Start from the decision, not the model
Most attribution debates begin with a method and go looking for a use. The productive order is the reverse: list the decisions marketing makes that depend on knowing what caused what, and only then ask which instrument each one needs. The decisions tend to fall into a short list. Which campaigns, ad groups and creatives inside a channel should get more of that channel's budget? Which channels assist and which close, and does the creative and sequencing reflect that? Should the budget for a channel go up, down or to zero? How should next year's budget be split across channels, including the ones that run on television or in print? What should the board see as the return on marketing?
Each of these has a natural instrument. Within-channel questions are answered by the platform's own attribution, because the platform sees all of its own ads and its bidding optimises towards that number. Channel role questions are answered by comparing analytics models, as the attribution models comparison describes. Up-or-down budget questions on a single line need an experiment, because every model describes correlation along a path. Channel-level planning with unobserved media needs a mix model. The board needs the system of record, with the analytics split as supporting detail.
Working from the decision also sets the standard of evidence. A choice between two ad creatives can run on a noisy attributed figure because the cost of being wrong is small and the decision is reversible next week. A choice to cut a channel that costs a large share of the budget cannot, because it is expensive and slow to reverse. The more money and the more irreversibility a decision carries, the further up the stack, towards experiments and models calibrated by them, the evidence has to come from. That single rule resolves most arguments about whether a measurement is good enough. The one-way and two-way door framework sets out how to classify a decision by its reversibility and route it to acting and watching, a test, or research first.
The options and what each is good for
| Option | Answers | Good for | Not for | Cost to run |
|---|---|---|---|---|
| Platform attribution | Which of this platform's ads were associated with conversions | Bids, budgets and creative within the platform | Comparing platforms; reporting totals | None beyond the settings |
| Analytics last click | Which channel produced the final recorded click | Reconciliation across systems; a stable baseline series | Valuing upper-funnel channels | None; every tool produces it |
| Analytics data-driven | How credit splits across recorded touches, estimated from the property's paths | Channel roles; assisted value; default reporting where volume supports it | Causal claims; unobserved media; low volume | None beyond tagging and identity hygiene |
| Model comparison and path analysis | How credit moves between models; which channels appear where in paths | Forming hypotheses about assists and sequencing | Any final answer | Analyst time |
| Incrementality testing | What a channel caused at a spend level in a period | Budget moves on large or contested lines; calibrating other methods | Many channels at once; small channels; impatient calendars | Withheld spend, design effort, weeks per test |
| Marketing mix modelling | Each channel's aggregate contribution and response curve | Annual and quarterly channel planning; unobserved media | Campaign-level choices; new channels; businesses with little history | Data assembly, modelling expertise, months to build |
Two things stand out from the table. The first four options cost almost nothing beyond discipline and are available to any business with an analytics property, so there is no reason for even a small team to rely on one platform's number alone. And the two causal options, tests and models, are complements rather than alternatives: the incrementality testing guide produces the points that calibrate the model, and the marketing mix modelling guide identifies which channels most need testing.
Beside these sits the cheapest signal of all, asking buyers how they heard about the business, and the guide to self-reported attribution shows how to compare those answers with the tracked source and use the gap to decide what to test.
Matching to data, volume and maturity
The decision sets the ideal instrument; the business's situation decides what it can run well today. Three constraints dominate: the quality and coverage of its tracking, the volume of conversions it generates, and the maturity of the team that will own the result.
| Situation | Recommended approach | Why |
|---|---|---|
| One or two paid channels, modest volume, tracking recently set up | One analytics model with last click retained; platform attribution for optimisation | The variation between models is noise at this volume; the priority is clean tagging and a stable series |
| Several paid channels, steady volume, clean tagging | Data-driven as the reporting model; model comparison quarterly; a first geo or holdout test on the largest line | Enough paths for the estimate to settle; enough budget at stake for one experiment a quarter to pay for itself |
| Long consideration cycle, business-to-business or high-value purchase | Analytics attribution on a long lookback with offline conversions imported; tests on proxy outcomes | Paths are long and cross many sessions; purchase-level tests are underpowered, so test on a frequent upstream outcome |
| Meaningful offline or unobservable media | Mix modelling for channel planning, calibrated with geo tests; attribution for digital tactics | No path-based method can credit media that produce no click |
| Multiple platforms each reporting strong returns, total not growing | A reconciliation table, then incrementality tests on the largest lines | Overlapping platform claims are the usual cause; a test settles which claims are real |
| Consent-declined or cross-device share high | Lean on modelled figures where documented, geo tests on total outcomes, and mix modelling | Path-based attribution covers a shrinking population; aggregate methods do not depend on identity |
| Small team, no analyst | One model, one dashboard, one annual test on the biggest line | The approach has to be maintainable by the people who have it |
Maturity deserves a plain word. A data-driven split that nobody on the team can explain to the finance director is less useful than a last-click series everyone understands, and a mix model with no one to refresh it is a one-off report. Choose the most sophisticated approach the team can own, not the most sophisticated one available, and plan the roadmap so that ownership grows with the stack.
