First touch vs last touch vs multi touch attribution: how each model works, and which to use
First-touch attribution gives all the credit for a conversion to the customer's first tracked interaction. Last-touch gives it all to the final interaction before purchase. Multi-touch attribution splits the credit across several interactions on the path. First touch suits awareness analysis, last touch suits conversion analysis, and multi touch suits longer paths.
Each model moves credit toward a predictable set of channels, so the choice changes how channels compare without changing how many conversions there were. This page explains each model with what it is best used for and its main drawback, works through one example path under each model, covers the multi-touch variants including data-driven attribution, shows what the models have in common, and sets out how to choose one and when to use a different kind of measurement.
First-touch attribution
How it works. The whole conversion is credited to the first tracked interaction a customer had with the brand, for example the first ad click or the first visit from a search.
Best used for. Understanding which channels introduce new customers and generate demand at the top of the funnel.
Drawback. Everything after the first interaction receives no credit, so nurturing emails, retargeting and the channels that close the sale look unproductive. It also depends on the first interaction being tracked, which is less likely on long paths or across devices.
Last-touch attribution
How it works. The whole conversion is credited to the final tracked interaction before purchase. It is the default in many analytics tools, so it shapes more budgets than any other model.
Best used for. Understanding which channels close sales, and optimising the final steps of a funnel.
Drawback. Everything before the final interaction receives no credit, so awareness and consideration activity looks unproductive. Channels close to the purchase, such as brand search and retargeting, receive credit for customers who had already decided to buy.
Multi-touch attribution and its variants
Multi-touch attribution splits each conversion across several tracked interactions. The main variants differ in how they weight them.
Linear splits credit evenly across every interaction.
Time decay gives more credit to interactions closer to the purchase.
Position-based, also called U-shaped, gives a large share to the first and last interactions and divides the rest among those in between.
Data-driven learns weights from the paths that did and did not convert, instead of applying a fixed formula.
Best used for. Longer paths with several touchpoints, where single-touch models ignore most of the journey.
Drawback. It needs reliable tracking across the whole path, which consent limits and multiple devices make harder. Fixed-weight variants also rest on an assumption about which interactions matter most.
The three compared
An example. Take one customer who clicks a social ad, later reads a blog post from a search, opens an email and finally buys after clicking a brand search ad. First touch gives the whole sale to the social ad. Last touch gives it to brand search. Linear gives each of the four a quarter. Position-based typically gives 40% each to social and brand search and 10% each to the blog visit and the email. Same sale, four different pictures of which channel mattered.
Where the credit goes. First touch credits discovery channels such as broad-reach social and display. Last touch credits channels close to the purchase such as brand search, retargeting and email. Multi touch spreads credit between the two.
Complexity. Single-touch models are simple to explain and to reproduce. Multi-touch models need more data and, for data-driven attribution, a model that has to be trusted.
Tracking needs. All three depend on tracked interactions. Multi touch depends on tracking the whole path, single-touch models on tracking one end of it.
Stability. Single-touch models change little when a path gets longer. Multi-touch weights can shift as paths change, which can look like a change in performance.
What all three have in common
All three models start from the same set of tracked conversions and the same tracked interactions, and they distribute exactly that set. The model choice changes how channels compare with each other. It leaves the total unchanged.
Two things fall outside every model. Conversions that cannot be tracked, such as in-store purchases, those influenced by connected TV or radio, and those on untracked devices, are not credited to anything. And conversions that would have happened without marketing are still credited to whichever interaction appeared on the path.
For that reason, no attribution model can report that a channel added nothing. Every tracked interaction on a converting path receives some credit under some model.
Data-driven attribution
Data-driven attribution is often presented as the answer to the weaknesses of rule-based models. It does improve the split: credit follows the evidence in the paths instead of a formula, and the choice of weights stops being arbitrary.
It still works within the same tracked conversions and interactions, so the two gaps above remain. It also needs a large volume of conversions to learn stable weights, so it is less reliable for low-volume or long-cycle businesses, where each path is rare and slow to complete.
