How to Measure the Return on Structured Data

Two weeks before the March board pack, a DTC seller of portable power stations for US and German buyers presents one slide: rich results now cover nearly every product page, and AI-referred sessions are up over the same quarter. The CFO asks which number she can defend. Silence. That quarter also shipped a flagship, paid two creators, and replaced the spring landing page.

Structured data pays back through citations and better-qualified clicks, not through rich results a platform may withhold — so measure it with a control set and a time split, not a sitewide curve.

A rich result is a display slot you do not own

Most schema business cases rest on a fragile assumption: that the rich result is the return. Whether it appears, how big it is, and when it disappears are platform decisions. The only input you own is whether the facts are accurate and complete.

Models use markup differently. Assembling an answer, a model decides whether a price is plausible, whether stock is current, and whether a store takes returns. Markup turns those from “read the copy and guess” into “check a claim.” A claim that checks out earns a citation; one that fails produces no error — the model quietly swaps in a source whose numbers agree.

Return lands on citation first, display second. That explains the familiar slide: coverage green, orders flat — the slot was measured while revenue moved elsewhere. Markup on one URL is only the entry ticket, because what gets compared is the entity behind the page, the argument in how AI reads your catalog as a graph.

Three layers, and the order they move in

Report conversion alone and schema loses credit to content and promotions; report citations alone and you own a vanity metric. Keep the layers apart, then read them together.

Layer	What to record	The question it answers	How it distorts
Citation	Citations on a frozen question set; cited fact type; shortlist entry	Will a model stake an answer on your data?	Brand terms, where you would be cited anyway
Click quality	AI-referred landing-page type; non-brand share; add-to-cart rate; pre-sales tickets	Do visitors reach the decision?	Organic brand search counted as AI impact
Conversion	Assisted conversion, first orders, returns, price-dispute tickets	Did it become revenue?	Last-click, AI referrals filed under Direct

Citation moves first, conversion last; between them sits one question the model asks on the buyer’s behalf — is this order safe to risk? Citation up with click quality flat means your name appears without a comparable fact. Click quality up with conversion flat points past markup, toward price, stock, or review totals that disagree across systems.

Four steps, eight weeks

Step	Action	Owner	Passes when
1 Hypothesize	Rewrite “schema helps” as “transaction fields raise price and delivery citations”	Growth	Metric, question type, and window named
2 Build controls	Same template, price band, traffic; randomize treatment and control	Growth and engineering	Pre-launch baselines are close
3 Split time	Four weeks baseline, four weeks after launch; log promotions, algorithm updates, paid changes	Analytics	Every disturbance sits on the timeline
4 Read the gap	Direction and stability over decimals; gap lands on predicted question types	Growth	You can explain why that type moved

Step 2 is the foundation: without controls you get “it went up after the change,” with five other things moving. Staggering beats one release — half the pages change this week and half in four, so seasonality and paid spend are shared. Restraint matters at step 4: with a few dozen pages, variance can exceed the effect, so “same direction for three weeks, only on price questions” beats a precise percentage. Change the transaction layer and nothing else; the field breakdown is in Product schema in 2026.

Three attribution errors that inflate the result

Crediting sitewide AI growth to schema. If content shipped, creators ran, or the category recovered seasonally, the sitewide curve sums all of it; only control pages split it.

Counting only pages that got a rich result. That drops failed samples before computing a success rate, so the pages that never earned the slot go unseen.

Using last-click. A buyer sees you in an answer, remembers the brand, and orders two hours later via brand search. The order is credited to search, so AI referrals stay undercounted and “organic has not moved” survives the budget meeting.

Example: 120 pages, staggered

Example — hypothetical store, no real brand or figures: a portable power station seller with a few hundred SKUs, one template, and a narrow price band. The team takes 120 pages from one price tier, changes half at random, and holds the rest four weeks.

Four fields change: gtin, priceValidUntil, shippingDetails, and hasMerchantReturnPolicy, plus aggregateRating aligned with the review count on the page. Nothing else moves.

Eight weeks in: the treatment group is cited more often on delivery and returns questions; the control stays flat. AI-referred sessions start landing on product pages, not the homepage. Pre-sales tickets on shipping and returns fall. Orders do not move.

Not a failure. Structured data fixes whether a model dares to recommend you and a buyer dares to order — not whether anyone wants the product, which is decided by product, price, and reviews. Write that down; it beats forcing a conversion lift into the slide.

Where this breaks

The site has a few dozen pages. Too few for controls; sample citation sources by hand.

The change fixes an error instead of adding lift. A wrong price or a permanent InStock is a defect; fix it without an ROI case.

Sales concentrate inside marketplaces. Off-site referral data is invisible there; track listing field completeness and update latency.

Attributes already disagree across systems. Markup would propagate the contradiction; settle master data first, as in fix product master data before GEO.

The number becomes an engineering KPI. The cheapest way to hit it is filling every field and rounding ratings up, until markup disagrees with the page — destroying the only thing schema sells.

This month, pull 40 pages from one template in one price band, split them at random, and add those four fields to half. Freeze 20 price-and-delivery questions today, record who gets cited, ask the same 20 in week four, and write the difference on one page. Reuse the AI visibility baseline for the questions; the rest of this line lives in measurement and growth.