>` step markup or do not ship it. The 4.3-point lift collapsed to zero on pages where the steps were in prose or styled `<div>` blocks; the HTML is doing the work.
- 03Ship Organization JSON-LD with sameAs and description for every brand. The 11-point lift on brand-naming citations was the largest schema effect we measured, and diminishing returns kick in around five sameAs entries.
- 04Retire the "schema completeness" dashboard. The score moves up steadily and predicts approximately nothing; replace it with Organization-markup completeness and treat everything else as page-level optional decoration.
Where this argument breaks
For news publishers, NewsArticle schema continues to be a hard requirement for Top Stories and Discover surfaces, which we did not audit and where the calculus is different — the schema there is a gating mechanism for distribution, not a lever on AI Overview citation, and dropping it would lose the surface entirely. For Recipe, Event, Product and the other commerce-flavoured schema types we did not test, the rich-result eligibility argument still applies and the audit does not displace it; this is specifically about Article, HowTo, FAQPage and Organization in the AI Overview citation context. For very small sites with fewer than a dozen pages, the matched-pair design we used cannot run cleanly and the audit reduces to a qualitative judgement. In Chinese-language search, 百度 still extracts FAQPage-style Q-and-A content at materially higher rates than Google does in 2026, and the FAQPage negative does not transfer — there it remains a defensible investment. Outside those carve-outs the lesson holds: in 2026 most schema is not buying you citations, it is buying you machine-readable identity, and an editorial budget that confuses the two is allocating against the wrong target.