Marketing Attribution Models Explained: Which One Should a Small Team Actually Use?

Marketing attribution models are rules, or algorithms, that decide how much credit each touchpoint gets when a customer converts. Google describes an attribution model as a rule, a set of rules or a data-driven algorithm that assigns credit along the path to an important action. The model changes the report. It does not change what customers did.

That distinction matters because the same twelve conversions can make email look like the hero in one model and paid social the hero in another. Before choosing a model, decide which question you are answering: where do journeys start, what closes them, or what is worth more budget. Different questions point to different models, and no model answers all three.

Attribution also depends on what you count as a conversion. A model that credits a newsletter signup will reward different channels from one that credits a signed contract, because the paths to each look nothing alike. Pick the event that matches a business outcome, such as a qualified enquiry, and run the model on that rather than on every event the tag manager can fire.

Finally, consider the window. A lookback window is how far back from a conversion the model will search for touches. A short window suits an impulse purchase; a long one suits a quote that takes weeks to approve. GA4 lets you set the key event lookback window in attribution settings, and changing it changes which channels appear to matter.

First-touch gives all credit to the first recorded interaction, last-touch gives it all to the final one, linear splits it evenly, and position-based gives the ends more weight than the middle (a 40/20/40 split is a common convention, though the weights are a choice rather than a standard). Time decay weights touches nearer the conversion more heavily. Data-driven models estimate each touch's contribution from the paths of both converting and non-converting users.

Rule-based models are easy to explain and easy to argue with. Data-driven models adapt to your data but need enough of it, and they are harder to inspect. Google notes that its data-driven model is specific to each advertiser and each key event, and that it compares converting and non-converting paths using a counterfactual approach.

A quick way to feel the difference is to run one imagined path through each rule. Suppose, as an illustration, a buyer reads a blog post, clicks a social ad, opens an email and then searches the brand and enquires. First-touch gives the blog 100%. Last-touch gives branded search 100%. Linear gives each of four touches 25%. A 40/20/40 position-based rule gives the blog 40%, the ad and email 10% each and branded search 40%. One buyer, four different stories.

GA4 currently offers three models in its attribution reports: data-driven attribution, paid and organic last click, and Google paid channels last click. Google's help page states that the first click, linear, time decay and position-based models are no longer available as of November 2023, and that all models exclude direct visits unless the whole path was direct.

The practical result is that first-touch and linear views no longer come from GA4 reporting itself. If you want them, export the path data to BigQuery or build them in your CRM, where each contact's first source and every later touch can be stored. The reporting attribution setting also controls which channels can receive credit and the key event lookback window, so check those settings before comparing two reports and finding that they disagree.

Last-touch is the usual culprit. Imagine, as an illustration, a service business whose brand searches convert beautifully. Last-touch gives branded search the win, so a manager cuts the content and referral work that made people search the brand in the first place. Six months later branded search shrinks, and the report that caused the cut cannot explain why.

First-touch has the mirror-image fault. It credits the blog post a prospect read two years before a sales conversation, and it cannot see the demo request that actually moved the deal. Linear avoids the extremes but treats a ten-impression retargeting sequence as ten votes. Data-driven models can be the most defensible, yet a model trained on a few dozen conversions a month is mostly noise with a confident label on it.

All of them share a deeper limit: attribution assigns credit among things that happened, so it cannot say what would have happened without a channel. Only a holdout, a geographic split or a pause test speaks to that. Google's own description of its data-driven model leans on randomized-trial data for exactly this reason.

Attribution models also misread offline and untracked influence. A referral from a friend, a conversation at an event or a phone call that starts with "my neighbour recommended you" never creates a click, so every model sends that credit to whatever came last, often direct or branded search. If a large share of your business is referral-led, the report understates the thing that works and overstates the thing that merely recorded it.

Pick one default and keep it for a full quarter. For most small teams that means paid and organic last click for channel reports in GA4, plus a first-source field stored on the CRM contact, so you can see both who closed and who opened. Two simple views that disagree are more useful than one complex view nobody trusts. Switch to data-driven when you have steady conversion volume and someone able to explain it.

For decisions that move real money, add a test. Pause a channel in one region for four weeks, or hold back a slice of an email list, and compare. Write one line in your reporting notes stating what your chosen model cannot see. If you want the build steps for multi-touch tracking, our analytics and dashboards service covers the CRM and BigQuery setup. This is not the right approach for a team with only a handful of conversions a month; there, read the individual sales conversations and ask each customer how they found you.

If you need a rule of thumb for review meetings, use two columns: sourced by (first recorded source on the contact) and closed by (last non-direct touch before the enquiry). Where the columns agree, you can be fairly confident in the channel. Where they differ, you have found a journey worth looking at by hand, and a half-hour of reading real customer paths will teach you more than another dashboard tab.

Most attribution disagreements come from tracking gaps, not from the model. Check that every campaign link carries consistent UTM parameters, that form submissions and calls are recorded as key events, and that the CRM stores the source of each contact at the moment of creation. A model applied to broken inputs produces tidy charts of the wrong thing.

Then write a one-page measurement note: the model you use for each report, the lookback window, which events count, and what you know the data misses. Direct visits get no credit in GA4 unless the whole path was direct, and people who read on one device and convert on another may look like two users. Naming those gaps in advance stops a monthly meeting turning into an argument about tooling. Revisit the note each quarter, or when you add a channel, change your consent banner or move CRM.

A final check is cheap and underused: ask new customers how they heard about you, in a form field or on the first call, and compare the answers with the model. The two will disagree, and the disagreement tells you where the model is blind. Dark social, word of mouth and podcasts rarely show up in click paths at all.