To understand attribution models, start with one customer journey and run it through each model. The differences will become obvious.
Here's the journey: Sarah sees a YouTube pre-roll ad for a software product. Three days later she searches "project management software reviews" and reads a blog post from an organic result. A week after that, she sees a Facebook retargeting ad. Two days later she searches the brand name, clicks a branded search ad, and converts.
Four touchpoints: YouTube impression, organic search, Facebook retargeting, branded paid search. How much credit does each one deserve? Your answer depends entirely on which model you're using.
How five attribution models distribute credit across the same customer journey
Last-click attribution
In last-click attribution, branded paid search gets 100% of the credit. YouTube gets zero. Organic search gets zero. Facebook gets zero.
This is still the default in most analytics platforms, including GA4. It's easy to calculate and easy to explain. It's also systematically wrong in ways that have real consequences for how companies allocate budget.
Last-click rewards the channel that's closest to the sale, not the channel that caused the purchase. For a deeper look at why this matters, see why last-click attribution is misleading. Branded search in this example is capturing demand that already existed, Sarah had already decided to investigate this product before she clicked a branded ad. Crediting that ad with the conversion doesn't tell you much about what drove Sarah to become interested in the first place.
The practical damage: companies that optimize on last-click underinvest in awareness channels (YouTube, display, podcast) because those channels rarely appear in last-click attribution, and over-invest in retargeting and branded search because those channels sit at the bottom of the funnel where conversions are concentrated.
First-click attribution
First-click flips the logic: 100% of credit goes to the first touchpoint, which in Sarah's journey was the YouTube impression.
This is better for understanding what triggered initial discovery, but it has an obvious bias in the other direction. The YouTube ad may have introduced Sarah to the brand, but she still needed the review content, the retargeting reminder, and the branded search to actually convert. First-click ignores all of that.
First-click is occasionally useful for measuring top-of-funnel reach and for understanding where new customer discovery happens. It's not a good model for budget allocation decisions that span the full funnel.
Linear attribution
Linear attribution distributes credit equally across all touchpoints. In Sarah's journey, each of the four touchpoints gets 25%.
This feels fair. It acknowledges that multiple touchpoints contributed. But "fair" and "accurate" aren't the same thing. Linear attribution has no opinion about which touchpoints were more valuable, a one-second YouTube impression gets the same credit as a ten-minute review article that answered Sarah's specific questions and drove her decision.
It's better than single-touch models, but it treats all touchpoints as equivalent when they clearly aren't.
Time-decay attribution
Time-decay gives more credit to touchpoints that happened closer to the conversion. The branded search ad gets the most credit. Facebook retargeting gets a bit less. Organic search gets less still. YouTube gets the least.
This model has an intuitive appeal, if something happened right before the purchase, it probably had more influence. The problem is that it's encoding recency bias as a measurement truth. A brand campaign that took someone from "never heard of this company" to "actively interested" happened at the beginning of the journey for a reason. Time-decay systematically discounts that early influence.
Position-based (U-shaped) attribution
Position-based gives 40% credit to the first touchpoint, 40% to the last touchpoint, and divides the remaining 20% among everything in the middle.
This is a reasonable heuristic for marketers who believe that first discovery and final conversion are the two most important moments. It's more nuanced than single-touch models and less arbitrary than linear. But it's still a rule-based system with numbers that someone made up rather than estimated from data.
Data-driven attribution (DDA)
Data-driven attribution is the approach that sounds most scientific: it uses machine learning to assign credit based on observed conversion paths in your data, rather than fixed rules.
The basic idea is to compare paths that converted versus paths that didn't, and identify which touchpoints were present in converting paths but absent in non-converting paths. Those touchpoints get more credit.
DDA sounds like an obvious upgrade, but it has serious limitations in practice. First, it requires a lot of data, Google's implementation requires at least 3,000 conversions per month per conversion action to function properly. If you're below that threshold, the model falls back to last-click.
