The Cannibalisation Question: Does DSP Steal Your Own DTC Sales?
Sometimes, yes — and the honest response isn't a reassurance, it's a test. A geo-lift comparing total combined revenue (Amazon plus your own site) in markets running DSP against matched markets where it's paused is the direct way to measure whether the spend is genuinely additive or partly redirecting demand your own site would have captured anyway.
What this looks like in a real account
What makes this question harder than a standard incrementality test
A standard DSP incrementality test asks whether the spend caused a sale versus no sale at all. The cannibalisation question is subtler: it asks whether the spend caused a sale on Amazon versus a sale that would have happened on your own site instead — both are real revenue, just at different margins and with different customer-ownership value attached. That distinction is why the outcome metric has to be total combined revenue rather than either channel's revenue alone; measuring only one side of the ledger structurally can't distinguish genuine incrementality from a same-total-revenue shift between destinations.
Why this is a legitimate concern, not a question to wave away
A DSP campaign that reaches a shopper who was already going to buy from your own site — through a display ad that happens to remind them, or a retargeting impression that catches someone mid-checkout on Amazon instead — and redirects that purchase to Amazon instead of your own site isn't growing revenue, it's moving it, while adding Amazon's referral fee to a sale that would have happened at full margin on your own site regardless. This is a real mechanism, not a hypothetical, and any brand asking the cannibalisation question deserves a real test, not a reassurance that it doesn't happen.
The test: a geo-lift measuring total combined revenue
The critical design choice is measuring the right outcome. A geo-lift that only checks Amazon sales in test versus control markets will show DSP "working" almost by definition, since that's the sale it's built to drive — it can't detect cannibalisation because it isn't looking at the channel the sale might have been stolen from. The correct design measures total combined revenue — Amazon sales plus your own site's sales — in matched test and control markets, with DSP running in test markets and paused in control. If DSP is purely additive, total combined revenue in test markets should rise relative to control by roughly the incremental Amazon sales DSP drove. If cannibalisation is happening, the combined-revenue lift will be smaller than the Amazon-only lift, because some of what looks like new Amazon revenue is revenue your own site lost.
A worked example of what cannibalisation looks like in the numbers
Take 4 test markets and 4 matched control markets, tracking within 3% of each other historically. Pause DSP entirely in control markets, run it normally in test markets, for 4 weeks. Test markets show $210,000 in Amazon sales against a control-implied $165,000 — a $45,000 apparent DSP-driven lift on Amazon alone. Now check total combined revenue: test markets show $520,000 combined (Amazon plus site) against a control-implied $500,000 — only a $20,000 combined lift. The gap between the $45,000 Amazon-only lift and the $20,000 combined lift — $25,000 — is the cannibalisation: sales that moved from the brand's own site to Amazon rather than being genuinely new. In this example, DSP is still net-positive ($20,000 of real incremental revenue) but roughly 56% less valuable than the Amazon-only view suggested.
Why the honest answer is rarely all-or-nothing
Most real results land somewhere between "fully additive" and "fully cannibalising" — some share of the apparent Amazon lift is genuinely new demand, some share is redirected from the brand's own site. The test above quantifies where on that spectrum your specific account actually sits, which is a materially more useful answer than a yes/no verdict, because it tells you whether to scale, hold, or restructure the spend toward the destination-mix strategy covered elsewhere on this site rather than simply cutting DSP entirely on a worst-case assumption.
The common mistake, including ours
The mistake — in both directions — is answering the cannibalisation question with an assumption rather than a test. We've seen an agency (not us, though we've been guilty of the milder version of this) confidently assert DSP spend was "purely incremental" to a DTC-majority client without ever running a combined-revenue test, because the Amazon-only attributed numbers looked strong and nobody wanted to complicate a good story. We've also seen a brand assume, with equal confidence and equal lack of evidence, that any DSP spend must be cannibalising their DTC business and refuse to test it at all. Both are the same underlying error — a strongly held position with no test behind it.
What to do once you have a real answer
If the test shows meaningful cannibalisation, don't necessarily cut DSP entirely — consider shifting more of the spend to link-out campaigns pointed at your own site instead of Amazon-destination creative, which captures the same audience-targeting value while routing the resulting sale to the channel it would have gone to anyway, at your own site's margin rather than Amazon's referral-fee-adjusted one. If the test shows the spend is largely additive, that's the evidence worth carrying into a budget conversation instead of the unvalidated attributed ROAS number — it's a stronger, tested claim.
| What you measure | What it tells you | Risk if used alone |
|---|---|---|
| Amazon-only attributed ROAS | How much Amazon sales DSP appears to drive | Can't detect cannibalisation from your own site |
| Total combined revenue geo-lift | Whether the spend is genuinely additive across both channels | Requires more setup — matched markets, longer test window |
| Gap between the two | The cannibalisation rate specifically | None — this is the honest number to act on |
Which one you should actually pick
Any brand with revenue across enough distinguishable markets can run this exact combined-revenue geo-lift themselves — it's the same method as any geo-lift test, applied to the right outcome metric. reMKTR runs the cannibalisation-specific version of this test for any client selling in both places before recommending a destination split, having seen both an over-confident 'purely incremental' claim and an over-cautious 'must be cannibalising' assumption turn out wrong without a test, as part of the discipline behind Full Circle's $500M+ in managed Amazon spend across 100+ brands.
Shortlist on the job, not the feature grid. Pull your search-term report for the last 90 days and total the spend against terms that produced no orders — 33.6% on the account above. Then ask each vendor on your list what they would do about it in week one, and see who answers with a process rather than a screenshot.
Common questions
How do I know if my Amazon DSP spend is cannibalising my DTC sales?
Run a geo-lift test measuring total combined revenue — Amazon plus your own site — in matched test and control markets, not Amazon sales alone. The gap between the Amazon-only lift and the combined lift is your cannibalisation rate.
Does cannibalisation mean I should stop running Amazon DSP?
Not necessarily — even meaningful cannibalisation often still leaves a net-positive result. Consider shifting more spend to link-out campaigns pointed at your own site instead of cutting the spend entirely.
Is cannibalisation always a bad outcome?
It reduces the spend's real incremental value, but it's not automatically disqualifying — the honest response is quantifying it and adjusting destination strategy, not treating any cannibalisation as proof the spend should stop.
Can I test for cannibalisation without pausing spend in any markets?
Not reliably — a clean geo-lift requires a genuine control condition, which means pausing spend somewhere. A before-and-after comparison without a control group can't separate cannibalisation from normal demand variation.
We show the method before the number.
Claim the free auditRead next
- Skai Pricing: Five Published Tiers, Read CloselyPricing · skai pricing
- Criteo vs Amazon DSP vs reMKTR: Which Demand You BuyHead to head · criteo vs
- Pacvue Pricing: No Public Number — What to AskPricing · pacvue pricing
- Skai vs Pacvue: Contracts, Not Feature GridsHead to head · skai vs pacvue