Marketing attribution models and how to pick one for each decision

A marketing attribution model is a rule for deciding which marketing touchpoints get credit for a conversion. Someone sees a video ad, later clicks a search ad, then comes back through an email and buys. Attribution decides how much of that sale belongs to the video, the search ad and the email.

Teams often argue about which model is correct. That argument rarely ends, because no model is correct. Each one is a simplification built to answer a particular kind of question. The useful skill is knowing which model to look at for which decision, and knowing where all of them mislead.

The common attribution models

Last click

All credit goes to the final touchpoint before the conversion. It's simple, easy to explain and easy to audit. It also heavily favours channels that catch people at the end, like branded search and retargeting, and gives nothing to whatever made them aware of you in the first place.

First click

All credit goes to the first recorded touchpoint. It's the mirror image of last click, useful for seeing what introduces people to the brand, and blind to everything that closed the sale.

Linear

Credit is split evenly across every recorded touchpoint. It's fairer in spirit, but it treats a passing impression and a deliberate product search as equal, which they rarely are.

Time decay

Touchpoints closer to the conversion get more credit than earlier ones. This suits businesses with short consideration periods, where recent activity really does matter more. For long B2B sales cycles, it tends to undervalue the work done months earlier.

Position based

A larger share goes to the first and last touchpoints, with the rest spread across the middle. It's a compromise that tries to credit both discovery and closing.

Data-driven

The platform or analytics tool uses its own data to estimate how much each touchpoint changed the chance of converting, by comparing paths that converted with paths that didn't. It's usually more nuanced than rule-based models. It's also a black box, and it can only learn from touchpoints the tool can see.

Why platforms over-count

If you add up the conversions each ad platform reports and compare the total with what your CRM or store shows, the platform total can come out higher. This happens for a few structural reasons, none of which require anyone to be dishonest.

Each platform credits itself. Meta, Google, LinkedIn and TikTok each look at their own touchpoints and claim the conversion if their ad was in the path within their attribution window. When a customer saw ads on three platforms, all three can report the same sale.

View-through conversions add more. Some platforms count a conversion when someone saw an ad without clicking and later converted. Some of those people were influenced. Others would have bought anyway, and the platform can't tell the difference.

Modeled conversions fill gaps. When tracking is limited by browsers, consent choices or app privacy settings, platforms estimate conversions they couldn't observe directly. The estimates may be reasonable on average, but they're estimates, and they're produced by the party being graded.

Attribution windows differ. One platform might count a conversion days or weeks after a click, another uses a shorter window, and analytics uses its own rules. The same sale can be inside one window and outside another.

None of this makes platform reporting useless. Platforms need conversion data to optimize their bidding, and their numbers are useful for comparing campaigns within that platform. They just shouldn't be summed together or used on their own to decide how much each platform deserves.

What attribution can't tell you

Every model above answers the same underlying question. Of the touchpoints we could see, which ones were present before the conversion, and how should we share the credit?

None of them answers whether the conversion would have happened without the ad. A retargeting campaign aimed at people who already have items in their cart will look excellent under almost any attribution model, because those people were close to buying. Some of them would have bought regardless.

That causal question needs a different tool. Incrementality tests, such as holdouts, geo tests and conversion lift studies, measure what changes when ads are removed for a comparable group. Marketing mix modeling estimates channel effects from aggregate patterns over time. Attribution is fast and detailed, and it's most reliable when its blind spots are checked against those slower methods.

How to choose a model for each decision

Instead of picking one model for everything, match the model to the decision.

For optimizing within a single platform, like choosing between ad sets or keywords, use the platform's own attribution, ideally data-driven if you have enough volume. The platform is optimizing on that data anyway, and the comparison is like for like.

For comparing channels against each other, use a single neutral source, usually your analytics tool or CRM, with one model applied to every channel. Last click in a neutral tool is crude, but it's consistent. A data-driven model in your analytics tool is better if it's set up properly. Either way, never compare one platform's self-reported numbers with another's.

For understanding what introduces people to you, look at first-touch reports and at the self-reported source on forms, if you ask for it. "How did you hear about us" is messy data, but it captures word of mouth, podcasts and other influences no tracking pixel sees.

For deciding how much a channel deserves in the overall budget, use incrementality tests and, if your spend supports it, an MMM. Attribution is the weakest tool for this decision, because it can't separate influence from coincidence.

For reporting to leadership, show one consistent source of truth for conversions and revenue, usually the CRM or store, and explain the methods behind any channel split. Changing models between reports to tell a better story destroys trust quickly.

Writing these choices down in a measurement plan prevents the same argument from being had every month. That's a large part of what measurement planning is for.

Get the tracking right first

Every model depends on the data underneath. If conversions fire twice, if consent handling drops a large share of events, or if offline sales never make it back to the platforms, the choice of model is the least of your problems.

Check that every key conversion is recorded once, with the right value, in analytics and in each platform. Send qualified leads or sales back from your CRM where the platforms support it. Where browser tracking is losing data, server-side tracking can recover some of it. Only then does comparing attribution models start to mean something.

Questions people ask

What are the main marketing attribution models?

The common ones are last click, first click, linear, time decay, position based and data-driven. The first five follow fixed rules for sharing credit across touchpoints. Data-driven models use conversion path data to estimate each touchpoint's contribution.

Which attribution model is right for my business?

It depends on the decision. Use platform attribution to optimize inside each platform, one neutral source with one model to compare channels, and incrementality tests or marketing mix modeling to decide how budget should be split.

Why do Google, Meta and other platforms report more conversions than my CRM?

Each platform credits itself whenever its ad was in the path, so the same sale can be claimed several times. View-through counting, modeled conversions and different attribution windows add to the gap. Compare platforms using a neutral source instead of adding their numbers together.

Is data-driven attribution accurate?

It's usually more nuanced than rule-based models, but it only learns from touchpoints the tool can see and it measures association. It doesn't establish cause. Treat it as a better way of sharing credit, and check its conclusions with incrementality tests where the budget justifies it.