Why are existing schema ROI claims untrustworthy?
Because almost all of them commit the same three sins. First, no control group: pages get schema, citations change (or do not), and everything else that shipped that quarter gets ignored. Second, selection bias: teams deploy schema on their best pages first, then attribute the pages’ pre-existing strength to the markup. Third, moving targets: engines update mid-study, so the before and after measure different systems. Any article quoting a precise uplift percentage for schema without addressing these three problems is selling, not studying. We will not quote one either.
What would an honest design look like?
Cohort construction
Select 60-100 comparable pages — same content type, similar baseline visibility, same site section where possible. Pair them on baseline Citation Rate from a frozen question set, then randomly assign one of each pair to the treatment group (schema deployed) and keep the other as control (unchanged). Matching first, then randomization, controls both selection bias and page-quality confounds.
Staggered intervention
Deploy schema to the treatment group in one defined week — Organization, Article, FAQPage or Product as appropriate, validated, server-rendered, matching visible content exactly. Ship nothing else to either cohort during the study window: no rewrites, no new internal links, no refreshes. The intervention must be the only thing that changes, or the study measures your content calendar. Full deployment standards are in structured data for AI.
Measurement protocol
- Frozen question set of 50-100 questions mapped to the study pages, fixed before the baseline run.
- Monthly runs per engine — Google AI surfaces, ChatGPT, Perplexity — with Citation Rate and Visibility scored per page-cluster, screenshots archived.
- Minimum window: two full measurement cycles post-deployment, because indexing and grounding lag by weeks. A 90-day POC structure — baseline, intervention, re-measurement — is the natural container; see the anatomy of a 90-day POC.
- Pre-registered scoring rules: write down, before the baseline, how you will score citations, what counts as a win, and which results you will publish (answer: all of them).
What results should you expect — and report?
Honestly, a modest and noisy signal. Schema plausibly helps at the retrieval and grounding stages — machine-readable passage boundaries, unambiguous entities, consistent dates — but it competes with content quality, source-graph strength and engine updates, and its effect is unlikely to dwarf them. Report the deltas with their uncertainty: treatment versus control, per engine, per dimension, with the screenshot archive attached. A null result is publishable and useful — “schema alone did not move citations in 90 days on our pages” saves the next team a quarter of misallocated effort.
What are the honesty rules?
- Publish the question set and scoring rubric with the results.
- Publish null and negative deltas, not just wins.
- Disclose every other change that shipped during the window, even to control pages.
- Never extrapolate one study to “schema increases citations by X%” universally — engines, verticals and page types differ.
Why bother if the effect is modest?
Because the study’s real output is not a percentage; it is an organizational habit. A team that can run a controlled 90-day measurement on schema can run one on anything — content patterns, source-graph pushes, freshness rituals — and that capability compounds far beyond any single tactic. Schema is simply the cleanest first experiment: binary, deployable in a week, and surrounded by enough hype to make an honest answer valuable. See the GEO testing framework for the general method, and how we measure Citation Rate for the scoring details. Publish the protocol openly and the study doubles as original research — exactly the kind of citable artifact that strengthens your own source graph while it answers the question.