Marketing attribution attempts to answer which activity caused a sale. The honest position is that this is a causal inference problem being solved with correlational data.
The standard models
Last click assigns all credit to the final touchpoint before conversion.
Which systematically favours channels appearing late in the journey — brand search, retargeting, direct — and undervalues those that created awareness.
First click does the opposite, favouring discovery channels and ignoring what closed the sale.
Linear splits credit evenly, which assumes every touchpoint contributed equally, which is obviously false.
Time decay weights recent touchpoints more heavily, which is a guess dressed as a model.
None of these is derived from anything. They are conventions.
Why the whole approach is shaky
Attribution asks what caused a conversion, using data showing only what preceded it.
Which cannot distinguish between a channel causing a purchase and a channel being visited by people already intending to purchase.
Brand search is the clearest case — someone searching for your company name has already decided, and the channel receives credit for capturing demand it did not create.
Which is why paid brand search frequently shows excellent attributed returns and produces little incremental revenue.
Incrementality testing
The approach that actually answers the question.
Withhold activity from a randomly selected group and compare outcomes.
Which is a genuine experiment and produces a causal answer rather than a correlational one.
Geographic holdouts, where activity is suspended in some regions, are practical for channels that can be targeted geographically.
The results frequently differ substantially from attributed figures, generally showing less incremental effect than attribution claimed.
Media mix modelling
Statistical modelling of aggregate spend against aggregate outcomes over time.
Which does not require individual tracking and has become more prominent as tracking has been restricted.
It handles the awkward things attribution ignores — diminishing returns, carryover effects from previous periods, and interactions between channels.
Its limitations are that it requires substantial historical variation to identify effects, and results are sensitive to model specification.
The tracking restrictions
Privacy changes have removed much of the data attribution depended on.
Third-party cookie restrictions, mobile tracking permissions and privacy regulation have all reduced cross-site and cross-app visibility.
Which has pushed the field toward aggregate methods, first-party data, and modelled conversions where platforms estimate what they cannot observe.
Modelled conversions are estimates presented alongside measured ones, and the distinction is frequently not clear in reporting.
What to actually do
Use attribution for directional operational decisions where the alternative is nothing.
Run incrementality tests on the largest spend lines, since that is where being wrong costs most.
Track a small number of business metrics — total revenue, new customers, cost per acquisition at the business level — because those are real regardless of how credit is assigned internally.
And be suspicious of any channel showing extraordinary attributed returns, since that pattern generally indicates the channel is capturing demand rather than creating it.
Consent and data quality
Where measurement gets harder before it gets easier.
Consent requirements mean a proportion of visitors are not tracked at all, and the proportion varies by region and by how the consent request is presented.
Which produces systematically incomplete data, and the incompleteness is not random — people who decline tracking differ from those who accept.
Server-side measurement, where events are recorded by the site rather than by browser scripts, improves reliability and does not remove the consent requirement.
Offline conversion
Businesses where the purchase completes by phone, in person or through a sales process face an additional gap.
Which requires connecting the eventual outcome back to the original source, generally through a customer identifier carried through the process.
Doing this properly is unglamorous data work and it is what makes measurement possible at all for those businesses.
The simple check
Total spend against total new customers over a period, computed at the business level, is immune to attribution problems.
It is crude and it is real, and comparing it across periods when spend changed tells you more than most dashboards.
Brand search specifically
Worth a targeted experiment, since it is the clearest case of attribution overstating value.
Pausing paid brand search and measuring whether total traffic and conversions fall answers the question directly.
Results vary — where competitors bid on your brand name, the paid listing defends a position that would otherwise be taken, and where they do not, the organic listing captures most of the traffic.
Which means the answer is business-specific and testable, and it is tested far less often than the spend justifies.
Reporting internally
Presenting attributed figures alongside incrementality results, with the difference explained, is more honest than choosing one.
Which requires accepting that the numbers previously reported were overstated, and that political difficulty is the main obstacle to better measurement in most organisations.
Which is why measurement improvements tend to arrive with a change of leadership rather than through internal persuasion.