Attribution Modeling
Attribution modeling is the method of assigning credit for a conversion across the various marketing touchpoints a buyer interacted with before converting.
Key takeaways
- An attribution model divides one unit of conversion credit across a time-ordered touchpoint list.
- Lookback window and identity matching change results more than switching between rule-based models.
- Run several models in parallel; the disagreement between them is the useful signal.
- Untracked channels donate their credit to whichever recorded touchpoint happens to come next.
- Only holdout and incrementality tests establish causality; attribution merely describes recorded paths.
In depth
An attribution model is a rule set applied to a time-ordered list of touchpoints. The tracking layer collects events such as ad clicks, organic sessions, email opens and form submissions, stamps each one with a source, medium and campaign, ties them to a single identifier, and sorts them by time. The model then divides one unit of conversion credit across that list. Everything downstream, from channel return on ad spend to cost per qualified lead, is arithmetic performed on the fractions the model hands out.
Two settings move the numbers more than the choice of model itself: the lookback window and the identity resolution method. A thirty-day window discards touches from a ninety-day sales cycle, and cookie-only matching loses journeys that cross devices. Longer windows and logged-in identity produce richer paths, but they cost engineering effort and raise the share of credit flowing to old, cheap touches. The trade-off sits between a model simple enough for finance to trust and one detailed enough to describe how buying really happens.
Most teams run two or three models in parallel instead of picking one. A last-touch view stays in place for platform bidding decisions, a first-touch view for demand-generation targets, and a multi-touch view for quarterly budget planning. Where a scorecard quiz sits in the funnel, tagging its completion as a distinct event lets every model report on it separately, so you can see whether the quiz opens journeys, closes them, or does both. Disagreement between the views is the signal, not a defect to reconcile.
Attribution describes correlation along a recorded path; it cannot say what would have happened without a channel. Word of mouth, offline conversations, podcasts and dark-social sharing never enter the event stream, so their contribution is quietly reassigned to whatever tracked touch came next. Low conversion volume makes any model noisy, and consent banners now remove a growing share of paths entirely. Incrementality tests such as geographic holdouts or deliberate spend pauses answer the causal question that no attribution model can.
Example in practice
How to measure it
Check the plumbing before you judge the model. Track the share of conversions whose path contains at least two recorded touchpoints, the share with no source beyond direct, and the median number of touches per converting path. A high direct-and-unknown share means the model is allocating credit whose origin it cannot see, and no weighting rule will repair that.
Then read the model comparison. Put first-touch, last-touch and a multi-touch view for the same channel into one table and look at the spread. A channel whose first-touch conversions far exceed its last-touch conversions is creating demand; the reverse means it captures demand. Recompute cost per credited conversion under each view and note which budget decisions would flip.
Common mistakes
The frequent error is switching to a new model and then comparing the fresh numbers against last quarter's, which were produced under the old rules. Channel performance appears to jump or collapse for reasons that have nothing to do with the campaigns themselves. Recalculate at least two prior quarters under the new model before anyone presents a trend, and label every chart with the model and lookback window used to build it.
The second failure is letting each team choose its own model. Paid media reports last click, content reports first touch, sales reports whatever the CRM stamped, and the credited conversions across the decks exceed the conversions that actually happened. Agree on one model of record for budget decisions and treat the rest as diagnostic views. Then check the total: credited conversions across all channels must never exceed real ones.
Frequently asked questions
Why does last-click attribution undervalue quizzes?
Last-click gives all credit to the final interaction before conversion, which is usually a branded search or direct visit. Quizzes that introduce and qualify the lead earlier get no credit, making them look less valuable than they are.
How do I attribute a quiz in my funnel?
Tag the quiz as a tracked touchpoint with UTM parameters and pass its identifier into your analytics or CRM. Then compare first-touch and multi-touch views to see how often it starts versus finishes a conversion path.
Which attribution model should I start with?
Start with last touch and first touch side by side, because neither needs a modelling assumption and both can be built from the source stamped on each session. That pair already shows which channels open journeys and which close them. Move to a weighted multi-touch model once monthly conversion volume is high enough that per-channel numbers stop swinging week to week.
How long should my attribution lookback window be?
Set it to roughly the length of your typical sales cycle measured from first known touch to closed deal, then round up. A window shorter than the cycle truncates real paths and pushes credit toward late touches. A far longer one sweeps in unrelated visits. Look at the distribution of time to conversion rather than accepting the platform default.
Why do my ad platform and my analytics tool report different conversion numbers?
Because each applies its own model, window and identity rules. Ad platforms usually credit themselves for any click or view inside their window, count the conversion on the day of the click, and cannot see other channels. Analytics tools credit only sessions they can resolve and count on the day of conversion. Both can be internally consistent while disagreeing.
Does attribution modelling still work without third-party cookies?
Partially. Consent refusals and browser restrictions break the click-to-conversion chain for a growing share of visitors, so recorded paths look shorter and direct traffic swells. Server-side tagging, click identifiers passed in the URL, and asking for an email early, as a quiz form does, restore part of it. Handle the remainder with incrementality testing rather than more modelling.
What is the difference between attribution and incrementality?
Attribution divides credit for conversions that already happened across the touchpoints that were recorded. Incrementality estimates how many of those conversions would not have happened without a given channel, usually by withholding spend from one region or audience and comparing outcomes. Attribution tells you where credit lands; only incrementality tells you whether the spend caused the result.
How do I attribute conversions that involve offline or sales-led steps?
Stamp the original source onto the record at the moment the person first identifies themselves, then carry that field through the CRM to the closed deal. Log sales calls, events and referrals as manual touchpoints on the same timeline. The model is only as good as the discipline behind those stamps, so keep touch types few and clearly defined.