Skip to content
DocsSearch the docsEnglishBack to siteOpen console

Run a marketing mix model

A marketing mix model estimates how many conversions each paid channel actually caused, not how many it happened to be the last click on. This page covers when the model can run, how each channel's ROI is earned, and how to read the output.

The question this tool answers#

Attribution reports tell you which touch came before a conversion. That is a correlational question, and it over-credits channels that harvest existing demand, such as brand search and retargeting.

A marketing mix model asks a different question: how is the whole budget working? What share of conversions each channel drives, where it saturates and how the budget should be split. The model puts each channel's weekly spend next to the conversions of the same weeks and separates each channel's share from the baseline. Every estimate carries a 94% interval, and the width of that interval is a large part of the output.

Two other measurement tools feed this model: a Lift test measures one planned change, and Incrementality measures every spend change automatically. Their results calibrate the model.

This is a premium feature

Marketing Mix (MMM) is not in the Free plan. If the website is on Free you get a message with an Upgrade plan button instead of the page. Plans are set per website, so one of your websites can have this page while another does not.

Before you run: the data the model needs#

The model lives at /marketing-mix and reads these inputs:

Input Where it comes from Without it
Spend per channel The Cost & ROAS screen (ad-spend import) there are no channels to model
Conversions or revenue The website's purchase events, with the value parameter for a revenue target the run fails
Calendar events and outages The seasonality calendar, automatically a holiday wave is credited to your ads

A few quantitative conditions decide what you get back:

  • History. The model starts from 26 weeks of tracked data and only models the span your tracking code recorded. The TRUSTED tier needs at least 104 weeks (two years).
  • Weeks. Weeks are counted in the website's timezone. A website on Tehran time has weeks ending on Friday; every other website, on Sunday. A trailing partial week is left out.
  • Spend variation. A channel that spends roughly the same every week cannot have its effect separated out. A weekly-spend coefficient of variation below 0.30 marks it Not identified.
  • Collinearity. When several channels' budgets rise and fall together, the data cannot tell them apart. A VIF above 5 earns the same badge.
Spend in several currencies

Spend is converted into the website's currency before modelling (rial and toman at the fixed ×10; other currencies through the exchange rates a platform administrator maintains). If a currency carrying more than 5% of spend has no rate, the run stops and names it.

A channel's ROI arrives with calibration, not with time#

This is the most important point on this page. With less than two years of data, the model cannot tell on its own how much of your sales paid media causes. So:

  • A calibrated channel (by a Lift result, or a spend change Incrementality measured) shows its ROI.
  • Every other channel has its ROI withheld and reads Needs a lift test in the Informed by column. Hover it to see how to unlock it.
  • After two years, uncalibrated channels show an ROI too, but an uncalibrated ROI level leans on the model's prior, and calibration still makes it more accurate.

So the fast way to a channel's ROI is not waiting: change that channel's spend in a clear step and let Incrementality measure it, or run a Lift test. Calibrate the model, further down this page, shows how the result reaches the model.

Start a run#

1
Open Marketing Mix (MMM) under Marketing measurement.
2
Select Run model now to open the Configure a model run drawer.
3
Pick the Target metric: Conversions or Revenue. Conversions is right for most websites; revenue is worth it when order values vary widely.
4
Cadence is Weekly. The model is fitted weekly only.
5
Select Queue run. The result drawer stays open and shows the run's status.
The Marketing Mix run queue and each run's data-sufficiency trust tier.
The Marketing Mix run queue and each run's data-sufficiency trust tier.

Each row in the list is one run: its date, target, week and channel counts, and at the trailing edge either a status (Queued, Running, Failed) or the finished run's trust tier. Runs still in flight always sort to the top.

Fitting happens in the background and usually takes a few minutes; you do not have to keep the page open. Once a day AdPix also queues a run itself for every website that has imported ad spend and has had no run in the last 7 days; that is where a run you did not start comes from. A run that stays Running for more than 6 hours turns Failed with the model job stopped before this run finished — run it again.

How to read the output#

Select any finished run to open the result drawer. Read the cards in the order they appear.

The trust tier comes before the numbers#

At the top of the drawer sits the run's Trust tier with a one-line verdict. This label has the first word.

