Attribution tries to answer a question that has no clean answer: which touchpoint caused the conversion? Understanding what each model assumes is more useful than picking the one that flatters your channels.
The single-touch models and their bias
First-touch credits the initial interaction and systematically favours awareness channels. Last-touch credits the final one and systematically favours capture channels like branded search, which often harvest demand created elsewhere.
Both are simple and both are wrong in a predictable direction. That predictability is useful: if you know last-touch overstates branded search, you can reason about the gap rather than being misled by it.
Multi-touch models spread the credit
Linear splits credit evenly across touchpoints. Time-decay weights recent interactions more heavily. Position-based typically gives the first and last touches the majority and divides the rest between the middle.
None of these are discovered from your data — they are assumptions you impose on it. Choosing time-decay is a claim that recency indicates influence, which may or may not hold for your sales cycle.
- First-touch: good for understanding demand creation.
- Last-touch: good for understanding conversion capture.
- Linear: reasonable default for long, multi-contact cycles.
- Time-decay: suits short cycles where recency plausibly matters.
- Position-based: sensible when discovery and closing dominate.
What attribution cannot see
Attribution only records touchpoints you can observe. Word of mouth, a podcast mention, a colleague's recommendation and a conversation at a conference are all invisible, and they are frequently the actual cause.
This is why self-reported attribution — simply asking 'how did you hear about us?' on signup — remains valuable. It is messy and biased, and it captures an entire class of influence your tracking cannot.
Use holdouts for causal questions
If the question is genuinely causal — would this conversion have happened anyway — the answer comes from an experiment, not a model. Pause a channel in one region and compare against a comparable one.
Teams consistently find that some channels attributed large volumes were substantially harvesting demand that would have converted regardless. That finding only ever comes from a holdout.
Key takeaways
- Every model encodes an assumption; know which one you are imposing.
- Last-touch flatters branded search; first-touch flatters awareness.
- Ask 'how did you hear about us' to capture untrackable influence.
- For causal questions, run a holdout — models cannot answer them.
Ready to make the switch?
Try Zero Delay Analytics free. No credit card required, and no cookie banner needed.
Get Started
