Measuring what marketing actually adds: attribution, marketing mix modelling and experiments
Updated: Nov 3
First published 2nd October 2024. Updated 25th October 2026 with current measurement tools and research.
Every advertising platform reports its own success. Add up the conversions claimed by search, social and display and the total often exceeds what the business actually sold. At the same time, activity that builds future demand, such as video, offline media or content, often looks as if it does nothing. Both problems come from the same source: measuring marketing by what can be tracked rather than by what it causes.
Why attribution alone misleads
Attribution assigns credit for a sale to the touchpoints a tracked customer passed through. It is useful for day-to-day optimisation, but it has well-known limits:
It sees only what it can track. Consent choices, browser restrictions and cross-device journeys mean a growing share of activity is modelled or missing
It confuses correlation with cause. A branded search ad clicked by someone who was going to buy anyway receives full credit for a sale it did not create
Each platform grades its own work. Google, Meta and others measure within their own walls, so overlaps are counted more than once
It undervalues long-term effects. Activity that raises awareness or lowers price sensitivity over months rarely appears in a short attribution window
Google retired most rules-based attribution models in Google Ads and GA4 in 2023, leaving data-driven and last-click attribution. Data-driven attribution is better than last click, but it still describes paths. It does not prove what would have happened without the advertising.
The three methods that work together
Marketing mix modelling (MMM) uses statistical models on historical sales and spend, alongside factors such as price, seasonality and distribution, to estimate how much each channel contributes. It covers online and offline activity, needs no user-level tracking and is well suited to budget allocation. Open-source models such as Google's Meridian (generally available since January 2025) and Meta's Robyn have made it far more accessible. Its limits: it needs enough history and variation in spend and it gives estimates with ranges, not precise weekly answers.
Incrementality experiments compare a group exposed to advertising with a comparable group that is not. Geographic holdouts, conversion lift studies and on-off tests answer the most important question directly: what did this activity add? Experiments are the closest thing marketing has to proof, but they need careful design and enough volume to detect a real effect.
Attribution remains the right tool for in-flight decisions, such as which keyword, audience or creative is working this week, provided its numbers are calibrated against the other two methods.
Used together, they check each other. Experiments calibrate the model. The model sets budgets across channels. Attribution guides daily optimisation within them.
Getting the data foundations right
None of this works without reliable data. The essentials in 2026:
Consent-aware tagging. A consent management platform connected to Google Consent Mode and other platform tags, configured for each market's rules
Server-side measurement. Server-side tagging and conversion APIs, such as Meta's Conversions API and Google's enhanced conversions, recover signal lost in the browser and give the business control over what is shared
Offline and value data. Importing qualified leads, sales and margin back into ad platforms so bidding optimises to real outcomes
One source of truth. A clear definition of revenue, customers and conversions agreed with finance, so marketing and the business report the same numbers
Common misconceptions
"High ROAS means the channel is working." Not necessarily. Branded search and retargeting often show the highest ROAS and the lowest incrementality
"MMM is only for big advertisers." Cost and data requirements have fallen sharply. Many mid-sized advertisers with two or more years of consistent data can now benefit
"One test settles it." Results change with season, creative and competition. Measurement is a programme, not a project
A practical measurement plan
Agree the commercial outcome that defines success, with finance, before choosing any metric
Fix tracking and consent so the data is complete, lawful and consistent across markets
Run experiments on the largest and least certain lines of spend first
Build or refresh a mix model at least annually and use it to set budgets
Report one set of numbers each month, with the uncertainty stated, not hidden
Measurement will never be perfect. The aim is to be less wrong each quarter and to make budget decisions on evidence rather than on whichever platform reports the best-looking number.
Sources: Google Meridian general availability, Search Engine Land; Meta Robyn open-source project; Google Ads Help, About attribution models; Google Ads Help, About consent mode; Meta Business Help Centre, About Conversions API