Teams that move to it expecting an answer to the budget question across channels usually find that question still open, because it needs a measure of total revenue.
How to see what a switch would change
Before changing model, export the same period under the current model and the proposed one and compare the split by channel. The difference is the size of the decision.
Two patterns appear in most accounts. Brand search and retargeting lose share under models that credit earlier interactions, and broad-reach social and display gain it. If your account does not show this, either paths are shorter than expected or early interactions are not being tracked. Both are useful to know before the discussion starts.
When each model is the better choice
First touch is the better choice for questions about acquisition: which channels bring new customers into the funnel.
Last touch is the better choice for questions about conversion: which channels close sales, and where the last step of the funnel loses people. It is also the easiest to keep consistent over time.
Position-based or data-driven multi touch is the better choice for short, well-tracked digital paths with several steps, where it describes the whole journey more fairly.
A different method altogether is the better choice when the question is how much each channel added to total revenue, or how to split a budget across channels. That needs contribution measurement, usually a marketing mix model.
Business type matters too. Ecommerce paths are often short and mostly digital, so multi-touch models describe them reasonably well. B2B and high-consideration purchases run over months and across several people, and much of the path happens in meetings, events and word of mouth that no model tracks. There, single-touch models are easier to keep honest, and the budget question relies even more on contribution measurement.
What attribution models cannot tell you
No attribution model measures what a channel caused. Each one distributes credit among interactions that occurred on converting paths.
None of them sees untracked conversions or separates out the sales that would have happened anyway, so none of them can show that a channel added nothing.
None of them covers offline channels in a way comparable to digital ones.
For those questions, the tools are a marketing mix model, which estimates each channel's contribution to total revenue, and experiments, which measure the effect of one channel directly.
How to choose
Choose a model for operational reporting, and use it consistently.
For a short, fully tracked digital path, use data-driven attribution if your platform supports it, and position-based if not.
For a long or partly offline path, use the simplest model you can explain, usually last touch, and read it as an operational signal.
For budget allocation across channels, use contribution measurement calibrated with experiments.
Change the model as rarely as possible. Every historical comparison breaks on the day it changes.
Cassandra is a marketing mix modeling platform, and it answers the question attribution models leave open: how much each channel added to total revenue. It decomposes the whole outcome, with the baseline as its own component.
Its production engine is Bayesian, so each channel's share comes with its uncertainty stated. Geo experiments run in the same system, using the same method family as GeoLift. Reads sit at campaign level, in the planning layer, alongside whichever attribution model a team keeps for daily work.
A price is published.
Changing attribution model
The main cost of a change is comparability. Year-on-year channel performance, campaign benchmarks and targets set against last year's figures were all calculated under the old model and do not carry over.
If a change is worth making, run both models in parallel for a full reporting cycle first, so there is at least one period where the two can be compared directly and the change can be explained.
Decide in advance what the new model is for. A change made to answer a budget question across channels will not answer it, because that question needs a measure of total revenue.
Tell the people who read the reports before the switch, and show them the side-by-side period. A channel owner who sees their share fall overnight without warning will read it as a performance problem, when the only thing that changed was the model.
Questions
What are the key differences between first click and last click attribution?
First click gives all the credit to the first tracked interaction, so it favours channels that introduce customers. Last click gives it all to the final interaction, so it favours channels that close sales.
What are the four types of attribution?
The most common are first touch, last touch, linear and time decay. Position-based and data-driven are also widely used, and the linear, time-decay, position-based and data-driven variants are all forms of multi-touch attribution.
What is the drawback to using the last touch attribution model?
It gives no credit to anything before the final interaction, so awareness activity looks unproductive, and channels close to the purchase receive credit for customers who had already decided to buy.
Which attribution model is most accurate?
All of them divide the same tracked conversions, so none measures what a channel caused. Data-driven attribution divides the credit more defensibly than a fixed formula on well-tracked paths.
How often should we change attribution model?
As rarely as possible. Every historical comparison breaks on the day the model changes, and consistency over time is usually worth more than a marginally better model.