Attribution Is Broken and Everyone Knows It: Marketing Mix Modelling for Mid-Market Teams
Multi-touch attribution is broken — cookie loss and walled gardens saw to that. Here's how mid-market teams can run marketing mix modelling without an enterprise budget.
By Robin Deane — Founder, RD
Multi-touch attribution is unreliable now because cookie deprecation, dark social, cross-device journeys, and walled-garden self-reporting have all degraded the data it depends on. Marketing mix modelling (MMM) — an aggregate statistical approach that never needed individual-level tracking — is the alternative, and it has become newly affordable for mid-market teams through cheaper compute, open-source tooling, and AI-assisted model-building. Most teams should run a lightweight MMM alongside attribution, not instead of it, for at least two quarters before deciding which to trust more.
Every marketing team still running last-click or multi-touch attribution knows, at some level, that the numbers are wrong. Channels report conversions that do not add up to total revenue. Two platforms both claim credit for the same customer. A channel that clearly should not be working keeps showing strong ROAS. Most teams keep using the dashboard anyway, because the alternative has always sounded like something only a Fortune 500 budget could run. That is no longer true, and this piece explains why the old system broke and what to do instead.
Why Does Multi-Touch Attribution Fail Now?
Definition: Multi-touch attribution assigns fractional credit for a conversion to each marketing touchpoint a user encountered before converting, based on tracked individual-level journeys — cookies, click IDs, and device identifiers stitched into a single path.
That definition depends entirely on being able to see the whole path for the whole person. Four things now make that reliably impossible:
Cookie deprecation and identity loss. Browser-level blocking, shortening cookie lifespans, and privacy-focused defaults mean a meaningful share of user journeys break into fragments — the same person looks like two or three different "users" to your attribution model, each with a partial, misleading path.
Cross-device journeys. Someone sees a paid social ad on their phone at lunch, researches on a work laptop in the afternoon, and converts on a tablet at home. Unless all three devices are logged into the same authenticated account throughout, attribution sees three disconnected sessions and typically credits only the last one.
Dark social and untrackable referral. A link shared in a private message, a screenshot posted in a Slack channel, a recommendation in a WhatsApp group — none of it carries a trackable parameter. The conversion arrives as "direct traffic," and direct traffic silently absorbs credit that word-of-mouth and social sharing actually earned.
Walled-garden self-reporting. Search and social platforms grade their own homework. Each platform's ad manager reports conversions using its own attribution window and its own view of the user journey — typically the most generous plausible interpretation. Add every platform's self-reported conversions together and the total routinely exceeds actual total revenue, sometimes by a wide margin.
None of these problems are temporary glitches waiting for a fix. They are structural, and they are getting worse as privacy regulation and platform-level tracking restrictions tighten further.
What Is Marketing Mix Modelling and How Does It Differ From Attribution?
Definition: Marketing mix modelling (MMM) is a statistical technique that estimates the contribution of each marketing channel to business outcomes by analysing aggregate spend and results over time — weekly or monthly totals, not individual user paths. It asks "when we spent more on this channel, did outcomes move?" rather than "which specific touchpoints did this specific customer see?"
The practical difference is what data each method needs. Attribution needs to track individuals, which is exactly the capability that has been degraded. MMM needs channel-level spend and aggregate outcome data — sales, sign-ups, revenue — which nothing about cookie deprecation or dark social affects at all. It cannot tell you that customer #48213 clicked a paid social ad then converted from an email three days later. It can tell you, with statistical confidence, that a 20% increase in paid social spend in a given region was associated with a measurable lift in sales two weeks later, net of seasonality, pricing changes, and competitor activity.
| Dimension | Multi-Touch Attribution | Marketing Mix Modelling |
|---|---|---|
| Data required | Individual-level tracking across devices and platforms | Aggregate spend and outcome data by channel and time period |
| Affected by cookie/privacy changes | Directly and increasingly | Not at all |
| Captures offline and brand channels | Poorly or not at all | Yes — TV, out-of-home, sponsorship, PR included natively |
| Granularity | Individual touchpoint and campaign level | Channel and time-period level, not individual campaigns |
| Speed of insight | Near real-time | Typically refreshed monthly or quarterly |
| Historical data needed | Minimal — works from current tracked sessions | 12–24 months of consistent spend and outcome data ideally |
| Best used for | Tactical, in-flight campaign optimisation | Budget allocation across channels and strategic planning |
Neither method is strictly better — they answer different questions at different speeds. The mistake most mid-market teams make is relying entirely on attribution because it is the only tool they have, not because it is the right tool for budget-allocation decisions.
Why Has MMM Become Newly Accessible to Mid-Market Teams?
MMM has existed for decades, run by econometricians for consumer packaged goods brands with tens of millions in media spend. Three things have changed:
Compute is cheap. The statistical models underneath MMM — Bayesian regression, typically — used to require specialist infrastructure and days of processing time. The same models now run on a laptop in minutes.
Open-source tooling exists. Frameworks originally built inside large tech companies for their own internal media measurement have been released as open-source packages. A team with basic data science capability, or a consultancy that has built the pipeline once, can stand up a working model without building the statistical engine from scratch.