Tier What it means What to do with it
TRUSTED at least 104 weeks of history and every forecast check passed (below) you can move budget on the intervals
DIRECTIONAL the model fits, but the evidence is limited: a short history, few calibrated channels, or channels whose spend moves together take the direction from it; neither the exact numbers nor a guaranteed ranking of channels
INSUFFICIENT the data does not permit a claim no ROI is shown at all, and the optimizer is closed

DIRECTIONAL makes no promise about the order of channels. When two channels' intervals overlap, the data cannot say which is better, even if one point estimate is higher.

ROI and marginal ROI#

Per-channel ROI and mROI with the 94% credible interval, plus the response curve.
Per-channel ROI and mROI with the 94% credible interval, plus the response curve.

The ROI table has these columns:

Column Definition
Confidence that channel's Identified / Uncertain / Not identified badge
Informed by Data, Partly data, Prior or Needs a lift test; ✓ means the channel is calibrated
Incr. conv. the conversions the model attributes to the channel across the modelled period
Share that channel's share of total conversions
ROI incremental conversions per thousand of spend, with its 94% interval
mROI the return on the next thousand at today's spend level, with its 94% interval
CPA and Spend cost per incremental conversion and the channel's total spend in the period

ROI averages everything spent so far; mROI says what the next thousand buys. Because the response curve saturates, a channel can have a high ROI and a low mROI. Move budget on mROI, not on ROI.

A channel that is Not identified, or whose ROI is withheld, shows a dash. Hover it to see why.

Read the 94% interval correctly#

Under each estimate a small bar shows the 94% credible interval (HDI in statistical writing; every MMM figure uses it):

  • A narrow interval means the data has pinned the number down. "2.0 to 2.4" means you can plan on 2.2.
  • A wide interval means there is not enough signal, not that the effect is zero.
  • You may only conclude a channel has no effect when the interval is both narrow and close to zero.
A wide interval is not a licence to decide

If two channels' intervals overlap almost entirely, comparing their point estimates carries no information. Do not move budget on it; calibrate the channel.

Decomposition and response curve#

The decomposition chart splits every week into layers: the bottom layer is the baseline (organic and direct demand, trend and seasonality) and each layer above it is one paid channel's incremental contribution. The dashed line is that week's actual conversions. A large baseline share means most sales would have happened without advertising; that is the ceiling on what advertising can do.

The Response curve (diminishing returns) chart shows, per channel, what happens to incremental conversions as weekly spend moves. The vertical now line is today's level, the grey zone is beyond the highest spend you have ever run (there the model extrapolates), and the band around the curve is the same 94% interval.

Model health and forecast checks#

  • Model health: R² (fit), MAPE, Holdout MAPE, R-hat (max) and ESS (min) for convergence, and the Baseline share. The chart beneath puts actual conversions against the model's 94% predictive band; the actuals should mostly sit inside it.
  • Data sufficiency: per channel, the spend CV, VIF, spend levels and the reason behind every badge.
  • Model checks: the share of conversions the model credits to paid media, which channels are calibrated, and every experiment or channel it could not use, each with its reason.

Three forecast checks decide whether a run can be TRUSTED:

Check TRUSTED needs
8-week forecast error (worst of 3) the model is refitted three times without its last weeks and forecasts them; every one of the three within 25%
Held-out weeks inside the 80% band at least half of the held-out weeks inside the 80% band
Trend sensitivity with a stiffer trend, no channel's ROI moves by more than its own uncertainty

With a short history these checks may read History too short to test. If a Lift test disagrees with what the model learned from your history, the card says so and the run stays DIRECTIONAL until the two are reconciled.

Calibrate the model#

Calibration experiments (anchor the model to lift tests) sits at the bottom of the page. The next model run uses every active calibration. There are three sources:

Source How it reaches the model
A Lift test on one channel its result appears in the table as Waiting for review, its window and spend change under Measured change
Incrementality every measured episode that changed a single channel arrives the same way
Entered by hand an experiment you ran outside AdPix (geo holdout, budget holdout, PSA/ghost ads)

Turning one on is always your decision: Use in model activates a row. A row badged Recommended is a grade A or B Lift result, calibrated on this website's own history, with its full spend change recorded. Activate all recommended turns all of them on in one click. If a Lift result is later re-measured or removed, its calibration is turned off automatically.

