Attribution Models, Honestly: From Last-Click to MMM
Last-click, linear, data-driven, MMM: what each attribution model really credits, where each one lies to you, and which to run your business on.
By The Spend Report Editorial Team. Published June 22, 2026. · 6 min read
On this page
Open three reports for the same week and you get three different stories. Meta's dashboard claims 180 purchases. GA4's data-driven model credits Meta with 90. Your Shopify checkout survey says 40 people heard of you "on Instagram," and your bank account grew by an amount that fits none of those numbers. Nobody is lying on purpose. Every one of those tools is running a different attribution model, and each model decides who gets credit for a sale by a rule that was never going to match reality exactly.
The mistake operators make is treating this as a measurement problem with a correct answer somewhere out there. It is not. Attribution is a budgeting decision dressed up as math. The useful question is not "which model is true," it is "which model do I steer the business with, and which lies can I live with."
What an attribution model actually does
Every model takes one sale and splits the credit for it across the touchpoints that came before. A touchpoint is any ad click, email open, or organic visit the tracking saw. The model is just the rule for slicing the credit. That is the whole job.
The catch is that the model can only credit what it observed. If someone saw your TikTok ad, ignored the click, searched your brand on Google two days later, and bought, most setups credit Google. TikTok did the work. Google got paid. This is not a bug you can configure away. It is the structural limit of click-based attribution: it rewards the channels that sit closest to the cart and starves the ones that create demand upstream.
The models, side by side
Here is the honest version of the menu. The figures throughout this piece are illustrative, meant to show the shape of the distortion, not a precise claim about your account.
| Model | What it credits | Blind spot | Best for |
|---|---|---|---|
| Last click | The final touch before purchase | Erases every channel that built demand | Quick reads on high-intent search |
| First click | The first touch that started the journey | Erases the channels that closed the sale | Judging top-of-funnel reach |
| Linear | Every touch equally | Treats a throwaway impression like a closing email | A neutral default when you have no priors |
| Time decay | Recent touches more than old ones | Still favors lower-funnel, just less harshly | Short consideration cycles |
| Position-based | 40/20/40 to first, middle, last | The split is a guess dressed as a rule | Balancing discovery and closing |
| Data-driven | Modeled probability per touch | A black box that needs volume to be stable | Accounts with enough conversions to model |
| MMM | Channel-level spend vs revenue over time | Coarse, slow, and needs a year of data | Whole-budget allocation decisions |
Read that table as a spectrum, not a ranking. The single-touch models (first, last) are blunt instruments that are honest about being blunt. The multi-touch models (linear, time decay, position-based, data-driven) try to spread credit more fairly, but every one of them still only sees clicks, so they all inherit the same upstream blindness. Marketing mix modeling sits in a different category entirely, and we will get to why that matters.
How the same week looks under different models
Take one DTC brand, one week, one set of sales. Nothing about the actual spend changes. Only the model changes. Watch what happens to each channel's share of credit.
Same week. The numbers are representative, not measured, but the pattern is real and it shows up in almost every account: as you move from last-click toward MMM, credit drains out of the channels that sit next to the purchase (brand search, email) and flows toward the channels that create the demand in the first place (paid social, video). Brand search did not stop mattering. It just stopped getting paid for demand it captured rather than created.
If you allocate budget off the last-click column, you will keep cutting the channels in the MMM column, watch your branded search volume quietly fall a few weeks later, and never connect the two. This is the most common self-inflicted wound in DTC media buying.
Why no model is "right"
Three forces make a single perfect model impossible, and all three got worse in the last few years.
Tracking loss came first. iOS privacy changes, cookie deprecation, and consent banners mean a real and growing share of touches are simply never recorded. The model cannot credit what it never saw. A platform reporting "180 conversions" is increasingly filling gaps with modeling of its own, which means you are often comparing one black box to another.
Self-reporting bias came second. Every ad platform marks its own homework. Meta counts a sale if someone saw an ad in the last seven days, Google counts the same sale if there was a click, and both will happily claim it. Add up the platform-reported conversions across your channels and you will often "sell" 130 to 160 percent of what your checkout actually recorded. That overlap is not fraud. It is the predictable result of every referee scoring for its own team.
The third force is the one operators forget: incrementality. Attribution asks "who touched this sale." The question that actually matters is "would this sale have happened anyway." Your branded search campaign gets a gorgeous last-click CAC, but a chunk of those buyers were going to type your name in and buy regardless. The ad harvested a sale it did not cause. No attribution model, however sophisticated, answers the incrementality question. Only a holdout test does: turn the channel off for a region or a window, and measure what actually changed.
What to actually run the business on
Stop hunting for the model that tells the truth. Run a layered system instead, where each layer answers a different question and you never ask one tool to do another's job.
Run the business on blended
MER or blended CAC is the number that cannot lie to you
Steer the week with a model
Data-driven or platform numbers for direction, not gospel
Settle big bets with holdouts
Geo splits and on/off tests answer incrementality
Layer one is your source of truth: total spend against total new-customer revenue. Blended CAC or MER does not care which channel gets credit, because it counts every dollar in and every dollar out. It cannot double-count, because there is only one denominator. If your blended number is healthy and trending the right way, your mix is working even when the platform dashboards disagree about why. If you are unsure which of those two to anchor on, we wrote a whole piece on blended CAC versus MER, and the MER calculator will get you a working number in a few minutes.
Layer two is direction. A data-driven or platform model is fine for answering "is Meta trending up or down week over week," as long as you treat it as a compass, not a scale. Trust the direction, distrust the decimal. Put these directional reads on the dashboard your team checks every morning so the in-week steering and the source-of-truth number live side by side and nobody confuses the two.
Layer three is for the expensive questions. Before you cut a channel, double its budget, or kill brand search to "save money," run a holdout. Turn it off in two states for three weeks and watch the blended number. Geo-based incrementality tests are the closest thing to ground truth you can run without a lab, and they routinely overturn what the attribution reports swore was true. You do not run these weekly. You run them when real money is on the line.
The honest takeaways
Attribution models are budgeting tools, not truth machines. Pick the one whose blind spot you can live with for steering, and never let it set the whole budget alone.
Run the business on blended CAC or MER, because it is the one number that cannot quietly double-count or starve your upstream channels.
Use a directional model for in-week reads, and benchmark what you see against where DTC acquisition costs actually sit this year so you know whether a moving number is a problem or just the market.
And reserve incrementality tests for the decisions worth being right about. The perfect attribution model does not exist and chasing it costs you the months you could have spent measuring what actually changed when you turned a channel off.