Restaurant Promotion Tests: Write the Brief Before You Launch

A delivery promotion can receive credit for an order without causing that order. To test whether an offer creates additional business, decide what will change, what will stay comparable, how exposure will be assigned, and how the result will be judged before the campaign starts.

For a multi-location restaurant team, the useful first deliverable is a test brief. It should make one decision possible: expand the offer, stop it, or gather better evidence. A dashboard total alone cannot make that decision.

Start with three different questions

• Attribution: Which orders did the platform credit to the promotion or ad?
• Economics: What remained after the relevant discounts, fees, commissions and operating costs?
• Incrementality: What would have happened without the change?

These questions need different evidence. Removing duplicate campaign credit improves accounting for attributed activity. It does not create the missing counterfactual, the outcome that would have occurred without the offer.

Platform measurement methods illustrate the distinction. Uber's April 2026 merchant update describes an offer study comparing customers exposed to offers with a holdout group. It separately describes randomized ad removal. DoorDash's October 2025 enterprise restaurant announcement discusses Ghost Ads measurement for Sponsored Listings. Neither source establishes that every restaurant has a self-serve promotion holdout control. Ask your platform contact about eligibility, the exact campaign type and the study design. Uber methodology | DoorDash restaurant measurement

Write this brief before anyone changes the offer

Use the following ten fields in your existing planning document. The filled examples are a planning illustration, not a recommended discount, budget or universal test duration.

1. Decision and hypothesis. State the decision the result must support. Example: “Should we expand one dinner offer to comparable locations?” Specify the offer, platform and intended customer population. Avoid testing a new discount, new menu, larger ad budget and new hours together if you need to isolate the offer.
2. Primary outcome. Pick one outcome before launch and define its unit. For example, completed marketplace orders per assigned store over the fixed test window. If profitability is the decision, add an explicitly defined contribution outcome. State whether canceled and refunded orders are included, and retain the source definition.
3. Treatment and comparison. Describe the exact change and the comparison condition. “Business as usual” needs a list of offers and ads that will remain active. A comparison store accidentally receiving the same new promotion is not an untreated control.
4. Assignment. Record who or what is assigned, how assignment is made, and who can change it. If a platform can run an eligible randomized customer-level holdout, obtain its method. For a store-level design, have the analyst assess whether enough comparable, independently assigned stores are available. Matching two stores without randomization is an observational comparison, not a randomized experiment.
5. Spillover check. Consider whether guests can switch between nearby test and comparison locations or see the same offer elsewhere. Grouping overlapping delivery areas may be more appropriate than treating each storefront as independent. Record the tradeoff and ask the analyst to account for the assignment level.
6. Dates and reporting cutoff. Fix the launch date, end date, store-local time zones and final data cutoff. Cover the relevant weekday and daypart pattern. Choose duration and sample requirements from baseline variability and the smallest commercially useful effect, not a universal “seven-day test” rule. Allow time for late cancellations, refunds and fee adjustments.
7. Cost boundary. Record who funds discounts, which fees and commissions apply, food and packaging assumptions, and any extra variable labor. Decide how ad and offer costs will be allocated if they overlap. Do not subtract a discount or fee again if the starting measure already deducts it. Label estimated contribution as an estimate, not net profit.
8. Operating guardrails. Define what would require a safety or service intervention: unavailable items, excessive cancellations, prolonged pauses or another brand-approved limit. A necessary intervention takes priority. Record it and explain its effect on the study rather than quietly deleting the affected days.
9. Analysis and uncertainty. Write the comparison method, treatment of missing data, exclusion rules and review threshold in advance. Report a range of plausible effects, not only one lift percentage. Analyze uncertainty at the assignment level; thousands of orders from two assigned stores do not create thousands of independent store assignments.
10. Owner and decision date. Name the person who authorizes campaign changes, the analyst who checks the result, and the person who makes the rollout decision. Preserve the original brief and log deviations. An inconclusive result can justify redesigning the test instead of expanding the offer.

NIST's experimental-design guidance explains why grouping comparable conditions and randomizing assignments serve different purposes. Similar-looking groups alone do not remove every alternative explanation. The restaurant brief above is an application of that planning discipline, not a claim that Voosh runs the experiment for you. NIST randomized block designs

A worked comparison: the extra 100 orders are not automatically 100 incremental orders

Consider two hypothetical, equally sized store groups observed over equal periods. Assume the order definition, included locations and reporting completeness stay constant.

• Test group: 1,000 completed orders before the change and 1,100 afterward, a gain of 100
• Comparison group: 1,000 completed orders before and 1,080 afterward, a gain of 80
• Difference in changes: 100 minus 80 equals 20 orders

The first comparison shows a 10% increase in the test group. The second shows why attributing that entire increase to the offer is unsafe: the comparison group also grew.

The 20-order difference is an arithmetic contrast. It is not, by itself, a proven campaign effect. Interpreting it causally requires a defensible design and assumptions about what the test group would have done without the change. Different local events, delivery availability, customer movement between stores or divergent prior trends could explain the difference.

This example supplies no randomization record, variance, sample-size analysis or confidence interval. It therefore cannot establish statistical significance or justify a rollout. It also contains no cost data, so it cannot establish profit. Its purpose is to show what the test brief must resolve before a team labels a result “lift.”

Decide what each result actually permits

Credited orders increased, but no valid comparison exists: Report attributed activity and campaign economics. Do not rename the result incremental growth.

The comparison is credible, but uncertainty is wide: Keep the decision open. The evidence may be compatible with both a useful gain and no worthwhile effect. Review sample requirements and whether a redesigned study is practical.

The planned analysis supports a useful improvement: Check the cost boundary and operating guardrails before expanding. Retain the tested population and conditions in the conclusion. A result from one platform, offer and market is not an all-store guarantee.

The test was disrupted: Explain what changed. If the intervention altered the question, label the result exploratory and plan the next decision accordingly.

Where Voosh fits in the review

Voosh's True ROI release describes store and campaign comparisons, customer segments and estimated contribution margin using food-cost assumptions and available commissions and fees. That can help teams inspect campaign economics and choose questions worth testing. It does not make attributed sales a causal lift estimate.

Use fair restaurant benchmarking to identify plausible comparison groups, then apply the additional experiment checks in this brief. Keep campaign authorization, study design and rollout decisions with the responsible people. Confirm supported platforms, permissions and any managed-service scope for your account.

Before your next offer launches, ask one question: “What result would change our decision, and what evidence would make us trust it?”

Book a Voosh demo to review the campaign data and cost assumptions available for your operation.