How to Test Whether Product Information Changes a Business Outcome

Why the generic ROI shortcut fails

“Better product information” is not a treatment. It can mean a different label, richer fit information, an environmental disclosure, a repairability score, an alert, a product passport interface or a dozen other changes. Those interventions can alter different decisions through different mechanisms.

The controlled evidence makes the point.

A randomised trial of refrigerator energy-cost information tested a specific information intervention in a real retailer setting and found that product information can change search and purchase choices. That establishes causal potential for that treatment. It does not establish a general DPP conversion effect.

The Behavioural Insights Team work on France's repairability index includes randomised design testing and a large quasi-experimental evaluation. The later sales effect was overall positive but not statistically significant, with differences by channel. A null or uncertain result is part of the evidence, not something to edit out on the way to a business case.

A 2026 DPP-specific incentivised laboratory experiment found that education increased engagement and that engagement with favourable environmental information increased willingness to pay. The authors themselves frame the work as insight for field testing. It does not establish field conversion, revenue, margin, repair, resale or programme ROI.

Other product-information field experiments reinforce the need for treatment specificity. Research on fit information in online apparel retail found measurable effects on outcomes including conversion, order value and returns in that setting. A randomised returns intervention shows that returns can be measured causally with transaction data. Neither is evidence that a DPP produces the same result.

The practical implication is simple: do not ask “What is the ROI of product information?” until you can name what changed, for whom, compared with what and which outcome should move. The wider chain between information and a decision is owned by what has to be true before a passport changes a decision.

Maturity today

Step 1: define exactly what changes

The treatment should be specific enough that two people implementing the test would expose the same information.

“Add a DPP” is usually too broad. A passport may change the carrier, page layout, data fields, explanation, evidence, call to action and service flow at the same time. If the outcome moves, you will not know which part produced the effect.

A stronger treatment definition looks like this:

Treatment: on otherwise unchanged product pages, eligible visitors see an additional verified repairability explanation in a fixed position before purchase. Control: the same pages without that explanation.

The example is illustrative methodology, not a claim that repairability information will improve sales. The point is to isolate the information exposure.

Before launch, write down:

• the exact content or data shown;
• where and when the user can encounter it;
• which products and users are eligible;
• what the control experience receives;
• any other concurrent changes that would contaminate attribution.

If you cannot state the treatment cleanly, you are not ready to interpret the result causally.

Step 2: pre-specify one primary business outcome

Choose the outcome before looking at the data. Conversion, return rate, completed repair, recall response, resale listing or another observable event can all be legitimate, but they answer different questions.

Instrumentation makes many outcomes technically measurable. For example, Google Analytics 4's Measurement Protocol can receive server-side events with identifiers, timestamps and session context, while GA4 ecommerce measurement can record purchase transaction IDs, value and item data. Those capabilities help capture events. They do not establish that the information treatment caused them.

For the primary outcome, pre-declare:

If you are using scans as an exposure or intermediate event, use the estate's scan evidence page and the dedicated denominator methodology before converting server requests into a rate.

Step 3: create a comparison that can support causal language

A before-and-after chart is often useful operationally. On its own, it rarely tells you what would have happened without the information change.

Randomised comparison is the cleanest option where it is feasible. Eligible users, products, stores or other units are assigned to treatment and control in a way that makes the groups comparable by design. The refrigerator, fit-information and returns studies cited here demonstrate why this matters: the intervention is defined, a comparison exists and outcomes are observed against it.

Quasi-experimental designs can be credible when randomisation is not possible. Matching, difference-in-differences or other designs can construct a comparison from observed data, but the assumptions must be explicit. The repairability-index evidence is useful precisely because the evaluation does not pretend that every observed difference is a universal treatment effect.

Whatever design you choose, document threats that could move the outcome at the same time: seasonality, promotions, stock availability, channel changes, pricing, assortment, fulfilment and other relevant confounders.

Do not upgrade “we launched it and sales rose” into “it caused sales to rise” without a defensible counterfactual.

Step 4: fix the population, exposure rule and period

A treatment effect without its population is hard to reuse and easy to exaggerate.

State who could have received the treatment, who actually qualified for analysis and how exposure was defined. Then fix the observation window. A one-session conversion outcome and a 90-day return outcome cannot be analysed as if they share the same clock.

At minimum, preserve:

• eligible population: who or what could enter the test;
• exposure definition: what counts as having received the information;
• analysis population: exclusions applied and why;
• time window: when outcomes count;
• sample size and denominator: visible alongside percentages;
• unit of analysis: visitor, product, order, item, store or another declared unit.

This is where many connected-product claims break. A server can count requests. A repair platform can count completed jobs. A recall team can count responses. None of those totals becomes a meaningful rate until the eligible denominator is defined. The detailed cross-funnel method is in The Denominator Problem.

Step 5: interpret the result without making it bigger than the test

A well-designed experiment can support a local causal statement:

For this treatment, in this population, over this period, the measured outcome differed from the comparison by this amount under this design.

That is already useful. It does not need to become “product information increases conversion” or “DPP delivers X% ROI”.

Three disciplines matter after the analysis.

Report null and negative findings. The repairability-index evidence is a useful reminder that an expected positive effect may be statistically uncertain overall or vary by channel.

Keep secondary outcomes secondary. If you inspect many metrics after the fact, label them exploratory rather than pretending they were the original success criterion.

Do not transfer the effect automatically. A favourable environmental-information result in an incentivised lab experiment does not establish field conversion. Fit-information effects at one apparel retailer do not become sustainability-information effects. A returns nudge is not a product passport intervention.

This is the difference between building an evidence library and building a library of attractive numbers.

What a credible test record should contain

A programme should be able to reconstruct the result without relying on the presentation deck.

ActivateDigital methodology guidance: if a future team cannot tell exactly what the treatment and denominator were, the result should not enter a benchmark library.

What would change the evidence base

This page should change when materially better controlled field evidence appears, especially studies that test DPP-specific or comparable product-information interventions with clear populations, denominators and business outcomes.

More field studies would improve the evidence base. A universal benchmark would require something harder: comparable treatments, comparable outcomes and comparable denominator rules across programmes. Until then, “the average DPP uplift” is not an evidence-backed number.

Sources

Sources as at 4 September 2026.

• d'Adda, Gao & Tavoni / Nature Energy: randomised refrigerator energy-cost information trial, 7 April 2022. Used as controlled evidence that a specific product-information treatment can causally change search and purchase choices.
• Behavioural Insights Team: repairability index design and evaluation. Used for randomised design testing and quasi-experimental sales evidence, including the overall non-statistically-significant result and channel heterogeneity.
• Dorner, Tucker, Zhang et al.: DPP green consumer information experiment, 19 August 2026. Used as DPP-specific experimental evidence with explicit laboratory and treatment limits.
• Gallino & Moreno: The Value of Fit Information in Online Retail, 23 May 2018. Supporting context for measurable economic effects from a specific information intervention in a randomised field setting.
• INFORMS / Marketing Science: Smart Green Nudging, 8 August 2024. Used as evidence that returns outcomes can be tested causally with digital transaction data; not treated as a DPP effect.
• Google Analytics 4: Measurement Protocol reference. Used for instrumentation capability, not causal proof.
• Google Analytics 4: Measure ecommerce. Used for transaction outcome instrumentation, not causal attribution.