Post-purchase surveys vs pixel data: use both, weight one

A brand ran a podcast sponsorship for two months and killed it, because the tracking showed eleven orders against a spend that needed a hundred.

Six weeks later somebody added one question to the order status page. Podcast came back as the fifth most named source, on orders that had all been recorded as direct. (Composite, drawn from post-mortems we have been walked through.)

Where the two signals disagree

Modelled on 1,000 orders in a single month, scored twice: once by pixel and resolved click data, once by survey responses scaled to the full base.

Channel	Pixel share	Survey share	Gap
Paid social	31%	18%	minus 13
Direct or unknown	30%	21%	minus 9
Email	19%	6%	minus 13
Search	14%	9%	minus 5
Creator or influencer	6%	17%	plus 11
Friend or word of mouth	0%	21%	plus 21
Podcast or offline	0%	5%	plus 5
Packaging insert	0%	3%	plus 3

Both columns sum to 100. Neither is inflating anything; they are describing different things.

Three rows carry zero in the pixel column and 29% between them in the survey column. Those are channels with no click surface at all, and no amount of tracking discipline will ever make them appear on the left.

What pixels are structurally blind to

A friend describing the product at a table. A podcast host saying the brand name out loud. A WhatsApp forward that gets retyped into a browser rather than clicked.

None of these produce a click, a session parameter, or a cookie. They produce a customer who arrives as direct traffic and gets recorded as demand that appeared from nowhere.

Cross-device is the fourth blind spot and the least discussed. A customer who discovers on a phone and buys on a laptop is a new visitor to your pixel and a continuing journey to the survey, because the person remembers what they saw even when the browser does not.

Email at 19% by pixel and 6% by survey is the reverse case and worth understanding. Email frequently touches the session in which an order happens and rarely creates the customer, so last-click pixel data credits it heavily while the customer does not name it, because from their side the email was a reminder rather than a reason.

What surveys are structurally bad at

Recall decays and it decays unevenly. People name what they saw most recently and what was most memorable, which systematically favours vivid channels over persistent ones.

Response rate is the second problem. A modelled 34% response rate on an unincentivised order status page question means every percentage above is measured on roughly a third of orders and scaled to the rest, and non-responders are not a random sample.

Free text is the third. Ask an open question and you get three hundred spellings of Instagram, needing manual coding every month by somebody who will eventually stop doing it.

And surveys cannot support campaign-level decisions. A survey tells you creators produced 17% of orders; it will never tell you which creator, at what cost, in which week, which is the question a budget meeting actually asks.

The weighting rule

State one rule and apply it to every report.

Pixel data leads for channels with a click surface. Paid social, search, email, affiliate, anything you bid on or link from. These have identifiers, campaign structure and cost data, and precision matters because money moves against them weekly.

Survey data leads for channels without one. Word of mouth, podcast, packaging inserts, offline. There is no competing measurement, so the survey is not being weighted against anything, it is the only instrument.

Creator sits in the middle and needs both. Six percent by pixel against 17% by survey is the gap between creator-driven sales that produced a trackable click and creator-driven sales that did not, and the fix is a better identifier rather than a choice between signals. That method is in tracking influencer sales without a discount code.

Applying that rule to the modelled month moves about 11% of orders between channels compared with using pixel data alone. That is a large enough shift to change a budget decision and small enough to be worth doing carefully rather than dramatically.

Running the survey without ruining it

One question. On the order status page. No incentive.

An incentive raises response rate and adds a population answering for the reward, which trades a measurable bias for an unmeasurable one. Three questions instead of one roughly halves completion in most implementations for information nobody uses.

A fixed list with a free-text other is the right shape. The list keeps the data codeable; the other field is where you discover the channel you did not know existed, which is the reason the survey exists at all.

Review the other responses monthly and promote anything that appears repeatedly into the list. A survey that never changes its options stops finding new channels after the first quarter.

Reporting the blend honestly

Publish three things together, always: the pixel number, the survey number, and the response rate.

Reporting a blended figure alone hides which signal produced which row, and six months later nobody can reconstruct it. The response rate belongs on the same page as the chart so that nobody reads a survey percentage as a census.

Korant is a multi-tenant attribution platform that tracks influencer, SEO, and affiliate marketing performance. Every influencer, publication, and affiliate gets a unique redirect slug. Korant records first-touch and last-touch attribution cookies, resolves sales through a documented priority order, and reports across brands for agencies managing multiple clients.

Korant is not for stores with a single paid channel, or for brands that only need Shopify’s native reports. Running one paid channel, the survey alone tells you most of what you need, and there is no resolution conflict to document.

The survey response also belongs on the order record itself rather than in a separate tool. A survey answer that cannot be joined to an order cannot be cross-checked against the pixel data for the same order, and that cross-check is where the useful disagreements surface, which is the structure described in building a single source of truth.

What neither signal answers

Both measure presence. Neither measures cause.

A customer naming a friend, and a pixel recording a paid social click, can both be true of the same order, and both can be irrelevant if that customer was buying regardless. Only a holdout answers the causal question, and neither instrument here substitutes for one.

The honest framing is that the survey exists to stop you cancelling channels you cannot see, and the pixel exists to stop you scaling channels you cannot cost. The podcast in the opening was killed on a precise number that was measuring the wrong thing, and no improvement in pixel accuracy would have saved it.