By Diosh — Founder, AHAeCommerce | eCommerce decision intelligence for $50K–$5M GMV operators
Open your Meta dashboard, your Google Ads dashboard, and your Klaviyo report side by side for the same month. Add up the revenue each one claims. If you operate a $2M–$20M brand, that sum is almost certainly 120–150% of what your Shopify settlement actually deposited. This is a mistake piece for operators who believe multi-touch attribution gives them the true picture of which channels drive revenue — when in reality every platform's attribution is engineered to over-credit itself, and the only honest measure of what a channel does is incrementality testing. By the end you'll know why the numbers don't add up, why "better" attribution models can't fix it, and how to run a holdout test that tells you the truth in four weeks.
Why Your Channel Revenue Sums To More Than You Sold
Here is the arithmetic that breaks most operators the first time they do it. A home-goods brand doing $480K in a given month pulls its platform reports. Meta Ads Manager claims $310K in attributed revenue. Google Ads claims $220K. Klaviyo claims $190K. That is $720K of "revenue" against $480K of actual sales — 150% of reality. Nobody fabricated anything. Every platform is reporting honestly by its own rules. The problem is that the rules overlap on purpose.
The mechanism is double-counting, and it is structural, not accidental. A single customer sees a Meta ad on Monday, clicks a Google branded search on Wednesday, opens a Klaviyo abandoned-cart email on Friday, and buys. That one $140 order gets claimed in full by Meta (it ran an ad in the window), claimed in full by Google (it got the last paid click), and claimed in full by Klaviyo (its email was the last touch before purchase). Three platforms, one sale, three full credits. Multiply that across a month and the inflation compounds.
Each platform uses a different attribution window and a different crediting rule, and none of them deduplicate against the others — because they have no access to the others' data and no incentive to look for it. Meta defaults to a 7-day-click / 1-day-view window. Google Ads uses its own click-based and data-driven windows. Klaviyo attributes any purchase within a configurable window (often 5 days) of an email open or click. These windows are wide and they overlap heavily, so the same conversion lands inside multiple platforms' nets simultaneously.
The first discipline is to stop treating platform-reported revenue as additive. It is not. Three numbers that each describe the same underlying sales cannot be summed into a total. The moment you internalize that, the entire "which channel is winning" leaderboard you've been managing to becomes suspect. The same self-crediting distortion shows up in your blended efficiency math — I walk through how it corrupts payback calculations in the truth about customer acquisition cost.
Attribution Is Marketing Collateral From Your Ad Vendor
Stop thinking of an attribution report as a measurement. Think of it as a sales document produced by the company that wants you to keep spending. That reframe is the single most useful thing in this article.
Meta and Google are not neutral referees standing between you and your customer. They are vendors whose revenue is your ad spend. Their attribution systems are built, tuned, and defaulted by teams whose commercial incentive is to demonstrate that their platform deserves credit — and therefore more budget. This is not a conspiracy claim; it is an incentive observation. When the same party both runs the auction and grades its own performance, the grade trends generous. The IAB, the industry's own standards body, has documented for years how platform-reported, self-attributed conversions diverge from independently measured outcomes, which is precisely why third-party measurement and audited standards exist at all.
The post-iOS14 shift to modeled conversions
Before 2021, platform attribution was at least anchored to observed events — a real pixel fire tied to a real click. Apple's App Tracking Transparency framework broke that chain by stripping the identifiers platforms used to deterministically connect an ad impression to a downstream purchase. The platforms' response was not to report less; it was to estimate more. A large and growing share of the conversions in your Meta and Google reports today are modeled — statistically inferred conversions the platform fills in where it can no longer observe the actual user.
Modeling is defensible as a technique. The problem is who controls the model. The same vendor that benefits from a favorable number is the one choosing the modeling assumptions, and those models are not open to your inspection. You cannot audit Meta's conversion modeling the way you can audit your own Shopify orders. You are being shown an estimate, generated by an interested party, presented with the visual confidence of a hard count. Treat "modeled" the way you'd treat a contractor's estimate of his own quality of work.
