INSIGHTS / RESEARCH EXPLAINED

Does Schema Markup Increase AI Citations? What the Evidence Shows

Adding schema is not a demonstrated universal route to more AI citations. Judge the specific defect, the kind of evidence available and the outcome you intend to measure.

7 min read

TL;DR

  • Schema describes page facts. It is not a proven shortcut to AI citations.
  • Read citation studies, retrieval tests and rich-result cases as different kinds of evidence.
  • Fund a defined fix and measure the outcome it is supposed to change.

What happened when pages added schema?

Ahrefs tracked 1,885 pages adding JSON-LD, against 4,000 controls, between August 2025 and March 2026. Its matched difference-in-differences estimates were −4.6% for AI Overviews, +2.4% for AI Mode and +2.2% for ChatGPT. The two positive estimates were statistically indistinguishable from zero. This was observational matched analysis, not a randomized trial. Ahrefs study, May 2026.

For a B2B team, the decision is narrower than “schema works” or “schema is useless.” Ask what job the markup must do. Clear business facts, richer search results and more AI citations are three distinct goals. Evidence about one cannot settle all three.

Schema is a vocabulary for describing information. JSON-LD (JavaScript Object Notation for Linked Data) is a common code format for those descriptions. Neither tells you whether an AI answer will cite your page.

Estimated citation change relative to matched controlsRelative change (%)
-6%-3%0%+3%+6%-4.6%AI Overviews+2.4%AI Mode+2.2%ChatGPT

Ahrefs matched analysis. Positive estimates were not statistically distinguishable from zero. Orange = negative estimate; blue = positive estimate. These are relative percentages, not percentage points. Ahrefs schema study.

What does the study leave unresolved?

The Ahrefs pages already had substantial AI citations. The analysis used a 30-day post-change window, pooled schema types and could not fully separate other edits. AI Overviews showed a small statistically significant decline relative to controls; it is not a general reason to remove schema. Uncited sites and longer-term effects remain open. Study limitations.

Our view: do not promise a gain. Do not promise no effect either. Ask the supplier which page needs work. Ask what is wrong and why the fix matters to buyers. “More schema” is not a complete scope.

A useful proposal separates facts from hypotheses. “This product page has the wrong currency in its markup” is a verifiable defect. “Correcting it will make ChatGPT recommend us” is a hypothesis. You can check the first after launch. The second needs data from later tests.

This makes reports clearer too. A team can finish a sound fix before it knows the effect on sales. Agree how to check the work and when to review its impact. That makes the scope and fee easier to judge.

Can models read facts that appear only in schema?

A separate searchVIU test used a synthetic product page on October 30, 2025. The JSON-LD-only price was absent from all reported system results. The report mixes direct retrieval and index-based checks, and states both five and ten repetitions in different places. Treat it as a limited diagnostic, not a platform-wide benchmark. searchVIU experiment

Fact extraction tests whether a system returns a specific fact. Citation-frequency research measures how often sources appear. Neither outcome is a substitute for the other. The historical single-page diagnostic cannot establish how all current products parse JSON-LD or execute JavaScript.

Google does not require special schema for generative search. Current Google guidance. Its structured-data policy requires relevant, current information that agrees with what people can see. Structured data policies. These policies describe Google, not every AI platform.

Start with your own page. Can a buyer find the scope, price basis and proof in its text? A rich block of code beneath a thin page does not help that buyer. Put the approved answer on the page before deciding how to describe it in markup.

Compare the evidence before approving work.

The sources ask different questions. Read each result against its unit of analysis and scope. None of these sources is a SemShift experiment.

Compare the evidence before approving work.
Source / typeQuestion / objectOutcomeSupportsDoes not support
Ahrefs / matched observational analysis1,885 already-cited pages adding schema; 4,000 controlsRelative citation estimates shown aboveNo clear positive lift in this measured sampleA randomized causal estimate or a claim about every site
searchVIU / historical single-page diagnosticWhether tested products extract facts placed in different page representationsPresence or absence of a returned factA reason to test visibility of important factsA citation-rate estimate or a current universal parser rule
Google / generative search guidance and markup policyEligibility and accurate page representationDocumented requirements, not experimental outcomesA defined Google implementation checkCross-platform performance predictions

When is a schema change still useful?

Use markup for a valid, applicable purpose such as accurately describing a page or supporting an eligible search feature. A correct validation result does not guarantee display. Google’s structured-data rules.

Google’s update log records that FAQ rich results stopped appearing on May 7, 2026. Documentation update. Keep useful questions and answers for readers; do not add FAQPage simply as a promised growth feature.

The decision is not “more markup” versus “no markup.” Ask whether the proposed change corrects a real inconsistency or supports a documented feature. Search click-through case studies do not establish an AI citation effect.

Illustrative fix: the page and markup disagree.

A fictional product page visibly offers a price in US dollars, while its JSON-LD labels the same offer as euros. This is a hypothetical defect, not a real offer or a recommendation to publish false prices.

The product owner first confirms the actual offer and currency. The developer aligns the visible price and all markup representations with that approved record. Check for stale plugins or duplicate scripts that could reintroduce the conflict.

Accept the fix when the published facts agree and the applicable validation passes. Whether citations change is a different observation requiring a stable prompt sample and preserved outputs.

Illustrative fix: the page and markup disagree.
Implementation checkAcceptance evidence
Visible factsRead back the approved offer, currency and conditions on the live page
Markup factsParsed structured data describes the same offer and currency
Syntax and applicable typeValidation result and relevant feature requirements reviewed
Duplicate or conflicting dataInspect all markup blocks and generated variants
Deployment recordURL, version, change date, approval and before/after record retained

What should a B2B team approve next?

Approve a small, testable change tied to a real page. Google requires markup to represent current, visible content, and a valid test does not guarantee a rich result. Use validation to check the implementation, not to certify future visibility. Structured data policies

Our suggested review has three steps. First, name the fault and who will fix it. It might be a wrong fact, invalid code, a missing field or a clash with the page text. Next, agree what “done” means. Save a record of what changed. Then decide if a separate study of the impact is worth its cost.

For an outcome study, define the questions and platform settings before launch. Keep dated responses and source URLs, including runs with no citation. Record other site changes and use a comparable unchanged group where feasible. Report sparse or mixed results as such. A screenshot chosen because it looks good is not a fair test.

For an outcome study, use the sampling rules and rate calculations. Keep no-citation answers, record concurrent changes and compare an unchanged group where feasible. Do not publish fake prices or remove sound markup to manufacture a control. The citation diagnostic guide helps interpret source errors.

Schema belongs inside a broader page review. See how to turn buyer questions into a content brief to identify missing facts and proof. Bring the page, the suspected conflict and its approved source to discuss a scoped review with SemShift.