The monthly report lands, and it says what it said last month. Branded search is the top revenue channel. Retargeting sits second. The prospecting campaigns you spent a quarter building sit near the bottom with a return that barely justifies the spend. So you trim prospecting and move the money into brand.
New customer revenue goes flat two months later, and nobody can point to the reason. The report didn't fail. It answered a different question than the one you were asking. It told you which touchpoints appeared before purchases. It never told you which touchpoints caused them.
That gap is where most budget mistakes start, and a better dashboard won't close it. Knowing which rule your report uses, what each rule can and can't see, and where to stop trusting all of them and run a test instead will.
An ecommerce attribution model is the rule that decides how much credit each marketing touchpoint gets for a sale. The sale is one event. The credit is an allocation, and the model makes that allocation repeatable.
Think about how a table splits a restaurant bill when four people claim they arranged the evening. One booked the table, one chose the restaurant, one drove, one paid. Every model below is a different answer to that argument, applied to marketing.
Single-touch attribution gives 100% of a conversion's credit to one interaction, usually the first or last click. Multi-touch attribution spreads credit across several interactions, such as paid search, display advertising, email, social media, marketplace visits, and direct traffic. The shift from one rule to many followed the fragmentation of the buying journey itself: a 2025 academic review of multi-touch attribution modeling screened 345 publications dated 2011 to 2024 and narrowed them to 62 original empirical studies, split across 29 conference papers and 33 journal articles.
Attribution Models at a Glance
That research volume has a practical edge for a brand selling on Amazon, eBay, Walmart, and its own D2C store. A last-click report can conceal the discovery, consideration, and retargeting interactions that made the sale possible, and the seller reads the concealment as a channel verdict.
The six rules below cover nearly every report you'll meet in a marketing dashboard. Read across each row for the rule, the bias, and the blind spot.
| Model | Credit rule | Tends to favor | Main blind spot |
|---|---|---|---|
| First-click | 100% to the first recorded touch | Awareness and discovery channels | Ignores what closed the sale |
| Last-click | 100% to the final touch before purchase | Branded search, retargeting, marketplace search | Hides every earlier touch |
| Linear | Equal share to every touch | Long journeys with many small touches | Treats a minor touch the same as a decisive one |
| Time-decay | Larger share to touches closer to purchase | Short consideration cycles | Underweights early discovery |
| Position-based | Typically 40% first, 40% last, 20% spread across the middle | A balanced view of discovery and close | Fixed split ignores actual influence |
| Data-driven | Fractional credit estimated from converting and non-converting paths | Accounts with enough conversion volume | Needs large data sets and is hard to audit |
Last-click is the default in most platforms, which is why it becomes the reference point everyone argues against. It isn't wrong so much as narrow: it answers "what closed the sale?" and stays silent on every other question.
Each model answers a different business question, so the right one is the one that matches the question in front of you, not the one your tool ships with. When you're ready to look past credit entirely, the wider set of digital advertising performance metrics shows what else belongs on the same page.
How Each Model Splits Credit on One Journey
Take one order and run it through each rule. Say a shopper spends $120 after five touches: a Meta video on day 1, a Google Shopping click on day 4, a brand email opened on day 9, an Instagram ad on day 12, and a branded search that ends in a purchase on day 14. Real paths are messier and usually include a direct visit or two, but this one is enough to watch credit move.

The table splits that same $120 five ways. The time-decay column uses a 7-day half-life, so a touch earns half as much credit for every seven days it sits before the order; those values are rounded and sum to about $120.
| Touchpoint | First-click | Last-click | Linear | Time-decay | Position-based |
|---|---|---|---|---|---|
| Meta video (day 1) | $120 | $0 | $24 | $10.76 | $48 |
| Google Shopping click (day 4) | $0 | $0 | $24 | $14.49 | $8 |
| Brand email (day 9) | $0 | $0 | $24 | $23.77 | $8 |
| Instagram ad (day 12) | $0 | $0 | $24 | $31.99 | $8 |
| Branded search and purchase (day 14) | $0 | $120 | $24 | $38.99 | $48 |
Reading the table row by row
First-click pays the entire $120 to the Meta video. Everything after day 1 earns zero, which is only true if that video alone made the sale. Last-click does the reverse and pays the branded search, the cheapest touch in the sequence.
Linear gives $24 to each. That's fair to the journey and unfair to the email, which sat five days out and carried the click that reopened the cart.
Time-decay pays $10.76 to the video and $38.99 to the branded search. Its logic is that recency signals influence. Its blind spot is a long consideration cycle where discovery did the persuading.
Position-based applies 40/20/40: $48 to the video, $8 to each middle touch, $48 to the search. It's the only rules-based model here that pays the opening and closing touches equally while still acknowledging the middle.