The tell: revenue that doesn't move when reality does
The cleanest evidence that attribution is self-serving is what happens when you change nothing real but the reported numbers move anyway — or when you change something real and the reported numbers don't. Operators routinely see attributed ROAS jump after a platform "improves" its attribution model, with zero change in actual deposited revenue. The platform got more generous; your bank account did not. When the report and the settlement disagree, the settlement is the truth and the report is the pitch.
The Models Don't Fix It — They Relabel The Same Lie
The standard operator response to double-counting is to go shopping for a "better" attribution model. Last-click feels naive, so you reach for multi-touch: linear, time-decay, position-based (U-shaped), or a platform's "data-driven" model. This feels like progress. It is not. You are reallocating the same inflated credit across more touchpoints — you have not removed a single dollar of double-counting between platforms.
Every multi-touch model is a rule for splitting credit, not a method for measuring cause. Linear says "give every touch equal credit." Time-decay says "give recent touches more." Position-based says "give the first and last touch 40% each." These are opinions dressed as math. None of them can tell you the one thing that matters: would this sale have happened anyway without the touch? Academic work out of Northwestern's Kellogg School and others in the marketing-science field has repeatedly shown that correlation-based attribution systematically over-credits channels that simply intercept customers who were already going to buy — branded search being the canonical offender. The model sees the touch, so it assigns credit; it has no way to see the counterfactual.
This is why a brand can switch from last-click to a sophisticated data-driven model, watch the leaderboard reshuffle, feel more scientific, and make exactly the same budgeting errors. The reshuffling is movement, not insight. You've changed how the pie is sliced without ever questioning whether the pie is real. Branded search is the clearest case: it captures people typing your name into Google — people who already knew you and intended to buy — and every attribution model on the market will dutifully credit those conversions to "paid search," tempting you to scale a line item that is largely harvesting demand other channels created. I unpack that specific organic-versus-paid confusion in the real math on paid ads versus organic.
The deeper trap is that attribution measures touches, and touches are not causes. A channel can be present at the moment of purchase without having caused it. Attribution cannot distinguish "this channel created demand" from "this channel was standing nearby when demand converted." That distinction is the whole ballgame, and no crediting rule — however many touchpoints it spreads across — can recover it from observational data alone.
Incrementality: The Only Question That Pays Rent
There is exactly one question that determines whether a channel deserves your money: if I turned this channel off, how much revenue would I actually lose? That number is the channel's incrementality. It is almost never equal to its attributed revenue, and the gap is usually enormous.
Incrementality is the difference between two worlds: one where the channel runs and one where it doesn't. Attribution can never measure this because it only ever observes the world where the channel ran — it has no control group, no counterfactual. Incrementality testing manufactures the missing comparison by deliberately withholding the channel from a matched portion of your audience or geography, then measuring the actual revenue difference between the group that got it and the group that didn't. That difference is causal. It is what the channel truly drives.
What the gap actually looks like
When operators run real incrementality tests, the recurring finding is that true incremental revenue lands somewhere between 40% and 70% of what the platform's dashboard claimed — meaning 30–60% of "attributed" revenue would have happened anyway. Treat that 40–70% range as a directional benchmark drawn from published incrementality work and practitioner reporting, not a guarantee for your specific account; your number depends on brand maturity, channel mix, and how much demand-harvesting your spend is doing. Nielsen's long-running marketing-mix and media-ROI research, along with the geo-experiment and Conversion Lift methodologies Meta and Google publish, all converge on the same uncomfortable conclusion: a meaningful share of credited conversions are not incremental.
The pattern within the gap is predictable. Prospecting and broad-reach campaigns tend to be more incremental than the dashboard shows partial credit for, because they create demand other channels later harvest. Retargeting, branded search, and email-to-existing-customers tend to be far less incremental than claimed, because they intercept buyers who were already converting. This is why blended efficiency can look healthy while you quietly overspend on the channels that flatter themselves the most. Knowing the true shape of your channel economics is also what tells you where the ceiling is — the logic for that is in when to stop acquisition spend.
