Reading affiliate platforms case studies: six gaps to check

A published case study about a partner program is written by whoever benefits from the conclusion. That does not make it false. It makes it unusable on its own, because the programs that went badly were never written up.

What follows is six findings that recur when somebody first examines a running program closely, and they are the six that affiliate platforms case studies tend to leave out. Each is described by mechanism and by the record that prevents it, so you can check it against a program you are already running — and against the influencer, creator and affiliate platforms an agency is deciding whether to keep or buy. None of the six needs a new tool; each needs a decision written down before launch.

What to take away

How to read somebody else's case study

Four things decide whether a published story about a program tells you anything, and each sits on the same chain: which route the order came through, which system counted the sale, when the click landed relative to the decision, and who owns the tracking domain.

Which system counted the sale is the one almost no published case carries, and it decides whether the story is about a program or about a reporting convention. What attribution does and does not establish is a modeling question rather than a measurement one, and the same caution applies to every channel report you have been shown.

The cases written by someone with no stake in the outcome read very differently. In the LeadClick Media affiliate network case, the U.S. Court of Appeals for the Second Circuit upheld a ruling requiring the network's operator to pay $11.9 million after finding that it recruited affiliate marketers who used fake news sites to drive traffic to the LeanSpa website. The FTC describes it as the first decision by a court of appeals holding the operator of an affiliate marketing network liable for deception by third-party marketers. No vendor case study would have recorded which partners sent the traffic and how; the court record did, and that is the standard to hold any program story to.

Finding one: the path nobody instrumented

The shape is always the same. A partner type appears to perform poorly and gets deprioritized. The real explanation is that a whole order route carries no tracking.

Phone orders, renewals, in-store redemptions and app purchases are the usual suspects. The missing route belongs to the part of the business added after the program launched.

You find it by placing a real order through each route yourself and looking for it, not by asking whether coverage is complete. The record that prevents it is a one-page list of every route a customer can buy through, with a named owner and the date each was last tested.

Finding two: two systems counting the same sale

Paid search and the partner program both claim a conversion, both report it, and nobody compares the two reports until a finance question forces it. This is not fraud and it is not a bug. It is two teams applying two defaults that were never reconciled.

The deduplication rule was never written down. Once written, it is boring. Until written, every quarter brings the same argument with the same two spreadsheets.

Agree it between channel owners before launch, put it where the reporting lives, and re-check it whenever a checkout changes. The habit of naming which system owns which fact belongs to the guide to measurement layers.

Finding three: the partner whose click lands after the decision

A partner with an unusually high conversion rate, treated as the program's best, gets a click seconds before checkout, after the customer decided.

That commission is a transfer from whoever did the persuading. Browser extensions, coupon aggregators and some cashback partners sit in this position by design. None of this requires anyone to behave badly.

The diagnostic is the gap between click and conversion, read as a distribution rather than an average. The remedy is a policy decision about what that partner type is worth, taken deliberately and applied consistently.

The variant that is genuinely dishonest, where a cookie is set for a visitor who never clicked, is cookie stuffing, and it looks identical in a conversion report. Telling the two apart is one of the narrow places where a detection product earns its price, which the guide to what detection can establish treats in more detail.

Finding four: the tracking domain that was never yours

This one only becomes visible at exit, which is why it is the most expensive. Links published in old posts, old videos and old newsletters keep sending traffic for years. If the tracking domain belongs to the vendor, all of that traffic goes on arriving at a system you no longer control.

Nothing about this is hidden. It is written in the setup and nobody reads the setup, because at signature everyone is thinking about launch. The record that prevents it is a single line in the implementation notes naming the registrant of the domain, checked in a WHOIS lookup rather than assumed.

Finding five: the fee line nobody read

Affiliate platforms meter differently, and the difference shows up in the invoice rather than the dashboard. Some bill a platform fee plus a per-event overage once you cross contracted event volume. Some quote a platform fee against tracked revenue with a separate line for managed services. Networks take a share of commission rather than a flat platform charge, sometimes with optional technology modules on top.

The terms behind those lines can move during a contract. Commission Junction's Publisher Service Agreement, filed with the SEC, lets advertisers change any payout rate on no less than seven days written notice, effective from the eighth day. A case study quotes the rate that applied when the story was written, not the one on next month's invoice.

The record that prevents this finding is a reconciliation date: one month of invoices checked against the contract before you sign the next term, with each metered line written as a named variable (platform fee, tracked revenue, overage rate, payment processing spread) so the arithmetic can be rerun when a rate changes.

Finding six: the data request that outlives the contract

A brand asks for the creator roster, the raw click log and the audience demographics. The agency sends a spreadsheet. Nobody records what was sent, to whom, or under what permission.

The FTC's Guides Concerning the Use of Endorsements and Testimonials in Advertising govern what a brief must require of a creator, and the FTC publishes them. State privacy statutes, including the California Consumer Privacy Act and the New York Stop Hacks and Improve Electronic Data Security Act, govern what personal data you may hold and for how long.

Neither question is settled on this page. Enforcement runs through the Federal Trade Commission, the California Attorney General and the California Privacy Protection Agency under the CCPA, and the New York Attorney General under the SHIELD Act. Route the specific question to a licensed attorney and to the regulator that covers your client.

The record that prevents it is a data map: what you hold, where it came from, who can see it, and the date it is deleted.

What the six have in common

None of them was a product defect. Each was a decision that nobody made, which then got made by a default. That is the real lesson available from case studies in this category, and it is why the useful artifact from a program review is a short written record rather than a slide — the same record an agency needs when it weighs influencer and affiliate platforms for a client.

Keep one page per program. It lists routes and last test date for each. It lists the deduplication rule and who agreed it. It lists partner policy by type. It lists domain registrant and fee reconciliation date. It lists data map owner. It lists the last incrementality test date if ever run.

An agency running several programs will find that page is also the fastest handover document it owns. That is the same argument made for the decision record in the guide to structured supplier comparison.

Common questions

Our program's reported revenue looks strong. Is that not evidence?

It is evidence that sales were attributed to the program. Whether those sales would have happened anyway is a separate question, and only a holdout answers it.

Should we run a holdout if the client will not like the result?

The result exists either way. A test you did not run does not make the revenue incremental, it only makes the claim unchallenged, which is a worse position when somebody eventually asks.

Is a vendor case study ever worth reading?

Yes, for the mechanics. Skip the outcome and read how the thing was set up, which partner types were used, and what the implementation involved. That part is usually accurate because it is not the part being sold.

We found one of these in our own program. What now?

Write it down before you fix it, including how long it had been true. The duration is what tells you whether the review needs to be annual or continuous.