MMM vs last-click attribution: 104 weeks against one click

The usual pitch for marketing mix modelling is that it replaces last-click attribution now that cookies and consent have broken it: the same report, without tracking anyone. It is not the same report. Last click credits one order, the day after it happens. An MMM credits a channel, from two years of weekly totals at minimum, and with the publishers’ own rules of thumb, two years of data supports about ten variables.

So the choice is not old method against new method. They answer different questions on different clocks, and an account that swaps one for the other loses the question the first one answered.

The three MMM frameworks, by their own documentation

We read the data-requirement pages of three open-source MMMs, one each from Google, Meta and PyMC Labs: Meridian, Robyn and PyMC-Marketing. Publisher pages only.

	Meridian (Google)	Robyn (Meta)	PyMC-Marketing (PyMC Labs)
History	No fixed minimum. The guidance is “rough and directional”	Two years of weekly data, minimum. 4–5 years if monthly	“Ideally 2–3 years”
Granularity	Weekly, ideally by region	Weekly is best practice. Daily with extra checks	Weekly or daily
Data per variable	Its example: 4 data points per parameter is “too low”. About 15 gives “directional information”	10 observations per independent variable	Not stated on the introduction page
Inputs	Spend or impressions per channel, controls, a summable KPI	Aggregated data, no user-level data	Spend or impressions per channel, KPI series
Calibration	Incrementality experiments, “perhaps the strongest basis” for priors	Conversion Lift or GeoLift, “strongly” recommended	Not stated on the introduction page

None of the three asks for a click. Meridian accepts clicks as one possible exposure metric, summed per channel per week. The unit of data is a row per week (or day), per channel, per region if you have them.

The arithmetic

Meridian’s own documentation works one example through. An account with 12 media channels, six control variables and eight knots (the time-trend terms) is 26 parameters. Two years of weekly data is 104 rows. That is four rows per parameter, which Meridian calls “too low to estimate the model reliably”.

Apply Robyn’s ratio to the same account. Twelve channels and six controls are 18 independent variables, which need 180 weeks: about three and a half years of weekly data. The 104 weeks that Robyn sets as its floor support 10 variables.

Meridian’s fix is to shrink the question: combine the 12 channels into three, cut the knots to two, drop a control that is not a true confounder, and use three years of data. That leaves 10 parameters on 156 rows, about 15 each, where “you might be able to glean some directional information”.

Read what that fix gives up. Twelve channels became three. A four-week campaign is four rows out of 156, folded into one of those three channels. An MMM grades a channel over years. It never grades a campaign. Last click does the opposite: it names the campaign, the ad and the keyword behind each order, and it knows nothing about whether the order would have happened anyway.

Two ways out exist, and both have a cost. Meridian’s geo model multiplies the rows: 105 regions over three years is 16,380 data points. But it also multiplies the parameters. Without sharing information across regions, the same 12 channels, six controls and 100 knots become 2,094 parameters. Robyn’s way out is daily data, which it recommends when the window is as short as six months, with extra validation.

What each one counts, and what each one misses

Last click counts tracked clicks. In GA4, paid and organic last click gives 100% of the value to the last channel the customer clicked through from, and ignores direct visits unless the whole path was direct. It misses every exposure that did not end in a tracked click: an ad seen and not clicked, a TV spot, a podcast read, and an AI assistant’s answer that the reader acted on without clicking.

An MMM counts what moves together over time. It sees TV and offline sales, because it only needs weekly totals. It misses anything whose spend did not change. Robyn’s guide gives the case: if TV spend stayed flat for the whole period, the model cannot tell what TV did.

The finding worth remembering: both over-credit paid search, for different reasons. Last click does it because a search click so often comes last, right before the purchase. An MMM does it because search spend follows demand: when more people search for the product, more search ads serve, so spend and sales rise together whether or not the ads caused anything. Meridian names this directly: without search query volume as a control, the model “can lead to overestimation of the causal effect of paid search”. It also says to split brand and generic search into separate channels. A model that skips both inherits last click’s search bias through a different door.

What AI changed, and what it did not

Building an MMM got cheaper. Meridian now ships five Agent Skills for AI coding assistants that map CSV columns to KPIs, spend and controls, fit the model and draft the budget report, and stop for the analyst to confirm the priors and knots.

The data requirement did not move. An assistant can write the model code. It cannot produce the two years of weekly history the model needs, or the variation in spend that lets the model tell channels apart. AI assistants also widened last click’s blind spot: an answer that shapes a purchase without a click is invisible to it. Our test of GA4’s AI Assistant channel found it only catches the clicks that arrive untagged. An MMM can include that kind of exposure only as a weekly number you can sum, and few teams have one.

What to do

1. Keep last click for the daily decisions: which ad, which keyword, which landing page. Read its paid-search share as an upper bound, not a measurement.
2. Start keeping the MMM’s data now, even if the model comes later. Weekly spend and impressions per channel, weekly sales, by region if you sell in several. The model needs two years of it, and a week you did not record is a week you have to rebuild from exports later.
3. Count variables before you fit anything. Weeks of data divided by 10 is Robyn’s ceiling on the number of channels plus controls. If your channel list is longer, merge channels before the model does it badly for you.
4. Add search query volume as a control and split brand and generic search, or the model repeats last click’s search bias.

Both Meridian and Robyn tell you to anchor the model with incrementality experiments, a lift study or a geo test, so the model’s answer for a channel starts from a measured one. That makes the experiment the foundation. How much is a lift result worth when the account runs about a thousand conversions? Less than its certainty score suggests: ads that do nothing get a lift result half the time.