ComparisonGrowth, Marketing & Sales
Marketing attribution models compared: which method to trust for which budget decision
Attribution methods disagree because they answer different questions. Rule-based and multi-touch attribution share out credit among touchpoints, marketing mix modeling estimates how spend drives outcomes across channels, and incrementality experiments measure what a channel actually causes. This comparison sets out what each method needs, where each one breaks, and how to combine them into a measurement system a finance team will accept.
On this page
- Three different questions that all get called attribution
- Rule-based models: cheap, readable and biased in known ways
- Data-driven multi-touch attribution and the signal it is losing
- How marketing mix modeling turns aggregate data into budget curves
- Open-source mix modeling tools and what they still require
- Incrementality experiments and the design mistakes that spoil them
- Side by side: what each method needs and where it breaks
- Triangulation: running a measurement system instead of one model
- B2B attribution: long cycles, buying committees and offline stages
- Which method to rely on for each budget decision
- Questions and answers
- Sources
Three different questions that all get called attribution
The first question is about credit: which touchpoints appeared on the way to a conversion, and how should credit be shared among them? The answer helps a channel manager compare keywords, creatives or audiences.
The second is about causal effect: how many conversions would have happened anyway if the channel had been switched off? Only an experiment, or a model calibrated against one, answers this convincingly.
The third is about allocation: given diminishing returns in every channel, how should next quarter's budget be split? That needs response curves for each channel, including offline media that never appears in a click path. Most measurement arguments start when a report built for the first question is used to settle the third.
Rule-based models: cheap, readable and biased in known ways
Last-touch, first-touch, linear, time-decay and position-based models apply a fixed rule to each conversion path. They are transparent and need only the tracking you probably already have, which suits operational reporting and quick sanity checks.
Their biases are predictable. Last-touch over-credits branded search, retargeting and email, which tend to capture demand created elsewhere; first-touch over-credits awareness channels. Every rule-based model assumes the touchpoints in a path caused the conversion and ignores people who saw the same ads and did not convert.
Data-driven multi-touch attribution and the signal it is losing
Algorithmic multi-touch attribution estimates each touchpoint's contribution from the paths of converting and non-converting users, using techniques such as Shapley values or Markov-chain removal effects. It needs user-level paths that join impressions, clicks and conversions across channels.
Those paths are getting harder to assemble. Consent requirements under the GDPR and ePrivacy rules, browsers that block third-party cookies by default, mobile tracking permissions and walled-garden platforms all leave gaps, and the gaps are not random. The model also still describes correlation: appearing in many converting paths is not proof of causing them. Its best use is comparing options within one well-tracked channel.
How marketing mix modeling turns aggregate data into budget curves
Mix models need no user-level tracking. They relate outcomes to spend and other drivers over time, usually by week and by region.
- Media inputs by week
Spend or impressions per channel, including offline media, at weekly or regional granularity.
- Non-media drivers
Price, promotions, distribution, seasonality and competitor activity.
- Adstock and saturation
The model estimates carry-over and diminishing returns for each channel, often with Bayesian priors.
- Calibration with tests
Results from lift experiments constrain channel effects so the model does not drift from causal evidence.
- Response curves
Estimated contribution and marginal return for each channel at different spend levels.
- Budget scenarios
An optimizer compares allocations within constraints such as minimum brand spend.
Open-source mix modeling tools and what they still require
Two open-source packages have made mix modeling far more accessible. Meta's Robyn is an R package, MIT-licensed, that uses ridge regression with automated hyperparameter search to estimate adstock and saturation.1 Google's Meridian is a Python framework built on Bayesian inference, made available to all advertisers and data scientists in January 2025.23
Free code does not mean a free model. Someone still has to assemble a consistent history of weekly spend and outcomes, choose priors and controls, judge whether results are plausible and refresh the model. Spend that barely varies gives the model nothing to learn from, which is one reason experiments matter.
Incrementality experiments and the design mistakes that spoil them
Geo experiments switch a channel on or off in matched regions, user holdouts withhold ads from a random group, and platform conversion-lift studies apply the same logic inside one ad network. Each gives a causal estimate if the design holds.
The test is too small to detect anything
Early signalConfidence intervals wide enough to include both a large effect and no effect at all.
MitigationRun a power calculation before launch and lengthen the test or enlarge the regions until the expected effect is detectable.
Control regions are contaminated
Early signalNational campaigns, spillover from neighboring regions or sales teams working accounts in both groups.
MitigationChoose regions with little media overlap, freeze other changes during the test and document what ran where.
The outcome window is too short
Early signalA channel looks useless in a test that ends before most buyers would have converted.
