Your report says branded search drove most of last month’s sales. The team celebrates and moves more budget into that campaign.
Nobody asks who taught that buyer the brand name in the first place.
That is the blind spot. Last-click attribution does not measure influence. It measures proximity. It hands the credit to whichever channel stood closest to the checkout.
What the last click actually measures
The model is simple, which is why it survived so long. Someone clicks, buys, and the full value lands on that final recorded click.
It worked reasonably well when journeys were short and cookies saw almost everything. Today the journey runs through video, messaging apps, search, reviews and private recommendations that no cookie observes.
Blockers, third-party cookie limits and in-app traffic broke that single view. We wrote about that shift in our piece on first-party data strategy.
Last-click attribution does not fail out of bad faith. It fails because it only ever sees the end of the story.
One more problem makes it worse. Each platform reports its own conversions on its own time window. Add three platform reports together and the total beats real sales. That gap is not a tracking error. It is structural double counting, and last-click attribution hides it behind one clean number.
Three signs your report is rewarding the wrong channel
Branded search looks unbeatable. Its cost per acquisition is the lowest line in the plan, month after month. It usually is, because it harvests demand another channel planted.
Retargeting looks miraculous. It chases people who already decided. Switch it off for a week and the business tends to move far less than the report predicted.
Discovery channels never perform. Video, audio, media and public relations sit in the red all year. They rarely earn the final click, so they rarely earn credit.
Email claims sales it did not create. An abandoned-cart reminder closes a purchase that was already decided. The channel reads as the most profitable in the plan, and the team mistakes it for new demand.
If your media plan has drifted for two years toward whatever already converts, you are probably optimising the harvest while neglecting the planting.
Media mix modeling came back, and it is open source now
The alternative is not new. It models the whole business instead of chasing users one by one.
A media mix model takes time series — spend by channel, sales, price, seasonality — and estimates how much each lever contributed. It needs no cookies and no personal identifiers.
That is also why privacy teams tolerate it. The model never touches an individual profile. It works on aggregates, which is the same reason it keeps working as tracking gets harder every year.
What changed is the cost of entry. Google released Meridian, an open-source MMM that covers media mix modeling, reach and frequency for video planning, and paid search modeling using search query volume.
Meta maintains Robyn, its own open-source MMM toolkit in R and Python. Robyn estimates adstock rates and saturation curves by combining ridge regression with evolutionary algorithms.
Adstock and saturation are exactly the two ideas missing from every click report. The first accepts that today’s advertising also sells tomorrow. The second accepts that the tenth dollar returns less than the first.
Without experiments, a model is a well-dressed opinion
An MMM can produce an elegant, wrong answer. Give it few data points and many correlated variables, and it will find patterns that are not there.
That is why the current standard pairs models with experiments. You raise or cut a channel across a set of regions, compare against control regions, and measure the real difference in sales.
Google built that logic into Meridian GeoX, which converts experiment results into priors that calibrate the model. Its own documentation frames the goal as calibrating MMMs with experiments, across channels.
The short version for your board: the model proposes and the experiment confirms. Skip the second half and you have swapped last-click attribution for a fancier spreadsheet.
How to start without a data science team
You do not need statisticians to move past last-click attribution. You need discipline with your own records, and one honest test.
- Gather two or three years of weekly data: spend by channel, sales or leads, price and promotions.
- Flag the events that moved the business: launches, stockouts, price changes, long holidays.
- Run an experiment before you build a model. Switch one channel off in a few markets for four weeks.
- Compare the outcome against what your click report predicted for those weeks.
- Treat that gap as your baseline, and only then evaluate a formal MMM.
A mid-sized brand does not need a full Bayesian model in its first quarter. It needs to stop making decisions from a single number.
That first experiment usually costs less than one week of media. And it answers the question no dashboard answers: what happens to sales when you stop paying.
What changes in the conversation with leadership
Better measurement changes the questions. You stop arguing about which channel posted the best cost per acquisition.
You start asking how much of that sale would have happened anyway. Where each channel saturates. Which one is underfunded and has nobody defending it, because it never shows up in the report.
The calendar changes too. A media mix model gets reviewed quarterly, not daily. That forces decisions with a horizon and kills the urge to shift budget every Monday over statistical noise.
Retail media is a clean example of the tension. It converts close to the purchase, so it always looks excellent. We unpacked it in our article on retail media advertising.
No method replaces judgement. But a media plan held up only by last-click attribution ends up funding the final metre of the race and forgetting the first thousand.




