You've uploaded two years of revenue and channel spend. A few seconds later, the model has a chart, a channel ranking and a suggestion about where your money should go. After months of conflicting platform reports, that clarity feels like progress. It might be. Before moving the budget, though, ask how much independent information the data contains and whether the recommendation survives a reasonable challenge. A precise answer is only useful if the data can support it, which means someone needs to challenge what the model actually learned before its recommendation reaches a budget meeting.
Twenty-four monthly observations give you 24 rows of history, which can help you explore patterns and expose useful questions. Once you're trying to separate several channels, seasonality, underlying demand, promotions and delayed advertising effects, however, those same rows can contain very little independent evidence. The responsible starting point is to ask what this particular dataset can support. The existence of a coefficient or a budget recommendation doesn't answer that question.
Two years of history can contain very little variation
Suppose paid search and paid social rise together every November and fall together every January. If revenue follows a similar pattern, we'd need to ask whether the extra spend created the increase, seasonal demand encouraged the business to spend more, or both happened. Although a regression can detect a relationship, that relationship alone won't resolve the ambiguity. Adding a seasonal control may help represent the business, but it doesn't manufacture independent channel variation where none exists.
Here's an intentionally simplified mocked up. In every month, social spend is exactly half of search spend. Revenue above a fixed baseline follows the equation 4 × search spend. You could represent that same pattern as 4 × search plus 0 × social, or 2 × search plus 4 × social. Because social is always half of search, both equations produce identical predictions, leaving the data unable to choose between those explanations. A model might still return one split because of a constraint, penalty or prior assumption, in which case that dependence needs to be visible. Presenting the split as something the observed data settled would hide the very uncertainty the budget owner needs to understand.
| Search spend | Social spend | Revenue above baseline | Split A: 4 × search | Split B: 2 × search + 4 × social |
|---|---|---|---|---|
| £10,000 | £5,000 | £40,000 | £40,000 | £40,000 |
| £20,000 | £10,000 | £80,000 | £80,000 | £80,000 |
| £30,000 | £15,000 | £120,000 | £120,000 | £120,000 |
These coefficients are mathematical illustrations. Real data is messier, and channels don't need to move in perfect lockstep for separating their effects to become difficult. The example shows why a strong fit can coexist with an unreliable channel recommendation. Google's Meridian guidance treats collinearity as something to inspect during exploratory analysis.\[1\]
The model has to represent the business you ran
Revenue moves for reasons beyond advertising, including price, product availability, discounts, distribution, holidays and changes in customer demand. Unless the model accounts for the relevant influences, it may attribute some of those movements to media. That doesn't mean adding every available variable will help, because doing so can exhaust a small dataset and introduce its own problems. The question is which factors are necessary to represent the decision credibly, with a clear explanation of the assumptions.
Time adds another constraint because marketing can have delayed effects and its response may change at different spend levels. To estimate those patterns, we need to make additional choices and often introduce more parameters. There isn't a universal row count that makes every specification reliable. As a model-specific rule of thumb, Meridian's collection guidance recommends at least two years of weekly history for geo-level models and three years for national-level models.\[2\] That's a useful warning against treating two years of monthly data as equivalent evidence, rather than a blanket eligibility rule for every possible MMM.
A national model uses country-wide totals, such as one UK revenue figure and one spend figure per channel each week. A geo-level model breaks those figures down by region, so it can learn from differences between places as well as changes over time. That extra variation can help separate marketing effects and reduce the amount of history needed, provided the regional data is reliable and the differences are informative. Simply splitting the same information into more rows won't solve the problem.
Challenge the recommendation before the business pays for it
I'd start by testing whether the decision stays broadly consistent under plausible alternative assumptions. For example, we could represent a major promotion differently, exclude an unusual peak month, simplify the channel grouping or change a defensible carryover assumption. If those reasonable choices make the proposed budget move swing dramatically, the uncertainty belongs in the recommendation. A stable answer is reassuring, although stability alone still doesn't prove that the model has identified causal effects.
Predictive checks can expose weaknesses, too. Where the data permits, compare performance on periods held out from fitting with a simple benchmark. With only 24 observations, those checks are themselves fragile and use up scarce history. Alongside those checks, examine residual patterns, coefficient sensitivity and whether the recommendation requires spending outside the observed range. If it proposes doubling a channel whose spend has barely varied, we'd need to understand what supports that jump, since a smooth response curve doesn't mean the business has observed the return at twice the investment. Any reliance on extrapolation needs to be clear to the person making the decision.
A useful tool should sometimes tell you to stop
By organising data and flagging channels that move together, a lightweight model can help identify where a better test would be valuable. Where the assumptions and evidence are proportionate, it may also support a directional hypothesis, although that doesn't justify turning every upload into a causal channel ranking or an optimised budget. “This dataset can't reliably separate these effects” is a valuable finding if it prevents a costly reallocation.
We don't get an exemption because it's our tool. If the data can't support a channel recommendation, we should say so, because giving you an answer anyway would make the tool easier to sell and harder to trust. Depending on the gap, the next step might be to combine inseparable channels for analysis, obtain better history or regional variation, investigate an external demand driver, or design an experiment. Whether any of those is worth doing depends on the investment at stake. You shouldn't have to commission a major measurement programme to make a small, reversible choice, and a material budget deserves more scrutiny than a persuasive demo.
The attraction of MMM is the possibility of getting beyond competing platform claims. Keep that ambition, and ask the model to earn the confidence you place in it. The most useful output may be a better question, a bounded test or a clear explanation of why the budget should stay where it is until stronger evidence arrives.
If you're considering an allocation decision based on a lightweight model, GV Marketing can help assess what the evidence supports and what you'd need to know before acting.
Read next: work through conflicting budget evidence or check what your ROAS target rewards.