Skip to content
allanninal.dev/build
Collaborate

Part 4 of 7 · Lead source attributor series ~5 min read

How credit gets assigned

Every argument about attribution rules is really an argument about which oversimplification to accept, and it goes better when everyone involved knows that. There is no rule that recovers the truth from the data, because the truth — why somebody actually decided — was never in the data.

Key takeaways

  • Last click is defensible and simple, and undervalues everything that creates awareness.
  • First click values discovery and ignores everything that closed the deal.
  • Reporting both, side by side, is more useful than either and takes no extra work.
  • Linear and time-decay look fairer and mostly move the confusion around.
  • Algorithmic attribution on small volumes is a confident number with nothing behind it.

The four rules

RuleGives credit toGood forBlind to
Last clickThe final touchClosing channels, ad spend decisionsEverything that created the demand
First clickThe first known touchDiscovery channels, content, SEOEverything that converted it
LinearAll touches equallyLong considered purchasesWhich touch mattered
Time decayRecent touches moreShort sales cyclesEarly awareness, mostly

None of these is a model of how people decide. They are conventions for dividing a number, and their value lies almost entirely in being stated. A last-click report labelled “last click” is a useful instrument; the same report labelled “lead sources” is a source of bad decisions.

Run two, not one

The same quarter reported under two attribution rules side by sideA stacked bar chart with two bars, each totalling two hundred and twenty-one enquiries. Five series: Paid search in orange, Retargeting in pink, Organic search in green, Referral and social in purple, and Everything else known in grey. The left bar, Last click, shows paid search at seventy-five, retargeting at forty-nine, organic search at thirty-one, referral and social at twenty, and everything else at forty-six. The right bar, First click, shows paid search at twenty-four, retargeting at three, organic search at sixty-nine, referral and social at forty-two, and everything else at eighty-three. A note says same two hundred and twenty-one enquiries, same data, two rules, and the disagreement is the finding.0100200300400~221Last click~221First clickPaid searchRetargetingOrganic searchReferral and socialEverything else knownSame 221 enquiries, same data, two rules. The disagreement is the finding.
Fig 1. The same quarter under two rules. Paid search takes a third of the credit on last click and a tenth on first click; organic does the reverse. A single-rule report deletes that disagreement entirely.

Reading the two bars together gives the analysis neither gives alone: paid search closes and rarely starts, organic starts and rarely closes, and cutting either one damages the other. That costs nothing extra to produce, because the path is already stored and both rules run over the same rows at report time.

It also changes the conversation from “which channel wins” to “which channel does what”, which is the question somebody actually wanted answered.

The classic mistake it prevents

A last-click report shows organic search at fourteen per cent and paid search at thirty-four, so the content budget gets cut and the ad budget grows. Six months later the ads perform worse and nobody connects the two, because the mechanism — the content was creating the demand the ads were harvesting — is invisible in last-click and obvious in the pair.

The rule that looks most sophisticated

Why algorithmic attribution usually fails at small volumesA vertical chain of five steps entered by a box labelled Use algorithmic attribution, data-driven, it says. Step one asks how many conversions there are per month. Step two asks whether it is fewer than a few hundred, which covers most businesses; if so it exits to Not enough data, because the model fits noise. Step three asks whether you can explain a change when a number moves; if not it exits to Unexplainable, and budgets move on it. Step four asks whether it sees offline touches such as the podcast; if not it exits to No, the same blind spot with more confidence. Step five recommends using the simple rules and saying which. A note says a sophisticated method on thin data produces a precise number, not an accurate one.AWS ACCOUNT"Use algorithmic attribution"data-driven, it saysHow many conversions?per monthFewer than a few hundred?most businessesNot enough datathe model fits noiseyesCan you explain a change?when a number movesUnexplainableand budgets move on itnoDoes it see offline touches?the podcastNosame blind spot, more confidencenoUse the simple rulesand say whichA sophisticated method on thin data produces a precise number, not an accurate one.
Fig 2. Why algorithmic attribution is usually the wrong choice at small scale. Each gate is a practical question rather than a theoretical objection.
  • Compute
  • App integration
  • Machine learning
  • Security & identity
  • Management
  • Analytics

The specific problem with algorithmic attribution at small volume is not that the mathematics is wrong. It is that it produces a number nobody in the business can interrogate, over data with the same blind spots as everything else, and the unexplainability makes it harder rather than easier to notice when it has gone wrong.

A last-click number that somebody disagrees with can be argued about productively, because everyone knows what it means. That is an underrated property.

Where the rule lives

In the report, never in the stored data. A stored attributed_source field is the thing that makes all of this impossible later, because the moment it exists something starts reading it, and then the rule cannot be changed without changing history.

Next: what the report actually says.

All posts