Incrementality testing vs attribution: which should UK brands use?

What to take away

Attribution assigns credit across the touchpoints that appeared before a sale. Incrementality testing measures the sales that would not have happened without the spend.
Attribution is observational, so marketing attribution bias creeps in through consent and tracking coverage.
Lift tests add causal evidence, but only for the channel, audience and period you tested.
Document the lawful basis and the analysis plan before a holdout goes live, not after.

Two questions that look similar but are not

Marketing attribution software reads the paths recorded against conversions, then divides credit using rules or a model. Its output is a share of conversions, expressed as revenue or as a count.

Incrementality testing asks a harder question: what changed because of the spend? A lift test compares an exposed group with a held-back group and reports the difference as conversion uplift.

The two are complements, not rivals. Attribution tells you where to look; a test tells you whether what you found is doing work.

When you line up channels for a budget meeting, start from Comparing channels on common definitions, because an attribution report will not tell you which channel caused the sale.

Why attribution looks more certain than it is

Attribution bias has a simple mechanism. Channels with clean click data collect credit. Channels without it, such as out-of-home or podcasts, collect none. The model is not wrong about the clicks it sees; it is silent about everything else.

Tracking coverage keeps shrinking. Apple's App Tracking Transparency framework requires apps to ask permission before tracking activity across other companies' apps and websites, so users who decline vanish from many mobile paths (Apple Support). Those gaps move into modelled conversions, which are estimates rather than observations.

A dashboard can show a channel taking most of the credit for conversions while its true incremental contribution is much smaller. That is not a reason to abandon attribution. It is a reason to stop reading the percentage as a causal claim.

What each method can and cannot answer

Marketing attribution

What it measures: Credit across observed touchpoints
Evidence type: Observational, historical
Main weakness: Coverage and selection bias
Time to first read: Continuous
Marginal cost: Platform fee and implementation
Best decision: Allocating budget across many channels

Incrementality test

What it measures: Difference between exposed and held-back groups
Evidence type: Experimental, prospective
Main weakness: Sample size and contamination
Time to first read: Two to eight weeks
Marginal cost: Revenue forgone in the holdout
Best decision: Confirming one channel, campaign or audience

Running a lift test under UK data rules

Holdouts are experiments on people, so UK GDPR still applies. Work through these steps in order.

Running a UK-compliant lift test

Write the decision and smallest effect
Pick the randomisation unit
Set sample size and duration
Record the lawful basis
Freeze the analysis plan and read once

Write the decision. State what you will change if the result is positive, flat or negative, and the smallest effect worth acting on.
Pick the randomisation unit. Geographic holdouts work for broadcast and out-of-home; user-level splits suit email and paid social.
Set sample size and duration before launch, using your baseline conversion rate and the effect size from step one.
Record the lawful basis. Consent and legitimate interests are both arguable, but the reasoning must be written down; the ICO's guidance on lawful basis explains how to choose and document one.
Freeze the analysis plan, then read the result once and act on it.

The holdout is also a cost line. Revenue forgone from the control group comes out of the media budget, which is one of the attribution costs the headline charge tends to omit.

If you publish the uplift figure, the CAP Code requires evidence for the claim before it runs, under the advertising codes enforced by the Advertising Standards Authority.

Reading a result without overclaiming

A positive lift is not proof that a channel works everywhere. It is evidence that it worked for that audience at that spend level.

Watch for contamination. If the control group still sees retargeting elsewhere, the holdout is not clean and the measured lift will be understated.

Statistical significance is a low bar. Ask whether the effect is large enough to change next quarter's spend.

Report the interval, not the point estimate. A result of minus two to plus eleven per cent is honest; a headline of nine per cent is not.

Common questions

Does incrementality testing replace attribution?

No. Attribution still shows which campaigns and audiences appear in converting paths. Tests check whether those appearances deserve the credit.

How long does a lift test need to run?

Long enough to reach the sample size set in advance, which usually means several weeks. Stopping early because the chart looks good invalidates the result.

Can a small team run this without a data scientist?

Yes, at a simpler level. A geographic holdout, a fixed duration and a pre-written decision rule will answer most single-channel questions without specialist modelling.

Which should a UK brand start with?

Start with attribution if you have no reliable view of channel mix. Start with a test if you already have a strong performer whose contribution you doubt.