Data-driven produces no fixed row at all. Its split comes from comparing thousands of converting and non-converting paths, so the number depends on your data and can shift between quarters without anything changing in the campaign.
The point of the exercise: the model doesn't describe your customer. It describes a rule you chose, applied to your customer.
One journey tells five stories. A seller reading only the last-click row sees branded search as the hero, cuts the Meta video that started the sale, and then wonders why prospecting stops working. The same argument runs per creator when you pay for posts, and this per-creator ROI breakdown for DTC brands works through how to compare them without double counting. Find out which row your current report follows before you sign off on any budget.
Then group the channels themselves into a paid, owned and earned media structure, because each group earns credit differently and a single rule rarely serves all three.
Why Attribution Misleads and How Incrementality Checks It
Every model above shares one defect: none of them can say whether the sale would have happened without the ad. Each one observes the path a buyer took and divides credit among the touches on it. A click proves contact. Proving the click caused the purchase is a separate question, and no allocation rule answers it.
Consider the shopper who types your brand name plus "reviews" into Amazon and buys through the sponsored brand ad at the top of the results. That person had already decided; the ad intercepted a purchase in progress. The model can't see the decision, only the click, so it books the sale as advertising's work and the budget follows.
This isn't a modeling flaw you can fix with a better rule. It's a statistical one. Buyers who see or click an ad are often more likely to purchase in the first place, which puts correlation in the seat that belongs to cause. The gap between credited and causal effect can be wide, so any credit split that drives a budget move should be checked against a controlled experiment before the spend shifts.
Where the causal story breaks
Three patterns repeat across ecommerce accounts, and each one inflates a channel that isn't doing the work.
- Retargeting claims conversions it followed. Someone who viewed a product page and bought four days later was already in motion. Retargeting appears at the end of those paths and collects the credit.
- Identity breaks at the channel boundary. A shopper who discovers on Meta, researches on Walmart, and buys on your D2C site looks like three unrelated people when each platform only sees its own slice.
- One order gets claimed more than once. Search, paid social, and retail media can each report the same purchase, so adding their totals produces a number larger than revenue.
Three questions that need three tools
Confusion usually comes from asking one method to answer all three questions. Attribution describes observed paths. Experiments estimate causal lift. Media-mix modeling estimates how much each channel contributes over months.
| Method | Question it answers | Setup effort | Main limitation |
|---|---|---|---|
| Attribution | Which observed touches came before the sale | Low, runs continuously | Splits credit among correlated touches and rewards existing intent |
| Geo holdout test | Did regional spend add sales compared with matched regions | Medium, needs matched regions and four to six weeks | Noisy in small markets |
| Conversion-lift test | Did exposed users buy more than a control group | Medium, built into some ad platforms | Covers only the platform running it |
| Media-mix modeling | How much each channel contributes over months | High, needs months of weekly data | Coarse and slow to react |
Disagreement between tools is itself diagnostic. Measurement guidance published for 2026 reports that platform-reported ROAS and blended revenue-based ROAS routinely differ by 20% to 40% (Silverback Strategies). A channel dashboard is one interested party's account of the sale, and treating several of them as independent evidence is how budgets drift toward whoever claims loudest.
Start with one test you can afford
A brand without a research budget can still run a real experiment. Pick the channel whose credit swings most between models, reduce or pause it in matched regions for four to six weeks, and compare sales against control regions that kept spending. Matched regions work like a control group in a trial: similar shoppers, same season, same promotions, with one variable removed.
What happens next is the answer. A channel that mostly intercepts buyers already on the way shows almost no change when it goes dark. A channel that creates demand leaves a visible gap, and the size of that gap is your estimate of incremental revenue, the figure a platform report never isolates.
Attribution tells you where to look. Only a test tells you whether to spend.
That gives a workable division of labor: attribution for daily diagnosis and optimization, a holdout test before any large budget shift, and media-mix modeling for quarterly allocation. The model becomes your diagnostic layer, and measured lift becomes the approval gate for spend. If you're building the reporting structure that holds all three, our guide to measuring marketing effectiveness covers where each number belongs.
Choosing a Model for Your Data and Channel Mix
No model is better in the abstract. It's better or worse for your volume, your channel count, your purchase cycle, your consent situation, and the decision in front of you.
If you want the mechanics walked through before you pick, this walkthrough covers the same ground visually.
Five criteria and the model each one points to
- Conversion volume. Data-driven attribution is a supervised, path-level model rather than a fixed allocation rule. Google Ads compares converting and non-converting paths across Search, Shopping, YouTube, Display, and Demand Gen campaigns and assigns fractional credit, but it recommends at least 200 conversions and 2,000 ad interactions within 30 days for supported networks (Google Ads Help). Below that, aggregate conversion actions, lengthen the evaluation window, or stay with a rules-based model.
