How to vet a media buying agency's attribution method

Choosing between media buying agencies means testing the attribution method each one proposes, by the decision it must support and the data it can honestly use. Use platform attribution for work inside that platform. A shared rules-based view can reconcile channels. A causal question needs a suitable experiment, provided the design can isolate a meaningful difference. These approaches can coexist because they answer different questions.

What to take away

Compare the question before the model

Platform-reported attribution asks which interactions its system associates with a conversion. It is quick enough for campaign management and retains platform-specific signals. Its boundary is also its weakness: another publisher may claim the same outcome, while neither report proves that advertising caused it.

Common platform reports include Google Ads last-click, GA4 data-driven attribution, and Meta Ads 7-day click and 1-day view.

A cross-channel rule, such as assigning an outcome to the last recorded eligible contact, gives the organisation one consistent convention. Other common rules are first-click, which credits the first recorded eligible contact; linear, which spreads credit evenly across recorded contacts; time-decay, which weights contacts nearer the outcome more heavily; and position-based, which favours the first and last contacts. Each is reproducible when identity, eligibility and the lookback period are fixed, and each still expresses a choice. None of them discovers the true influence of every exposure.

Illustrative example: a UK retailer spends 40,000 pounds a month on paid search, paid social and email. A last-click rule assigns 65 percent of recorded sales to paid search, 20 percent to paid social and 15 percent to email.

A geo test holds paid social off in half of matched regions for four weeks. If sales per capita in test regions fall by 8 percent against control, the last-click figures overstate paid social's incremental effect.

An experiment asks whether outcomes differed under a planned comparison. That is the appropriate direction when the decision concerns incrementality rather than credit allocation. It needs a written hypothesis, suitable comparison, sufficient data, contamination checks and pre-agreed analysis. If those conditions cannot be met, describe the result as attribution or association instead of causation.

As a rough planning range, an experiment needs at least 100 conversions per arm to detect a large effect and several hundred per arm to detect a modest one. Below that, treat the result as directional.

Check whether the data can support the choice

Inventory the observable events before selecting a model. Google's global conversion tracking guidance documents website, app, phone and offline routes in Google Ads. That first-party record shows configuration options within the product. It cannot validate the advertiser's event definition or settle credit across other media.

Mobile data has platform conditions too. Apple's global user privacy and data use documentation explains when AppTrackingTransparency permission is required for tracking across other companies' apps and websites in Apple's ecosystem. Do not fill a resulting observation gap by assuming unobserved users behaved like observed users. Record what the dataset excludes.

For organisations operating in England, the practical requirement is a lawful basis for any storage or access on a user's device. The Data Protection Act 2018 complements the UK General Data Protection Regulation, and the Privacy and Electronic Communications Regulations (PECR) sit alongside them. Under UK GDPR and PECR, user consent is already needed before behavioural advertising can take place. Non-essential cookies and similar identifiers used for advertising or measurement must not be set before consent, unless a strictly necessary exemption applies.

The ICO's online advertising guidance, including its guidance on the use of cookies and similar technologies, explains how these rules apply to online advertising. A qualified UK reviewer should confirm the proposed identifiers, consent approach and data sharing. A model's technical availability does not establish that a particular use is lawful.

Use a method-selection sequence

First, state the decision. Daily bid or creative management may use the relevant platform report with its label intact. Budget allocation across channels needs a common outcome definition and a reconciled cost basis. A claim that media generated additional outcomes calls for an experiment or a marketing mix modelling (MMM) design, not a more complicated credit rule. Ask each agency you are considering which of these questions its proposed method answers.

Next, grade coverage. Note which media, devices, conversion routes and customer records can be observed under the permitted process. Mark reporting delays, duplicate claims and unmatched events. Then choose the least complex method that answers the question within those limits.

Check timing as well as event coverage. A short lookback may omit later outcomes, while a longer one can admit more competing activity. The agreed period should follow the decision and customer journey, not whichever default produces the largest total.

Finally, write a method card. Include the eligible touchpoints, outcome, lookback period, identity rule, model or allocation rule, exclusions, source versions and change date. Where the design is econometric, record the outcome measure, the media and control variables, the calibration period and the validation approach. Run a sensitivity check where another reasonable rule could alter the decision.

Questions to put to a shortlisted agency

Ask each media buying agency on your shortlist to hand over the method card, not a slide that summarises a dashboard. A platform-based answer should name the platform reports and keep their labels; a cross-channel answer should name the allocation rule and the lookback period; an incrementality answer should name the design, the comparison group and the pre-agreed analysis. Ask what the dataset excludes, and how the agency would know if its own figures overstated its effect. An agency that cannot answer those questions is offering a convention, not evidence.

State the result without overreach

Report platform totals as platform-attributed outcomes. Call a rules-based allocation an allocation. For a test, publish the design, period, exclusions and uncertainty with the finding. If two methods disagree, do not average them into a smoother number. Explain the different questions and decide which one governs the action at hand.

Before you act

Common questions

When is an experiment the right choice over a rules-based attribution model?

An experiment is appropriate when the decision concerns incrementality rather than credit allocation. It asks whether outcomes differed under a planned comparison. It requires a written hypothesis, suitable comparison, sufficient data, contamination checks and pre-agreed analysis. If those conditions cannot be met, describe the result as attribution or association instead of causation.

What data coverage does an experiment need?

A rough planning range is at least 100 conversions per arm for a large effect and several hundred per arm for a modest one. Below that, treat the result as directional. The test also needs a stable comparison group and checks for contamination.

How should you handle gaps in mobile data due to privacy restrictions?

Do not fill a resulting observation gap by assuming unobserved users behaved like observed users. Record what the dataset excludes. Apple's documentation explains when AppTrackingTransparency permission is required for tracking across other companies' apps and websites in Apple's ecosystem. A qualified UK reviewer should assess identifiers, consent and data sharing.