How to read offline campaign measurement benchmarks: a method, not published values

What to take away

Benchmark each layer against a comparable mechanism and definition.
Publish raw counts, denominators, windows, and uncertainty.
A high response rate can coexist with weak incrementality or poor economics.

Offline campaign measurement benchmarks can include on-time delivery, usable placement, estimated exposure, response, qualification, appointment, purchase, adoption, retention, complaint, waste, incremental margin, and cost. None is portable without the channel, audience, offer, place, period, and rule that produced it.

This article gives a method, not benchmark values. The numbers shown below illustrate the form of a card; they are not targets to copy.

Use a layered benchmark card

Layered benchmark card

Layer	Numerator	Denominator
Delivery	Units delivered under rule	Eligible units entered
Availability	Units or locations usable	Scheduled units or location periods
Exposure	Credibly exposed units	Eligible audience opportunity
Response	Valid campaign responses	Defined delivered or exposed base
Qualification	Responses meeting criteria	Valid responses
Mature outcome	Completed valuable outcomes	Qualified people or eligible audience
Incremental effect	Estimated outcomes caused	Comparable eligible population
Net value	Incremental value less cost	Campaign, customer, or unit basis
Guardrail	Adverse events	Relevant people, actions, or periods

A filled card for one layer looks like this:

Response layer example

Layer: Response
Numerator: 1,840 valid responses
Denominator: 24,000 delivered units
Window: 6 weeks from first delivery
Rate: 7.7%
Interval: 7.3% to 8.0% (95%, binomial)

The interval is narrow because the denominator is large. Group the same units by store or carrier route and the interval widens according to the design effect, which the analyst has to estimate and state.

GAO's 2012 guide to designing evaluations covers evaluation questions, constraints, designs, data collection, and analysis for federal programs. It supplies no marketing targets, but it supports choosing evidence from the decision rather than copying an outside percentage.

Include interval and sensitivity

NIST's statistics handbook introduces confidence intervals in a statistical reference context. A valid interval depends on design and assumptions; it does not repair selection bias, poor event definitions, uncontrolled spillover, or an invalid comparison.

Show numerator and denominator beside every rate
Label observed, modeled, attributed, and incremental values
Separate new, existing, and unknown customers
Wait for outcomes to mature and subtract reversals
Report match rate, missingness, and contamination
Compare value with full lifecycle cost and guardrails

Build internal reference classes by mechanism, audience, offer, geography, season, distribution quality, outcome window, and design. Preserve definition changes and series breaks. An industry average can support a planning question, but it should not become a target without compatible definitions and business economics.

Use both a decision threshold and a practical-importance check. A statistically distinguishable change may still be too small, costly, delayed, fragile, or uneven to justify scale. A large estimate with wide uncertainty may require another test rather than immediate expansion.

Make the comparison reproducible

The GAO data reliability guide treats reliability as fitness for an intended use and requires documented assessment. Use that test for offline campaign measurement benchmarks; the federal guide does not certify the local data.

The NIST experimental design selection guidance begins design choice with the objective and practical constraints. It supports separating offline campaign measurement benchmarks reporting from controlled effect estimates, not turning observation into causation.

State the benchmark population, period, inclusion rules, calculation, currency, maturity window, and uncertainty before comparing results. Keep market context separate from an internal target. Recalculate the figure from the cited source or governed dataset, then record any mismatch. A benchmark becomes useful only after the local team explains why the comparison group and method fit the decision.

For offline campaign measurement benchmarks, keep the evidence record beside the decision. A reviewer can then reproduce the reasoning without relying on memory.

Practitioners in the United States and Canada can build comparable reference classes from their own delivery, response, and outcome records. Industry associations, syndicated research providers, and national postal operators publish pooled figures and service standards, which work as an outside check only when the definitions match.