MitigationMeasure over the realistic decision cycle, or use a validated leading indicator such as qualified pipeline created.
Side by side: what each method needs and where it breaks
| Criterion | Rule-based | Data-driven MTA | Marketing mix modeling | Incrementality tests |
|---|---|---|---|---|
| Question answered | Who gets credit, by a fixed rule | Who gets credit, estimated from paths | How spend drives outcomes across channels | What a channel or tactic causes |
| Data needed | Click and conversion tracking | User-level paths across channels | Long weekly history of spend, outcomes and controls | A test design, matched groups and patience |
| Granularity | Campaign, keyword, creative | Campaign, keyword, creative | Channel, sometimes region or tactic | One channel or tactic per test |
| Resilience to signal loss | Low | Low | High, as it uses aggregates | High for geo tests; platform studies vary |
| Main bias | Assumes touches cause conversions | Correlation read as contribution | Confounding when spend follows demand | Results apply to the tested conditions only |
| Best used for | Operational reporting | Optimization inside a channel | Annual and quarterly budget splits | Settling disputed channels; calibrating the mix model |
MTA is multi-touch attribution. The ratings describe typical behavior, not a specific vendor's implementation.
Triangulation: running a measurement system instead of one model
- Run lift experiments
Test the channels with the most spend or the most disagreement first.
- Calibrate the mix model
Feed experiment results into the model as priors or constraints.
- Set budgets from curves
Use calibrated response curves for the quarterly or annual split.
- Optimize within channels
Use attribution reports to tune keywords, creatives and audiences day to day.
- Flag disagreements
Where methods diverge sharply, schedule the next experiment.
B2B attribution: long cycles, buying committees and offline stages
Business-to-business measurement breaks most consumer assumptions. A purchase involves several people over months, many touches happen in meetings, events and sales emails that ad platforms never see, and the outcome that matters is revenue recorded in the CRM, not a form fill.
Attribute to opportunities and pipeline stages at account level, using contact roles to connect marketing touches to the buying group. Give mix models lags long enough for the cycle, and use qualified pipeline as the experimental outcome when closed revenue would take too long. A free-text question about how the buyer first heard of you catches word of mouth and communities that no tracking system records.
Which method to rely on for each budget decision
- If
You are splitting an annual or quarterly budget across channels, including offline media.
ThenUse a marketing mix model calibrated with recent experiments.
It is the only method that sees every channel and models diminishing returns.
- If
You need to know whether one channel or tactic is worth its cost.
ThenRun an incrementality test designed for that channel.
A controlled comparison answers the causal question directly.
- If
You are tuning keywords, creatives or audiences inside a channel every week.
ThenUse platform or data-driven attribution and accept its bias.
Relative comparisons within one channel suffer less from what the model cannot see.
- If
Spend is small, channels are few and history is short.
ThenKeep simple rule-based reporting and add a few well-designed holdouts.
A mix model needs variation in spend over a long period to learn anything.
- If
You sell B2B with long cycles and buying committees.
ThenMeasure opportunity-based attribution in the CRM, account-level holdouts and self-reported source together.
Most of the buying journey happens where ad tracking cannot follow it.
Questions and answers
Is multi-touch attribution still worth doing after privacy changes?
Yes, for a narrower job. With consent gaps, cookie blocking and walled gardens, user-level paths are incomplete in ways that skew results toward the channels that track best. Multi-touch attribution remains useful for comparing options inside a single channel. It should no longer be the basis for splitting budget across channels; a calibrated mix model and experiments do that job better.
How much history does marketing mix modeling need?
Enough weekly observations to separate channel effects from seasonality, which in practice usually means a couple of years of consistent data. Variation matters just as much: if spend in a channel never changed, the model cannot estimate what happens when it does. Teams with short histories often start with experiments and add the mix model later.
Can a small marketing team run incrementality tests?
Often, yes. User-level holdouts in email and owned channels cost little, and many ad platforms offer conversion-lift studies. Geo experiments are harder with a small budget because the effect may be too small to detect, so a power calculation before launch tells you whether the test is worth running or needs a longer period.
Should we trust the conversions reported inside ad platforms?
Treat them as one input, not the answer. Each platform counts the conversions it can link to its own ads under its own rules, so platform totals usually add up to more than the conversions you recorded. Use them for in-channel optimization and check spend decisions against your own CRM or order data, ideally with experiments.
Sources
- Robyn: semi-automated marketing mix modeling package — Meta Marketing Science (GitHub) · checked 10 October 2026
- Meridian marketing mix model documentation — Google for Developers · checked 10 October 2026
- Meridian is now available to everyone — Google · checked 10 October 2026