Attribution for a UK product launch: what one method can and cannot prove

What to take away

Reporting attribution shows which touchpoints were credited before a sale, not what the launch caused.
Multi-touch attribution splits credit across channels but still cannot remove overlap or selection bias.
Randomised incrementality testing, through geo holdouts or split tests, is the only route to a causal launch figure.
Marketing mix modelling covers the whole launch window but cannot isolate one placement or one creative.
Decide the question first, then pick the method, then record which question the number answers.

The fixed case: one UK launch, four questions

Take a UK homeware brand launching a new product in March. The mix covers its own website, one marketplace listing, two retail chains, paid search, paid social and email.

The launch raises four questions. Did it sell? Which channel added sales? What would have happened without the campaign? Where should the next pound go?

Each question carries a different evidence standard. Last-click, multi-touch attribution, incrementality testing and marketing mix modelling each answer one question well and the rest poorly.

What last-click and platform reporting can prove

Last-click in GA4 credits the final click before purchase. Google Ads data-driven attribution and Meta Ads Manager models spread credit across impressions and clicks in the path.

Both describe delivery. They show that a paid social impression or a brand search click sat near a sale. They cannot show the sale depended on it.

Platform numbers also move with consent rates. Under PECR, cookies and similar technologies need clear information and consent, and the ICO direct marketing and cookies guidance sets out what UK marketers must record. Fewer consented users means fewer observed paths, not fewer sales.

Multi-touch attribution and where its evidence stops

Multi-touch attribution in GA4 offers linear, time decay and position-based models. Each is a rule for splitting credit, and each produces a different channel ranking from the same data.

Wikipedia's overview of multi-touch attribution notes the model dependence and the limits of rule-based credit. A launch with heavy upper-funnel activity flatters whichever channel the rule weights most.

Rule-based credit cannot handle cross-channel overlap, cross-device journeys or users who decline tracking. A position-based model has no control group, so it cannot report a counterfactual.

Incrementality testing for the causal question

A controlled experiment compares a group that sees the launch campaign with a matched group that does not. Wikipedia's article on A/B testing covers the design logic behind that split.

For a launch, the practical options are geo holdouts, platform lift tools such as Google Ads Conversion Lift and Meta Conversion Lift, and public service announcement holdouts in paid social. Each needs a clean control and enough volume to detect an effect.

Steps for a launch test:

Incrementality testing

Choose the outcome, for example new-customer orders in the first 28 days.
Split matched regions or audiences into test and control groups.
Hold the control group free of the campaign for the test window.
Compare the difference and report the interval, not only the point estimate.

A holdout proves the effect of the campaign as a package. It rarely separates one creative from another when spend is modest.

What marketing mix modelling adds over the launch window

Marketing mix modelling regresses sales on spend, price, seasonality and distribution across months or years. It reads the whole launch period, including channels that carry no click data, such as television and print.

The method needs long clean history and real variation in spend. A single product launch inside a wider brand gives it little to work with. It guides budget allocation rather than crediting one campaign.

Example: the same launch read three ways

The same launch read three ways

Method

Which channels sat in the path?: Last-click in GA4
How should credit be split?: Multi-touch attribution
Did the campaign add sales?: Geo holdout

What it can show

Which channels sat in the path?: Credited touchpoints before purchase
How should credit be split?: A ranking under a chosen rule
Did the campaign add sales?: Incremental orders in test regions

What it cannot show

Which channels sat in the path?: Causation, or sales lost without the campaign
How should credit be split?: A counterfactual, or full cross-device coverage
Did the campaign add sales?: Which single advert or keyword did the work

Other launch cases worth comparing

For a studio or content-led launch, four ways to attribute studio outcomes ranks four methods by evidence need, bias and strength of conclusion.

A parallel case, women founder funding, weighs experimental, quasi-experimental and contribution methods for funding outcomes.

Paid media teams weighing reporting tools against tests will find platform attribution, experiments and quasi-experiments compared by assumptions and data needs.

For agency work in England, attribution for a website project holds one case constant across credit rules and causal limits.

Common questions

Can last-click ever prove a launch worked?

No. It reports credited conversions. A claim that the launch caused sales needs a control group or a defensible quasi-experimental design.

How long should an incrementality test run?

Long enough to cover the purchase cycle and reach stable daily volumes. Four weeks is a common minimum for fast-moving consumer launches, and longer for considered purchases.

Do UK privacy rules stop incrementality testing?

No. Holdouts compare aggregated regions or audiences, which needs far less personal data than path-level attribution.