Second, DDA can only work with what it can see. It doesn't see the TV ad Sarah watched. It doesn't see the podcast she listened to. It sees the touchpoints that left a digital trace, which means it's still systematically undercounting offline and cookieless channels.
Third, and most importantly: DDA identifies correlation between touchpoints and conversions. It can't prove causation. If your target audience happens to click many branded search ads before converting, DDA will give branded search high credit, even if those clicks aren't causing conversions, but are just a behavior common to people who were already committed to buying.
Multi-touch attribution
Multi-touch attribution (MTA) is not a separate model. It is a category. Every model covered above, except last-click and first-click, qualifies as multi-touch: it assigns credit to more than one touchpoint rather than to just one.
In practice, "MTA" usually refers to a specific kind of measurement infrastructure: a cross-channel data collection layer that captures touchpoints from multiple platforms and applies a unified attribution model across all of them. Tools like Northbeam, Triple Whale, and Rockerbox are built on this premise.
The value of MTA tools is not the model itself. It is the data layer. Relying on each platform's native attribution means comparing Facebook's self-reported numbers against Google's self-reported numbers, with no unified view and no way to deduplicate credit. An MTA tool collects first-party pixel data and shows you the full multi-channel journey in one place, which is already a significant step forward from siloed platform dashboards.
That said, the limitations are the same as any attribution approach: MTA can only measure what it can track. Ad blockers, cross-device journeys, iOS privacy restrictions, and offline touchpoints all create gaps. Those gaps tend to be larger than most marketers assume, often 30 to 50% of actual customer interactions leave no trackable trace. The model is operating on an incomplete picture of the journey, and it cannot know what it cannot see.
Unified marketing measurement
Unified marketing measurement (UMM) is an attempt to solve attribution's structural limitations by combining it with media mix modeling. The idea: use multi-touch attribution for granular, near-real-time optimization signals and use MMM for strategic budget allocation, then reconcile the two so they tell a consistent story.
In theory, this is the right approach. MMM gives you channel-level ROI estimates grounded in causal modeling and aggregate data, no cookies required. MTA gives you touchpoint-level signals useful for creative testing, audience segmentation, and intra-channel optimization. Together, they cover both the strategic and the tactical layer of measurement.
In practice, UMM is hard to do well. The two methods often disagree, sometimes significantly. When MMM says Facebook contributes 8% of sales and MTA attributes 23% of revenue to Facebook, reconciling those numbers requires not just technical integration but judgment about which method to trust in which context.
The vendors who offer UMM (Analytic Partners, Nielsen, and others at the enterprise end of the market) typically use MMM outputs to calibrate or constrain the MTA model, anchoring the attribution to a more statistically robust baseline. This is better than either method alone. But it is also expensive, and the quality of the output depends heavily on the quality of both the MMM and the MTA inputs.
For most companies, the practical path to unified measurement is not a single UMM platform. It is building the measurement stack incrementally: get attribution working first, add incrementality testing to calibrate it, then layer in MMM for strategic budget decisions. That sequencing delivers most of the benefit at a fraction of the cost of a full UMM implementation. Full UMM makes sense once the components are mature and the budget justifies the integration investment.
The honest bottom line
Attribution models are useful as optimization signals within paid channels. If your DDA model shows that email performs well in the middle of customer journeys, that's worth knowing, it might influence your sequencing or messaging strategy. If last-click shows that a specific ad creative has a high conversion rate, that creative feedback is valuable even if the attribution of cause is imprecise.
What attribution models should not do is drive major budget allocation decisions. The question "should we invest more in YouTube awareness or in retargeting?" cannot be answered by attribution alone, because attribution systematically biases toward the channels that are easiest to measure and closest to conversion.
For budget allocation, you need MMM and incrementality testing. Attribution gives you the operational data to optimize within campaigns. It's not the right tool for the strategic decisions where the most money is on the table.
The mistake most marketing teams make isn't using attribution models, it's using them for the wrong questions.