Before a mix model's recommendations move budget, the guide to evaluating a marketing mix model sets out the review questions on data, specification, stability, calibration and uncertainty that decide whether it has earned that trust.
A staged roadmap
Adding layers in order protects the reporting the business already relies on while it gains the ability to answer harder questions.
Stage one: make the paths trustworthy. Consistent campaign tagging, a documented channel grouping, key events defined once and matched to the business's key performance indicators (KPIA metric tied directly to an objective.), and offline conversionsUploading outcomes recorded outside the website, such as closed deals, to an ad platform. imported where they exist. Nothing above this stage works without it, and most attribution disputes trace back to a tagging gap.
Stage two: one reporting model, one baseline. Choose the analytics reporting model, data-driven where the tool offers it, restate history under it, and keep last click as the reconciliation baseline. Build the reconciliation table between each platform and analytics and publish it with every report, so that the gaps are explained rather than argued about.
Stage three: read the roles. Run model comparisons and path analysis on a schedule to form hypotheses about which channels assist and which close, and feed those into creative, sequencing and tagging decisions. This stage costs analyst time only, and it produces the shortlist of channels whose budgets are most in question.
Stage four: test the biggest lines. Design and run incrementality tests on the largest or most contested channels, derive an incrementality factor per tested channel, and apply it to attributed figures in routine reporting until the next test. Retest the biggest lines on a cycle; test new channels before they scale.
Stage five: model the mix. When unobserved media are a meaningful share of spend and the history is long enough, build a mix model calibrated against the test results, and use it for annual and quarterly planning. Keep attribution for tactics and tests for calibration. The marketing ROI hub covers how the model's outputs and the test results combine into the return figures the business reports.
A software company's two-year roadmap
A business-to-business software company spending £2.4 million a year across paid search, paid social, events, content syndication and a new podcast sponsorship starts in year one with stage one and two: it fixes campaign tagging across 14 inconsistent naming patterns, imports closed-won deals from its CRM as offline conversions, sets data-driven as the reporting model and restates the previous four quarters under it. The restated series shows paid social credited with 31 per cent of pipeline against 12 per cent under last click, which raises the question of whether to double its £600,000 budget. Rather than act on the split, the team spends the second half of the year on stage four: a geo holdout on paid social across six regional pairs, measuring qualified leads rather than closed deals because deals are too rare to power the test. The holdout finds an incremental cost per qualified lead of £410, within the range the business can afford, and an incrementality factor of 0.7 on the attributed figure. Social spend rises by half rather than doubling. In year two the podcast and events, which produce no clicks, become the reason to assemble weekly history for a first mix model, with the social test result as its calibration point.
Governance: one definition of truth
A measurement stack with several layers produces several numbers for the same channel, and that is correct. What turns it into a problem is letting those numbers circulate without labels. Governance here means a short set of written rules, kept in the measurement plan and enforced in every report.
- One system of record per outcome. Revenue comes from the order system, pipeline from the CRM, conversions for optimisation from each platform, channel splits from analytics. The plan names the source for each figure, and no report substitutes one for another.
- Every figure carries its attribution source and model. "Paid social conversions" is not a number; "paid social conversions, GA4, data-driven, closed month" is. Dashboards label this in the title, not in a footnote.
- One reporting model, changed deliberately. A change of model is a project with a date, a restated history and a note in the change log, never a quiet switch in a settings page.
- Incrementality factors are dated and owned. Each factor records the test it came from, the spend level and the period, and it expires on a schedule.
- The reconciliation table is a standing section. Platform figure, analytics figure, explained gap, every period, for every paid channel.
- Disagreement is recorded, not resolved by seniority. When a platform and analytics disagree about a channel, the table explains why; when the question matters enough, it goes on the test schedule.
The attribution hub holds the model comparator and the overview that new team members read first; the measurement plan holds the rules above. A business that keeps both current rarely has to have the "which number is right" meeting, because every number already says what it is. To check how far a business's current practice is from these rules, the marketing measurement audit scorecard rates the KPI tree, the system of record, the attribution approach and the tests behind the largest budgets, and ranks which gap to close first.
Before booking a first experiment, the comparison of geo holdout tests and platform conversion lift studies helps decide whether the test should be run by the platform or measured independently by region.
How experiments and a mix model divide the work, and how test results calibrate the model, is set out in the comparison of marketing mix modelling and incrementality testing.
Common mistakes
- Choosing the method before the decision. The decision sets the standard of evidence.
- Buying sophistication the team cannot own. An unexplained model loses to a simple series everyone trusts.
- Skipping the tagging stage. Every layer above inherits the gaps.
- Moving budget on a model comparison. Credit shifts are hypotheses; tests are evidence.
- Building a mix model to answer a single-channel question. A geo test is faster and cheaper.
- Circulating unlabelled numbers. The argument about which is right is really about what each is for.
- Letting incrementality factors live forever. They expire with the creative, the audience and the competition.
The attribution models comparison and the incrementality testing guide cover the first layers of the roadmap in detail, and the marketing ROI hub covers how the results become return figures.
