Attribution determines where marketing budget goes, and every model available makes assumptions that affect the answer.
The problem
Customers encounter several touchpoints before converting.
Which means assigning credit requires a rule.
Every rule is a choice rather than a measurement.
Last click
All credit to the final interaction.
Which is simple and systematically overcredits channels that appear late.
Search on a brand name captures credit for demand created elsewhere.
First click and linear
Credit to the first interaction, or spread evenly.
Which correct one bias and introduce others.
None of them represents actual causation.
Incrementality testing
Turning spending off in some regions or audiences and measuring the difference.
Which measures causal effect rather than correlation.
It is the only approach that answers what would have happened otherwise.
Geographic holdouts
Suspending activity in matched markets.
Which produces a control group.
This is the most practical incrementality method for most companies.
Media mix modelling
Statistical modelling of aggregate spending against outcomes.
Which does not require individual tracking.
It has returned to prominence as tracking became more constrained.
Privacy changes
Restrictions on cross-site tracking have reduced attribution accuracy substantially.
Which has pushed the industry toward modelling and experimentation.
Platform-reported conversions increasingly rely on modelled estimates.
Platform self-reporting
Advertising platforms reporting conversions attributable to their own advertising.
Which systematically exceeds what independent measurement shows.
Summing platform-reported conversions frequently exceeds total actual conversions.
What to actually do
Use attribution for directional signal, run incrementality tests for decisions, and treat platform reporting as a claim rather than a measurement.
Brand search
Paid search on your own brand name.
Which attribution credits generously and incrementality testing frequently finds adds little.
This is one of the clearest cases where the two methods disagree.
Long consideration cycles
Purchases considered over months.
Which attribution windows generally do not span.
Business purchases are systematically under-attributed to early-stage activity.
Offline activity
Television, outdoor and print.
Which produce effects that click-based attribution cannot capture at all.
Media mix modelling and geographic testing address these.
Organisational effects
Attribution determines budget, which determines team size and priorities.
Which gives people an interest in the model chosen.
The honest approach
Test rather than attribute where decisions matter, and treat platform figures as claims.
Running a holdout test
Suspending activity in matched markets for a defined period.
Which requires enough scale to detect an effect.
Statistical power calculations determine how long a test needs to run.
Interpreting results
Incrementality frequently shows lower effect than attribution.
Which is uncomfortable and is the more accurate figure.
Budget decisions based on attribution alone systematically overspend on some channels.
Aggregate measurement
Modelling total marketing spend against total outcomes.
Which does not require individual tracking.
Open-source implementations of these models are available.
Practical cadence
Attribution continuously for direction, incrementality tests periodically for decisions.
The honest position
Nobody knows exactly which marketing works, and testing narrows the uncertainty considerably.
What attribution models do
Assign credit for a conversion across the touchpoints preceding it.
Which is a modelling choice rather than a measurement.
Different models produce different answers from identical data.
Last-click
All credit to the final touchpoint.
Which is simple and systematically favours channels that appear late.
Branded search consistently benefits from this.
Multi-touch
Credit distributed across touchpoints by rule or model.
Which is more defensible and still a set of assumptions.
It cannot distinguish influence from coincidence.
The incrementality question
Whether the conversion would have happened without the spend.
Which no attribution model answers.
Only a controlled test does.
Tracking limitations
Cookie restrictions, app tracking rules and privacy regulation.
Which have removed a substantial share of the data attribution relies on.
Modelled gaps now fill much of what was measured.
Offline and unmeasured effects
Word of mouth, physical presence and brand memory.
Which attribution never captured and which drive real behaviour.
Their absence biases spending toward what is measurable.
The measurable bias
Channels that report well attract budget regardless of true effect.
Which is a structural distortion in digital marketing.
Retargeting is the clearest example.
Platform-reported figures
Each platform claims credit under its own rules.
Which is why claimed conversions frequently exceed actual ones.
Summing platform reports is a common error.
Where to start
Pick one significant channel and run a holdout test on it.
Which will produce an uncomfortable and useful number.
One honest test beats a year of attribution dashboards.
A closing caution
None of this is prescriptive. Businesses differ by sector, by scale and by stage, and practices that work well in one context fail in another for reasons that are not always visible from outside.
What is consistent is that the businesses handling these questions well tend to have written something down, measured it in a defined way, and reviewed it on a schedule rather than when a problem forces the issue.
Where a decision carries legal, tax or employment consequences, professional advice specific to your jurisdiction is worth the cost, and this article is general description rather than advice.