The almanac and the storm
How to turn marketing mix modelling from an annual report into a live instrument feeding this week's bids
On the afternoon of 15 October 1987, Michael Fish stood in front of the BBC weather map and announced that a woman had rung in worried about a hurricane, and viewers were not to worry, there wasn’t one coming. That night the worst storm since 1703 took down fifteen million trees. The Met Office’s failure was cadence as much as physics. The models and observations arrived too sparse and too slow to see a storm assembling over the Bay of Biscay. The fix, over the years that followed, involved better data on a faster clock feeding the same science.
Marketing mix modelling (MMM) is the advertising industry’s Met Office, and most brands still run it on the 1987 clock. The discipline works. Sixty years of econometrics can decompose your sales into channel contributions, saturation curves and carryover effects, and the post-cookie signal fallout has powered a full MMM renaissance. Meta open-sourced Robyn and Google’s Meridian, two free, open-source tools for building these models, significantly reducing entry cost into MMM.
Here’s how.
Google’s 2025 measurement report recorded a 212% rise in MMM adoption since 2023, eMarketer finds nearly half of US marketers planning fresh MMM investment this year, and a 2025 MediaPost survey found 74% of marketers reporting that privacy rules have carved costly blind spots into their attribution.
Then the model reports in March on what happened last summer, the annual plan gains direction and re-looked qith quraterly refreshes at best. A model that describes last year is an almanac. The category advantage goes to the brand that turns it into a forecast feeding this week’s bids, and that conversion, from settlement document to live instrument.
In a nutshell
Marketing mix modelling usually lands once a year as a report: an almanac that describes last year.
The category edge goes to whoever turns it into a live instrument, refreshed weekly and wired into this week’s bidding strategies.
Build one turns sales MMM from slow to fast moving data.
Build two prices brand equity into the same bid using the brand’s own survey data.
When sales MMM and brand MMM run together, brands grow their organic base, rely less on using paid channels to capture demand that already exists, optimise MMM through agents and as a result, gain category share over time.
For agencies it sells as an always-on optimiser, not a study with recurring and higher margin fees.
Build one: convert your sales MMM into an agentic optimiser
Start with what you already have: three years of sales, spend and channel performance, plus the category data (competitor spend, pricing, seasonality, distribution) that most MMM briefs skip and most categories can buy. A Bayesian open-source model refreshes weekly rather than annually, and geo holdouts and lift tests bring validation; Meridian can take your experiment results as starting assumptions for each channel, so the annual MMM ‘almanac’ and the real‑world lift tests reinforce one another instead of contradicting each other in the quarterly review.
The larger half of the build translates model output into auction bidding strategies without a human in the loop for each decision: channel elasticities become budget reallocation rules, saturation curves become spend caps that trigger before diminishing returns, and bidding optimises towards true channel contribution. D2C brands such as Gousto and HelloFresh for instance, have talked publicly about rebuilding MMM in-house on exactly this logic: the model stops being a report the marketing director reads and becomes structured data and a signal for real time bidding. Your competitor’s machine chases the commodity of default platform’s number. Yours chases a number the platform has no way to compute, that gives brands a category advantage in the auction.
Build two: turn your brand health scores into a brand MMM optimiser
The second build brings together category, channel, and brand strength. Binet and Field’s IPA research showed the core idea years ago: brand building creates the long-term sales base that doesn’t come from direct last-click activity, and that base grows over time, while activation effects fade within weeks. The challenge is that brand impact has usually been tracked in a separate report, on a separate timeline, by a separate team, which means media buying systems cannot act on it directly.
Here’s how.
Pricing brand equity into the bid
Resonance modelling closes that distance, and it runs on far more than share of search. The model fuses three years of channel investment with the brand’s own survey data: the consumer tracker waves already measuring awareness, consideration and preference that most brands commission and pay for. Where that tracking runs thin, a brand can commission fresh studies in its priority local markets and feed those in too, which sharpens the read where the category is contested.
The model estimates two relationships: which channels move which brand metrics, and how movements in those metrics convert into future base sales. Combine them and each channel acquires a second value alongside its short-term return, the demand it creates rather than harvests. Share of search fills the gaps between survey waves so the signal keeps moving weekly (Binet again, from the 2020 EffWorks research). Feed that second value into the same bidding layer as build one, and the machine starts optimiding in near real time.
Think about what this means.
In a category where competitors use the commoditised default optimisation setting, their systems bid only for this week’s conversions. Your system bids for this week’s conversions plus an estimated value of next year’s brand base - ie the budget and bidding strategy required to lift consideration by 3 points - based on surveys of your actual customers that neither the platform nor your competitors can see. The auction itself does not change. What changes is how you steer it toward brand growth signals.
