Every analytics tool, ad platform, and CRM will happily tell you which channel 'caused' a sale, and each one will give you a different answer for the same customer journey. This is the central problem of attribution: it is a method for allocating credit across touchpoints, not a measurement of what actually caused the purchase. Understanding that distinction is the single most important thing before you choose a model.
What attribution models actually do
An attribution model is a rule for splitting credit for a conversion among the touchpoints a customer interacted with before converting. If someone sees a Meta ad, later searches your brand on Google, clicks an email link, and then buys, four different systems could each claim 100% of that sale, because each only sees its own touchpoint. Attribution models try to resolve this by applying a consistent allocation rule across a single data set (usually inside one analytics platform).
The common model types
- Last-click: gives 100% of credit to the final touchpoint before conversion. Simple, but ignores everything that built awareness earlier.
- First-click: gives 100% of credit to the first touchpoint. Useful for understanding discovery channels, but ignores what closed the sale.
- Linear: splits credit evenly across every touchpoint in the journey. Fair in theory, but treats a passing glance the same as a decisive click.
- Position-based (U-shaped): weights the first and last touchpoints more heavily, with the rest sharing the remainder.
- Data-driven: uses statistical modeling on your own conversion data to estimate each touchpoint's relative contribution, when enough volume exists to support it.
Why none of these prove causality
Every model above only sees touchpoints that were tracked, inside whichever system is reporting. It cannot see word of mouth, a billboard a customer passed, a review they read on a site without tracking, or the fact that they were already planning to buy regardless of any ad. Platforms also tend to report attribution in a way that flatters themselves: Meta's dashboard will claim credit for conversions that Google Ads' dashboard also claims, because each platform models in isolation. Add these numbers from different platforms together and you will routinely see more attributed revenue than you actually generated.
| Model | Strength | Weakness |
|---|---|---|
| Last-click | Simple, matches most ad platform defaults | Undervalues awareness-stage channels |
| First-click | Highlights discovery channels | Ignores what closed the sale |
| Linear | Spreads credit fairly across touchpoints | Treats minor and major touchpoints equally |
| Position-based | Balances discovery and closing | Still arbitrary weighting of the middle |
| Data-driven | Uses your own data patterns | Needs volume; still platform-bound and correlational |
A pragmatic framework for choosing
Start with the decision you need to make, not the model you've heard is 'best'. If you run a single acquisition channel, last-click inside that channel's own reporting is usually good enough, since there's little cross-channel journey to misattribute. If you run several channels and buyers typically interact with two or more before converting, a position-based or data-driven model inside one consistent analytics tool (not blended across separate platform dashboards) will usually be less misleading than last-click alone.
Build in a reality check
Whatever model you pick, treat it as a hypothesis generator, not a verdict. Validate it periodically with methods that don't rely on self-reported attribution: a holdout test (pause a channel for a defined group or region and watch what happens to overall conversions), a simple brand-lift survey asking new customers how they heard about you, or a geo-based incrementality test. If your attribution model says a channel drives 40% of sales but pausing it barely moves total conversions, the model is overstating that channel's incremental value.
A model-selection checklist
- Write down the specific budget or channel-mix decision the attribution data needs to inform
- Count how many distinct channels a typical customer touches before converting
- If one dominant channel, use last-click within that channel's own reporting
- If multiple channels, centralize conversions in one analytics tool and use position-based or data-driven modeling
- Never sum 'attributed revenue' across separate ad platform dashboards as if it were total revenue
- Schedule at least one holdout, geo test, or brand-lift check per quarter to validate the model against reality
- Document the model's known blind spots (offline, word of mouth, cross-device) so stakeholders don't over-trust it
Common mistakes
- Adding attributed revenue from multiple ad platforms together and presenting it as total company revenue.
- Switching attribution models mid-quarter to make a channel look better, without disclosing the change.
- Treating 'data-driven attribution' as objective truth rather than a statistical estimate built on incomplete, platform-bound data.
- Never validating the chosen model against an experiment, so errors compound silently for years.
- Using a complex multi-touch model with too little conversion volume to produce a statistically stable result.
When this is not the right tactic
If you have very low conversion volume (for example a handful of sales per month), any multi-touch or data-driven model will be unstable and misleading; a simple last-click view plus direct customer conversations will serve you better. If your business sells almost entirely through one channel with no meaningful cross-channel journey, investing analyst time in attribution modeling is not the best use of that time; it will not change any decision. And if leadership intends to use the model's output as a precise financial truth rather than a directional signal, pause and realign expectations before implementing anything more sophisticated than last-click.
Where to go next
Attribution sits downstream of clean tracking and a working measurement vocabulary. If you haven't already, review how conversion tracking is set up and the core measurement glossary, then come back to this framework once you can see a believable (if imperfect) multi-channel data set.
