There are six attribution models you will actually run into, they will give you six different answers about the same conversion, and the disagreement is the useful part. The spread between them tells you how much of your result depends on a bookkeeping choice rather than on what your media did.
Most arguments about attribution are really arguments about which bias someone prefers. So the fastest way through is to price one journey under every model and look at the numbers side by side.
The six models
Single-touch models give everything to one interaction:
- First-touch. All credit to the first interaction in the path. Answers “what introduced this customer to us.”
- Last-touch. All credit to the final interaction. Answers “what was in front of them when they converted.” This is the default almost everywhere, and it has a specific and consistent bias.
- Last non-direct click. Last-touch, but direct traffic is skipped and credit passes to the last identifiable marketing source. It exists because “someone typed the URL” is not an actionable finding.
Multi-touch models split credit across the path:
- Linear. Equal share to every touch. Makes no claim about which mattered more.
- Time-decay. More credit the closer a touch is to the conversion, usually on a seven-day half-life. A compromise that leans bottom-funnel.
- Position-based (U-shaped). Typically 40% to the first touch, 40% to the last, 20% spread across the middle. Built on the assumption that discovery and closing are the hard parts.
- Data-driven. Credit assigned by a model trained on your own converting and non-converting paths, so a touch that appears in both earns less. The only one that is fitted to your business rather than asserted in advance.
The same conversion, priced six ways
One customer, a $1,000 purchase, twenty-one days, five interactions:
- Day 1 — sees a TikTok ad, does not click
- Day 3 — clicks a Google Display ad, browses, leaves
- Day 10 — finds a blog post through organic search, leaves
- Day 20 — clicks a link in a lifecycle email
- Day 21 — searches the brand name, clicks the branded search ad, buys
Four clickable touches. The TikTok view gets nothing in any click-based model, which is the first thing worth noticing.
| Model | Display | Organic | Branded search | |
|---|---|---|---|---|
| First-touch | $1,000 | $0 | $0 | $0 |
| Last-touch | $0 | $0 | $0 | $1,000 |
| Linear | $250 | $250 | $250 | $250 |
| Time-decay | $70 | $140 | $375 | $415 |
| Position-based | $400 | $100 | $100 | $400 |
| Data-driven | $340 | $210 | $300 | $150 |
Same customer. Same revenue. Display is worth either $1,000 or nothing depending on a dropdown.
The data-driven row is the interesting one. It gives branded search the least credit, which looks wrong until you remember what the model is fitted on: branded search appears in nearly every path, including the ones that did not convert, because people who are going to buy search your name either way. A touch that fails to discriminate between converters and non-converters is not predictive, and a data-driven model prices it accordingly.
What each model is biased toward
Pick a model and you have picked a direction to be wrong in:
- First-touch over-credits discovery and tells you nothing about what closes. Optimise to it and you will buy a lot of cheap top-funnel traffic that never converts.
- Last-touch over-credits capture. Branded search and retargeting look spectacular; the demand creation that fed them looks like a cost centre. Cut that, and the capture numbers stay beautiful for about six weeks while the pipeline empties.
- Linear is the only model that makes no assumption, which also means it has no opinion. It flatters long paths: add two more low-value touches and the channel’s share grows.
- Time-decay is a reasonable default for medium-length cycles and still leans bottom-funnel.
- Position-based is the best of the rule-based options when you genuinely do not know, because the 40/20/40 assumption is roughly true for most considered purchases.
- Data-driven is the best-informed allocation you can get, and it is still blind to everything unclickable — the TikTok view, the podcast mention, the colleague’s recommendation.
None of them measure causation. Every one is a credit-sharing rule applied after the conversion already happened. That distinction is the whole game, and it is why last-click lies even when you replace it with something more sophisticated.
How to choose
Three questions, in order.
How long and how crowded is the path? Median time-to-conversion under a day and one or two touches: stay on last-touch, a multi-touch model will cost you machinery and return the same answer. Weeks and five-plus touches: you need multi-touch to allocate at all.
What decision are you making? Splitting budget between two Google Ads campaigns is a fair last-touch question — both sides are measured the same way. Deciding whether upper-funnel paid social deserves to exist is not, and no attribution model will answer it.
How much observable data do you have? Data-driven models need volume. Below roughly 600 conversions a month, a data-driven model is fitting noise and position-based is the more honest choice.
A note on tooling: GA4 no longer offers the choice. First-click, linear, time-decay and position-based were removed from GA4 reporting in 2023; what remains is data-driven and paid-and-organic last-click. If you want to compare models against your own paths, you need event-level data and your own SQL, which is the practical reason the BigQuery export stops being optional at this point.
What to put around the model
A model allocates. It does not prove. Three things belong around whichever one you choose:
One number no platform can inflate. Total revenue over total media spend — MER. Neither input comes from an ad platform, so attribution cannot touch it. Watch it month over month as you shift budget; details in how to calculate ROAS.
One method that establishes causation. Geo holdouts, or turning a channel off in a controlled way and measuring total conversions. This is the only thing on the list that answers “what did this cause.” It costs real revenue to run, and it is still the cheapest answer available.
Revenue connected back to the platforms. Especially in B2B, where the conversion event is not the revenue event. Offline conversion imports and server-side tracking so the optimisation target is the closed deal, not the form fill.
The short version
Attribution models are bookkeeping rules, and the gap between them is a measure of your own uncertainty rather than a question with a correct answer. Use position-based or data-driven to allocate inside paid, depending on your conversion volume. Judge the program on blended efficiency. Establish causation with holdout tests. And treat any single model’s output as one input to a decision rather than a statement about what caused your revenue.
If your platform reports and your bank statement disagree, measurement is the first thing I audit.