The flywheel : sales MMM and brand MMM work better together.
Sales and Brand MMM grow together. The resonance layer directs money toward equity-building work the sales MMM alone would under-credit. Equity lifts the base: direct traffic, branded search, repeat purchase, the revenue that arrives unbought. A rising base reduces dependence on paid harvesting, which frees budget, which the model reallocates into whatever it has learned builds the base fastest, and each weekly refresh sharpens the estimates. Brands running the annual almanac experience this loop once a year, if the March meeting goes well. On a weekly clock it turns roughly fifty times in the same period, and small compounding advantages taken fifty times a year are how category positions change hands.
How they work better together.
The resonance layer shifts money toward brand-building activity that a sales-only MMM may undervalue. That activity lifts the base: direct traffic, branded search, repeat purchases, and other revenue that arrives without being directly bought. As the base grows, the business becomes less dependent on using paid channels to capture demand that already exists, which frees up budget for the model to reinvest in the channels and tactics it has learned build the base fastest. Because the model refreshes weekly, the predictions constantly improve - cleaner data, better predicitons, better outcomes.
By contrast, brands running an annual or quarterly planning cycle only see this feedback loop once or a few times a year.
Those small gains, repeated week after week, are how category positions gradually change hands in the auction.
The agency economics: sell MMM as an optimiser
Econometrics has traditionally been priced by analyst hours. Since the 1960s, more clients meant more people and every refresh meant another fee. Instead of a labour-heavy fees, the pipeline becomes a machine-learning product, powered by agentic agents: built once, refreshed by software, validated through scheduled experiments, monitored by a small central team and powered by agentic agents.
That changes the economics.
An agency running it for ten clients does not need ten times as many analysts, and adding the eleventh client brings in new licence revenue with almost no extra cost. Revenue grows while headcount stays flat, so margins stay at software-like levels rather than being dragged down by service delivery. It also makes the fee harder for procurement to compare with an hourly rate card.
Agentic automation is what drives the marginal cost down. Building the system above used to require a scarce econometrician at every stage: cleaning the panel, choosing priors, fitting the Bayesian model, and then manually translating the elasticities into bidding rules. Now agents handle most of that assembly. They ingest the sales and spend data, retrain the MMM on each weekly refresh, and build the optimisation models that turn channel elasticities and saturation curves into the budget rules used for bidding strategies.
The econometrician’s role shifts from building the model to reviewing it: setting the levers and spotting where the agent’s fit looks wrong. One senior person can now oversee an entire client book that once would have needed ten separate teams.
That changes the client conversation too.
The agency starts selling an always-on system rather than periodic dips, wired into the client’s own brand surveys. That makes the relationship much harder to unwind than a standard retainer, because switching agencies now means switching off the machine that directs budget, which creates stickiness through infrastructure.
One honest caveat.
A weekly MMM is not automatically better just because it updates more often. More frequent refreshes also create more ways for stuff to go wrong: data feeds can break, the model can react too strongly to noise, and Bayesian models still rely on priors chosen by someone who ‘gets’ the category.
The real value of weekly updating is that it creates a tighter feedback loop. But that feedback loop only works if it is grounded in reality. That is why lift-test calibration matters: geo-tests, holdouts, and similar experiments give the model a direct check on what actually caused incremental lift, so it can be corrected and validated.
In simple terms, weekly MMM can become confidently wrong unless it is kept in touch with real-world experiments. Lift tests stop the model from becoming a polished but misleading story about the past.
The same logic applies to agentic systems. Humans are essential, but their role changes. Agents can handle repetitive work, while people still need to set goals, choose assumptions, check the outputs, and take responsibility for decisions.
So the fastest systems are not fully human-free. They are systems where agents execute quickly and humans provide judgment where context, uncertainty, and mistakes matter most.
The Met Office learned the 1987 lesson and rebuilt around cadence across more observations, faster assimilation, and the same physics. British weather forecasting is now among the best in the world, while Michael Fish became a lasting symbol of the limits of the old system.
The tools now exist for weather and for media.
Brands still running annual econometrics have chosen the 1987 setup on purpose, even though the storm is already building in their category.
If your MMM still reports once a year, I would love to know what’s stopping the weekly version - data, priors, or politics. Subscribe for the next issue, and forward this to whoever owns the forecast at your company.
This is Erfan Djazmi, fortnightly notes on the commodity gap in marketing, from someone who ran the money.