Data Analytics

Marketing mix modelling for small brands: is it worth it yet

Attribution is getting less reliable and marketing mix modelling is being sold as the replacement. For a brand spending a few lakh a month, it usually is not.

Key takeaways
  • MMM needs years of weekly history and real variation in spend. Most small brands have neither.
  • Without variation in your spending pattern, the model cannot separate channels no matter how good the maths.
  • Geo holdout tests answer the same question at a fraction of the cost and with fewer assumptions.
  • Revisit MMM when spend is large enough that a ten percent misallocation is worth more than the model costs.

Marketing mix modelling has become a common recommendation for Indian D2C brands, and the reasoning behind the trend is sound. Platform attribution has become less reliable as privacy changes have reduced tracking, cross device journeys have grown more complicated, and the gap between what ad platforms claim and what shows up in the bank has widened.

MMM is offered as the answer, and for large advertisers it often is. For a brand spending a few lakh rupees a month across three or four channels, buying one is usually premature and occasionally harmful.

What the technique needs to work

MMM is a statistical model that attempts to explain sales as a function of media spend, price, seasonality, distribution and external factors. It infers each channel’s contribution from how sales moved historically when spending patterns changed.

That inference imposes two hard requirements.

The first is history. Weekly data over two to three years, so the model has enough observations to estimate a fair number of parameters. Anything much shorter and you are asking it to separate many effects from very few data points.

The second, and the one that gets ignored, is variation. The model learns from change. If you have spent roughly the same amount, in the same proportions, on the same channels, week after week, then those channels moved together and no technique can tell them apart. Stable spending feels like good discipline and it makes your data uninformative.

Why a bad model is worse than no model

If insufficient data simply produced an error message, this would be a smaller problem. Instead it produces output, formatted attractively, with contribution percentages per channel.

Those numbers carry authority. They get presented to a board, they inform a budget reallocation, and nobody in the room can easily challenge them because the method is unfamiliar. A brand can shift a meaningful share of its budget on the strength of a model that was fitting noise.

The tell is usually confidence intervals. If they are absent from the output, or so wide that a channel’s contribution could plausibly be anywhere across a huge range, the model is expressing uncertainty that the summary slide is hiding. Ask for them explicitly, and treat reluctance to provide them as an answer in itself.

The cheaper test that answers the real question

Most brands do not actually need a full attribution picture. They need an answer to a specific question: is this channel worth what we are spending on it.

That question is answerable by experiment. Switch the channel off, or reduce it substantially, in a set of regions while keeping it running in comparable regions, and compare what happens to sales. This measures incremental effect directly rather than inferring it, and the logic is simple enough that a non technical audience can interrogate it.

India’s regional structure suits this well. Cities and states differ enough to give you meaningful test and control groups, and most platforms allow geographic targeting fine enough to run it. The cost is the revenue you deliberately forgo in the test regions for a few weeks, which is real but usually far below the cost of a modelling engagement.

Run one channel at a time, on the largest spend first, and you will accumulate a genuine incrementality picture over a few quarters.

Simple signals that carry real information

Alongside experiments, a few cheap indicators do more work than they get credit for.

Branded search volume is the most useful. It is the clearest available proxy for whether upper funnel activity is creating demand, it is free to track, and it responds to real awareness rather than to tracking pixels.

A post purchase survey asking how customers first heard of you is imperfect, biased towards recent and memorable touchpoints, and still frequently more informative than a platform’s self reported attribution. Treat it as directional and watch how the mix shifts over time.

And the blunt one: total marketing spend against total new customers acquired, tracked monthly. It tells you nothing about allocation between channels, but it tells you immediately whether the overall machine is getting more or less efficient, which is the question that actually determines whether the business works.

When it becomes worth doing properly

The threshold is economic rather than technical. MMM is worth commissioning when a ten percent misallocation of your budget costs meaningfully more than the model and the analyst time.

For a brand spending a few lakh a month, ten percent is a small number and the model costs more than the error. At substantially larger spend across many channels, with several years of history and genuine variation in it, the arithmetic reverses and the investment makes sense.

Until then, the most valuable thing you can do is build towards being ready: keep clean weekly records of spend by channel, sales, price and promotions from now on. The single biggest obstacle when brands eventually want MMM is not budget, it is that nobody kept the history in a usable form, and that is a problem you can solve today at no cost.

The daily brief

Never miss a move

The moves that move money, every morning.

One email a day. No spam, ever.

FAQ

Quick answers.

As a working rule, two to three years of weekly data, which is a hundred or more observations, with meaningful variation in spend across the channels you want to separate. The variation matters as much as the length. If you have spent a steady proportion on the same two channels every week for two years, the model has nothing to learn from, because the channels moved together and cannot be distinguished.
You can fit a model, and it will produce confident looking numbers. Whether those numbers mean anything is a different question. With roughly fifty weekly observations and several channels plus seasonality and price effects to estimate, you are close to fitting noise. The danger is not that the model fails visibly, it is that it succeeds visibly and sends you in a wrong direction with apparent authority.
You switch a channel off, or substantially down, in a set of matched regions while keeping it running in comparable control regions, then compare sales. It measures incremental effect directly rather than inferring it statistically, which makes it far easier to trust. It is cheaper, faster and more honest than a model for most single channel questions, and the main cost is the revenue deliberately forgone in the test regions.
Open source implementations remove the software cost, which was never the main barrier. The binding constraints are data history, spend variation and the analytical judgement to specify and validate the model. A free tool applied to insufficient data produces a free wrong answer, which is more dangerous than no answer because it carries the credibility of having been modelled.
Run structured incrementality tests on the channels where most of the money sits, one at a time. Track branded search volume as a proxy for upper funnel effect. Ask customers at checkout how they heard about you and accept it as directional. These are unglamorous and they answer the practical question, which is usually whether a specific channel is worth its budget, rather than the theoretical question of how all channels interact.

Related insights

From the wire

India's Commerce Engine

Put it
to work.

hello@zane.marketing

Book a meeting