Why this beats every dashboard you own
A holdout test does not care about windows, view-through, modeling assumptions, or which vendor is grading its own homework. It compares real dollars in a treated group against real dollars in a control group. There is no crediting rule to argue about because there is no credit being assigned — only a measured delta. That is the difference between a number you can defend in a board meeting and a number your ad vendor handed you.
Run This Test On Your Biggest Channel In 30 Days
Here is the concrete action. Take your largest spending channel — for most brands at this scale that's Meta — and run a spend-down holdout. You don't need a measurement vendor or a data scientist to get a directionally honest first answer.
The geo holdout, step by step
Split your country into two matched sets of markets — comparable in baseline revenue, customer profile, and seasonality. For platforms that support geo-level controls, this is the cleanest design; Google's geo-experiment framework and Meta's lift tools exist specifically to support it. In your test markets, cut the channel's spend by 30% (or go to zero if you have the nerve and the budget headroom). In your control markets, hold spend exactly flat. Run it for four full weeks so you clear weekly purchase cycles and short-term noise.
Then measure the total revenue delta between test and control markets — from your own Shopify and bank data, not from the ad platform. If you cut Meta spend 30% in the test markets and total revenue there falls 12% relative to control, that 12% is the real incremental contribution of the spend you removed. If revenue barely moves, you just discovered that a large slice of your "attributed" Meta revenue was demand that would have converted regardless — and you've found budget you can redeploy or pocket.
If you can't do geo, do a time-based spend-down
Smaller operators without enough geographic volume can run a clean before/after instead: cut the channel 30% for two weeks, restore it for two weeks, and compare actual revenue across the periods while controlling for any promotions or seasonality. It is weaker than a true geo holdout because the world changes between periods, but it is still vastly more honest than reading a dashboard, and it costs you nothing but a temporary spend reduction.
What to do with the answer
Whatever delta you measure is your channel's real incrementality factor. Apply it as a haircut to that channel's reported revenue across all your planning. If Meta claims $310K and your holdout says the spend is ~55% incremental, plan against ~$170K of real contribution — and re-run your acquisition-cost and payback math on that honest figure, not the inflated one. Then repeat the test on your next-largest channel the following month. Within a quarter you'll have an incrementality factor for every major channel, and your budgeting decisions will rest on measured causation instead of vendor-graded estimates.
Build A Measurement Stack That Distrusts Itself
One holdout test is a wake-up call. A measurement discipline is what keeps you awake. The goal is a stack and a habit that treats every platform number as a claim to be verified, not a fact to be summed.
Anchor everything to your settlement data. Your Shopify orders and your bank deposits are the only ground truth in your business — they are the one set of numbers no vendor has an incentive to inflate. Every platform's attributed revenue should be reconciled against that total, with the explicit expectation that the platform claims will over-sum. When they sum to 140%, you don't panic; you apply your measured incrementality haircuts and reconcile down to reality. The architecture of a measurement stack that holds settlement as its source of truth is something I lay out in the eCommerce analytics stack.
Institutionalize incrementality as a recurring ritual, not a one-time project. Platforms change their attribution models, your channel mix shifts, and a brand's demand-harvesting ratio drifts as it matures — so an incrementality factor measured in Q1 is stale by Q4. Run a holdout on at least one major channel every quarter, rotating through your spend until each channel has a fresh causal read at least once a year. For larger brands, marketing-mix modeling — the discipline Nielsen and others have run for decades — is the natural complement to geo holdouts, because it estimates incrementality across all channels at once from historical variation in spend.
Finally, kill the leaderboard mentality on the rest of your reporting too. Attribution inflation is one species of a broader disease: optimizing to numbers that flatter the channel rather than describe the business. The same trap catches operators who chase reported conversion rate without asking what it actually predicts — I take that one apart in conversion rate as a vanity metric. The operators who win at this scale are not the ones with the most sophisticated attribution model. They are the ones who stopped trusting attribution and started measuring what actually changes when they turn a channel off.
The mistake this article exists to prevent is simple to state and expensive to make: treating your ad platforms' self-reported revenue as truth, summing it, and steering a real budget by it. The reports are marketing collateral from your vendors. The holdout test is the only document you wrote yourself. Trust the one you wrote.