To enter one by hand, give the channel and test type and fill the whole Measured spend change (calibrates the v2 model) section: Test start, Test end, Daily spend before, Daily spend during, Lift measured in (conversions or revenue), Total lift in the window and ± std. dev.. Fill all six, or leave them all empty. An ROI with a standard error on its own only anchors the legacy engine; the current model sets it aside. A row entered this way shows — under ROI (it has none) and its window, daily spend before → during and lift ± std. dev. under Measured change.

Which change teaches the most

A test that cuts a channel's budget and one that raises it pin that channel's curve best. Change spend by at least 20% and hold it for two weeks.

The budget optimizer and planner#

The Budget optimizer card uses the response curves to propose a weekly split: Maximize conversions for a fixed budget, or Hit a target CPA. Each channel moves only within 70%–130% of its current weekly spend (the average of the last 4 weeks) and never above 1.2× the highest weekly spend the model has seen. A minimum or maximum you enter for a channel is intersected with that range. A channel whose ROI is withheld stays locked at its current spend. When the budget is more than those limits allow, the difference is shown as Unallocated budget. On an INSUFFICIENT run the optimizer refuses.

The Budget planner card goes further. Choose where its curves come from (Marketing Mix model, Measured spend changes from the Incrementality page, or Best informed per channel), a Goal and an Attitude to risk, then select Build plan.

Goal What it does
Most conversions for a budget splits a weekly budget for the most conversions
Cut the budget, lose the least reaches a lower budget while giving up as few conversions as possible
Same conversions, less spend finds the smallest budget that keeps today's conversions
Spend while the next conversion costs less than a target keeps adding spend until an extra conversion would cost more than your target
Budget ceiling with a cost target the same, but never above a budget

Every plan is tested on each plausible version of the model. Chance it beats today is the share of versions where the plan does better, and Confidence is high only when that share is at least 80%. A plan's 14-day prediction accounts for carryover: a channel whose effect builds slowly shows less in the first two weeks than at steady state. The plan suggests moving at most 20% every two weeks. With Save and follow up, once the Activity Ledger finds the change you made it measures it, and the Follow-up shows whether the result landed within the predicted range.

Model settings#

Next to Run model now, Measurement v2 engine is on by default, and everything above describes it. Controls defaults to External factors only (total effect): ads also bring organic visits, so organic traffic is an outcome of your ads, not a control. Inflation-adjusted expresses spend in today's prices from the price index; the run stops if the index is more than 45 days old. Each change applies from the next run; finished runs keep the engine they ran with.

Legacy engine

Turning Measurement v2 engine off fits the next runs with the legacy engine. It does not run at all below 78 weeks, uses on-site organic and direct traffic as a control (so it understates a channel's total effect), and reads calibrations as an ROI with a standard error. Its TRUSTED tier needs 104 weeks, at least half the channels identified, and a baseline share between 45% and 92%. It is kept only to compare with older runs.

Exports#

On any finished run that is not INSUFFICIENT, the Export menu offers PDF report, CSV data (one row per channel with ROI and its intervals) and PNG chart. If your role may not see cost, the cost columns are absent from the CSV entirely.

Frequently asked questions#

How many weeks of data does the model need?

26 weeks of tracked data. With less, the run closes on the INSUFFICIENT tier and shows no ROI. If you have enough history and the run still fails, either no channel spend has been imported or the website has no purchase events to explain.

Why does a channel show no ROI and read "Needs a lift test"?

With less than two years of data, the model cannot tell on its own how much of your sales paid media causes. A channel's ROI arrives with calibration, not with time. Change that channel's spend and let Incrementality measure it, or run a Lift test, then turn the result on in the calibration card. There is no two-year wait.

Does DIRECTIONAL mean the channel ranking is right?

No. It means the evidence is limited and the numbers only show which way things point. When two channels' intervals overlap, the data cannot say which is better, and their ranking is not guaranteed either.

Does a wide interval mean the channel's effect is zero?

No. A wide interval means the data cannot pin the number down yet. You can only say a channel has no effect when the interval is both narrow and close to zero.

How often should I run the model?

You don't have to run it by hand. Once a day AdPix queues a run for every website that has imported ad spend and has had no run in the last 7 days. Run it manually after importing a backlog of spend or turning a calibration on.

Build with the APIUnderstand where revenue comes from.
Was this page helpful?