Self-Reported Attribution: Asking Customers How They Found You

A voluntary discovery question provides a customer’s recollection, not a complete exposure log or a causal estimate. Combine it with recorded source data while keeping unknown answers and nonrespondents visible. The useful outcome is a better hypothesis about discovery, not an automatic budget reallocation.

Ask one optional question

Illustrative prompt: “How did you first hear about us? Optional — you can skip this question.” Place it after a confirmed order and avoid making fulfillment depend on an answer. Open text and offered choices each have tradeoffs; neither guarantees unbiased recall. Use the same question version during a comparison.

A minimal response schema

Keep the survey optional and collect only what is needed under the applicable privacy basis. Tell customers how responses will be used. Do not request sensitive personal details. Limit access, redact free text where necessary and define retention. A response with an order ID should not be uploaded to an advertising platform merely because it is technically joinable.

Categorize without manufacturing certainty

Use an adjudication rule for ambiguous answers: two reviewers inspect a sample independently; disagreements remain unknown unless a permitted clarification resolves them. AI can suggest labels, but it must not invent a channel or reinterpret “Google” as an organic visit.

Reconcile 200 platform purchases and 120 responses

Synthetic month: 300 eligible paid orders, 150 answered, and 120 respondents mention Meta. Meta reports 200 attributed purchases under recorded settings. Response rate is 150 / 300 = 50%; Meta mention share among respondents is 120 / 150 = 80%. Neither 120 nor 200 proves what would happen without advertising. They answer different questions and may cover different events, cohorts or recall periods.

The eighty-count difference is not “sales that would happen anyway.” Inspect duplicated purchase events, return handling, nonresponse, first-discovery versus later assistance, channel confusion and reporting lag. Do not fill the 150 unanswered orders with the respondent distribution without a clearly labeled modeling assumption.

Use the insight as a test brief

Compare both source perspectives over equal matured order cohorts and keep count-weighted and revenue-weighted views distinct. A customer’s larger order can dominate a revenue share without increasing the number of customers. A controlled incrementality study is needed when causal contribution is the decision requirement; survey agreement alone is insufficient.

Prepare a creative hypothesis and chosen spend limit as a NovaStorm campaign draft and review it against mature net contribution. For example, test a message addressing a repeatedly mentioned discovery need while retaining the same offer. Survey storage, categorization and causal measurement remain separate from campaign preparation.