- Channel count. A single-channel store can read last-click next to the platform's own numbers. A funnel that runs from Amazon to a brand site to paid social needs a path-level view plus a holdout, because no single platform sees the whole shopper.
- Purchase cycle. Impulse buys with same-day decisions suit time-decay. Considered purchases with several research visits suit linear or position-based, where early discovery keeps its share.
- Privacy and consent. Lost consent, iOS measurement limits, and modeled conversions leave paths incomplete. Industry analysis of ecommerce app marketing found remarketing and re-engagement mechanics accounted for 27% of attribution-related questions, while SKAN and iOS privacy accounted for another 16%, or 43% combined (AppsFlyer). A higher match rate doesn't automatically mean better attribution, so favor consented first-party data strategy and estimates validated by lift.
- Decision use. Reporting a channel mix needs a stable, consistent model that doesn't move month to month. Moving budget needs incrementality evidence, because the model can't supply it.
When you do compare channels, check that the return formula matches on both sides. A walkthrough of how Chartsy calculates ROAS is a useful reference for spotting where two platforms are measuring different things under the same label.
Judge a model by whether it improves budget decisions and predicts performance on data it hasn't seen. Not by whether it hands upper-funnel media a bigger share.
Set Up, Calibrate and Confirm Before Moving Budget
The two methods answer different technical questions. Attribution estimates how conversion credit is distributed among observed touchpoints. Incrementality estimates causal lift, meaning the purchases that would not have occurred without the marketing exposure. Google's own guidance on holdback experiments in data-driven attribution describes testing in service of the model, not instead of it.

The seven checks before a budget shift
- Inventory every touchpoint and confirm each one is tracked. That means paid search, paid social, email and SMS, marketplace ads, organic traffic, affiliates, and offline channels. The common slip is a channel missing from the tag setup, which shifts credit to whichever channel is measured properly. For brands selling on Amazon, eBay, and Walmart, our marketplace services can help: product listing optimization, marketplace SEO and keyword strategy, inventory and fulfillment management, Enhanced Brand Content (EBC) and A+ Content, and multi-marketplace expansion all create touchpoints a web-only setup never records.
- Pick a primary model for daily reporting and a second one for comparison. Position-based beside last-click is a common pair, because the gap between them shows you how much of a channel's value is model choice rather than fact. Reading only the default is how teams end up defending a number nobody can explain.
- Set the lookback window to your purchase cycle. Seven to thirty days covers most ecommerce paths. Overlapping windows across tools double count the same order, so write the window down next to the model name.
- Combine low-volume conversion actions until you clear the sample threshold. Splitting purchases across several small conversion actions starves a data-driven model and produces output that looks precise and isn't.
- Run a holdout or geo test on your largest channel for four to six weeks. Skip this and the true lift stays unknown, no matter how clean the attribution report looks.
- Report attributed ROAS beside incremental ROAS with confidence intervals. Confirm each platform's ROAS formula matches before you compare them, since nonstandard retailer formulas make cross-platform numbers incomparable. Our performance reporting is built around showing what's driving sales rather than restating platform totals, and that distinction is the whole point of the exercise.
- Shift budget only when the lift holds, then recheck the model each quarter. Channel mix, privacy signals, and customer behavior all move, and a model calibrated in January can drift by summer.
The 30-day starting plan
Export your current channel report and write the model name at the top of it. That single act resolves most arguments.
Add one alternative model as a second column so you can see which channels swing. Then run one incrementality test on your largest channel, and hold the budget until it reports.
Confirm three things before any move: the model is documented, the volume clears its threshold, and the holdout agrees with the direction attribution suggests. When all three line up, you have earned the right to reallocate. When they don't, you're guessing with extra steps.
The decision rule that keeps you out of trouble
A working setup is one where attributed and incremental results point the same way within a stated margin. Anything else is a signal to test, not to spend.
Watch for the reshuffle that happens without lift evidence: money migrating toward upper-funnel or retargeting channels because a model change made them look stronger. Our custom reporting solutions exist to show what's driving a client's sales, and attribution belongs inside that picture as a diagnostic input, not as the verdict.
Attribution is worth doing well, and it's worth stopping at the right point. Next Point Digital builds the measurement setup around your actual channel mix, from Amazon, eBay, and Walmart listings through paid search, social, and email, then ties the reporting to decisions you can defend. If your reports currently disagree with each other and nobody can say which one to trust, visit Next Point Digital and we'll walk through your model, your windows, and the one test that would settle the argument.