Post-Purchase Survey Attribution, Weighted and Reconciled

A checkout survey where 40 percent of respondents name a social platform is not telling you that platform drove 40 percent of orders. Response rates on post-purchase surveys run roughly 20 to 45 percent depending on placement, and the people who answer are not a random sample of the people who bought: they skew newer, younger, more mobile, and more likely to recall a channel with a strong creative signature than one with none. A raw survey percentage measures who answered. Handing it to a media team unweighted is how paid search ends up looking dead. What we do: We design the instrument, install it, and build the weighting layer that turns responses into an estimate about the order population. Placement first: the post-purchase page inside the checkout extension rather than a thank-you page overlay, because the overlay competes with the order confirmation and loses. One question, not a branch tree. Every additional question costs response rate, and the second question is rarely worth what it costs. Question wording is where most implementations fail before they collect anything. How did you hear about us, and what made you buy today, are different questions with different answers, and shops routinely ask the first while reporting the results as though they had asked the second. We choose one deliberately, usually first-touch discovery, because that is the thing no pixel can see; last touch is already well measured by click data. Option order is randomised per respondent and we measure the residual position bias so it can be corrected rather than assumed away. The I don't remember option stays on the list — removing it does not produce better data, it produces confident guesses. For Gulf accounts we write the instrument in Arabic and in English as two instruments, not one translated into the other. The channel list differs. In Saudi Arabia, Snapchat is a primary discovery channel rather than a secondary one; forwarding between family and friend groups on messaging apps is a real acquisition path that no measurement system captures; and an influencer and a friend sent it to me have to be separable options, because they imply completely different budget responses. An English-first list translated late surfaces none of this. The weighting model: Every response is joined to its order identifier and then to the order table, so responses can be weighted back to the full order population by device, new versus returning, order value band, region and acquisition day. For subscription businesses — most of our pet work is subscription — we weight on first order and then track the surveyed cohort's retention separately, because the channel that brings discovery volume and the channel that brings customers who are still there in month four are frequently not the same channel. Then we triangulate. Survey estimates are read next to whatever incrementality evidence exists, and where they disagree we investigate rather than average. A channel the survey loves that no holdout can detect is usually a recall artefact attached to a memorable brand campaign. A channel the survey never names that a test shows working is usually mid-funnel and invisible to the customer. Both findings move budget. Neither is available from the survey alone. What we refuse: We do not present survey shares as attribution percentages. We do not run this as a standalone measurement system, and if that is what you are buying, we are the wrong practice. We do not build branch trees past two questions. We do not install on accounts under roughly 600 orders a month; the weighting cells go empty and the estimate becomes noise with a decimal point attached.