AI-assisted model-building lowers the expertise bar. Configuring an MMM correctly — choosing adstock decay rates, handling seasonality, avoiding common specification errors — used to require a dedicated econometrician. AI-assisted tooling can now guide non-specialists through reasonable default configurations and flag likely errors, closing most of the gap between "enterprise data science team" and "marketing analyst with clean spend data."
The result: a mid-market team spending a few hundred thousand a year across five or six channels can now build a genuinely useful lightweight MMM, not a toy version of one.
How Should a Mid-Market Team Start with MMM?
Spend by channel, and the outcome you care about (revenue, sign-ups, orders) for the same weeks. This is almost always messier than expected — finance's version of spend rarely matches the marketing platform's version. Reconciling this data is the least glamorous and most important step.
Seasonality, pricing changes, promotions, competitor activity if you can approximate it, and anything else that moves revenue independently of your spend. A model that ignores a major sale event will wrongly credit whichever channel happened to be running at the time.
If the model says your highest-performing channel is one you have always suspected is weak, do not immediately reallocate the whole budget. Investigate the specification first. MMM output should update your beliefs, not replace judgement entirely — especially on the first model.
A single MMM run is a snapshot; a quarterly cadence is a decision-making tool. Compare successive models to see whether a channel's estimated contribution is stable or noisy, and weight your confidence accordingly.
Run this alongside your existing attribution setup rather than switching immediately — see our marketing automation ROI benchmarks for what good channel performance looks like while you build confidence in the new model.
How Do You Reconcile MMM and Attribution When They Disagree?
They will disagree, and that is useful information rather than a failure. Attribution tends to over-credit lower-funnel, high-frequency channels (retargeting, branded search) because those are exactly the touchpoints closest to conversion and easiest to track. MMM tends to reveal that upper-funnel and offline channels are doing more work than any tracked path shows, because it captures the effect on total demand rather than only the last visible click.
The practical reconciliation: use MMM for the big allocation questions — how much to spend on paid social versus sponsorship versus content, at a quarterly or annual level — and use attribution for tactical, in-flight decisions — which specific ad creative or landing page is converting better this week. Treat a large, persistent gap between the two as a signal to investigate your tracking setup or your model specification, not as proof that one method is simply right and the other wrong. Our analytics and growth team builds this kind of dual measurement stack for clients who have outgrown attribution-only reporting but are not ready for an enterprise data science hire.
- Multi-touch attribution is structurally unreliable now due to cookie deprecation, cross-device journeys, dark social, and walled-garden self-reporting
- Marketing mix modelling estimates channel contribution from aggregate spend and outcome data, unaffected by individual-level tracking loss
- MMM has become affordable for mid-market teams through cheaper compute, open-source tooling, and AI-assisted model configuration
- MMM requires 12–24 months of clean weekly spend and outcome data, plus variables like seasonality and promotions
- Attribution and MMM answer different questions — tactical, in-flight optimisation versus strategic budget allocation — and work best used together
- A persistent disagreement between the two methods is a signal to investigate, not proof that one is simply correct
- Sense-check the first model's output against existing knowledge before reallocating budget on the strength of it alone
Frequently Asked Questions
Is marketing mix modelling only for large companies?
No — that was true when MMM required specialist econometricians and expensive infrastructure, but cheaper compute, open-source modelling frameworks, and AI-assisted configuration have brought a lightweight version within reach of mid-market teams spending a few hundred thousand a year across several channels. It will not have the granularity of an enterprise model, but it can meaningfully outperform attribution alone for budget-allocation decisions.
How much historical data do you need for MMM?
Twelve months is a workable minimum; 24 months produces materially more reliable results because it captures at least two cycles of seasonality and gives the model more variation in spend levels to learn from. Fewer than 12 months of data generally produces a model too unstable to trust for budget decisions.
Does marketing mix modelling replace attribution entirely?
Not for most teams. Attribution remains useful for tactical, in-flight decisions — which ad creative, landing page, or audience segment is converting better right now. MMM is better suited to strategic, channel-level budget allocation. Running both and comparing where they agree and disagree gives a more complete picture than either alone.
What data do you need to start building an MMM?
Weekly or monthly spend by channel, the business outcome you want to explain (revenue, orders, sign-ups) for the same periods, and any variables that also move that outcome independently of marketing — seasonality, pricing changes, promotions, and competitor activity where you can approximate it. Reconciling marketing platform spend data with finance's numbers is usually the most time-consuming part of this step.
Why do attribution platforms show more conversions than actual total revenue?
Each platform — paid search, paid social, email — reports conversions using its own attribution window and its own view of the customer journey, typically the most generous plausible interpretation of its own contribution. Add every platform's self-reported number together and the total routinely exceeds real total revenue, because multiple platforms are claiming credit for the same conversion. This is one of the clearest signals that attribution-only measurement cannot be trusted for budget decisions on its own.
How often should a marketing mix model be updated?
Quarterly is a reasonable cadence for most mid-market teams — frequent enough to inform the next budget cycle, infrequent enough that each run has meaningful new data since the last one. Compare successive models over time; a channel whose estimated contribution swings wildly between runs is a channel to treat with more caution than one whose contribution is stable.
Keep Reading
