Which Attribution Model Actually Fits Your Business
To define attribution marketing is easy. Choosing the right model is not. A decision guide that matches attribution models to data volume, sales cycle, and channel mix.
On this page
Search "define attribution marketing" and you get six near-identical definitions. Adobe calls it "the process of determining which interactions influence a customer to purchase from your brand." Marketing Evolution calls it the analytical science of determining which tactics are contributing to sales or conversions. Fine. Accurate. Useless on Monday morning.
Because the definition was never your real question. Your real question is: which model do I actually run? Last-click, linear, position-based, data-driven, or a full handoff to media mix modeling? That choice depends on three variables, and only three: your conversion volume, your sales cycle length, and your channel mix. Everything else is decoration.
So here is the chooser. Walk the branches, take the verdict, move on.
First, the thirty-second definition
Marketing attribution is the identification of the user actions, Wikipedia calls them "events" or "touchpoints", that contribute to a desired outcome, and the assignment of value to each. A model is just the rule for splitting that value. Last-click gives 100% to the final touch. Linear splits evenly. Data-driven lets an algorithm weight touches based on observed conversion paths.
If you want the full taxonomy with the limits of each model, we wrote that up in what marketing attribution is and what works now. This piece assumes you know the vocabulary and need the decision.
Branch 1: How many conversions do you generate per month?
This is the gating question, because algorithmic models are data-hungry and rules-based models are not.
If you're under roughly 300 conversions per month (per market, per conversion type), stop considering data-driven attribution. The algorithm needs enough distinct conversion paths to learn stable weights; below this range it will fit noise and reshuffle credit every week. You'll watch a channel's "contribution" swing 40% with no change in spend, and someone will make a budget decision off it. Bad outcome.
Verdict: run a simple rules-based model (last non-direct click, or position-based if you have real upper-funnel spend) and validate the big calls with a quarterly geo-holdout test. Holdouts measure lift directly, so they work at any volume.
If you're between roughly 300 and 3,000 conversions per month, data-driven attribution becomes viable inside a single ecosystem, Google Ads DDA being the obvious example, but cross-channel algorithmic models will still be shaky at the edges. Verdict: platform-level data-driven for bidding, rules-based for cross-channel reporting, and don't let the two get confused in the same dashboard.
If you're above 3,000 conversions per month, volume is no longer your constraint. Move to Branch 2.
Branch 2: How long is your sales cycle?
Sales cycle length determines how much the order of touches matters, and whether your lookback windows can even see the whole journey.
If your cycle is under two weeks, most journeys are short, and the difference between last-click and multi-touch is real but modest. Amazon Ads frames attribution as determining how tactics "and subsequent customer interactions" contributed to conversions; when there are only two or three interactions, the models mostly agree. Verdict: data-driven if you cleared Branch 1, last non-direct click if you didn't. Spend your energy on creative and offer testing instead.
If your cycle runs two weeks to three months, this is the zone where model choice genuinely moves budget. Last-click will systematically over-credit branded search and retargeting, because those are the touches that happen to sit nearest the conversion. Verdict: multi-touch, data-driven where volume allows, with lookback windows set to at least 1.5x your median cycle. If your windows are shorter than your cycle, your model is truncating journeys and you're defining attribution on a fiction.
If your cycle exceeds three months, typical for considered B2B, touch-based attribution starts to fray regardless of model. Cookies expire, devices change, committees form. Individual paths become unreliable evidence. Verdict: keep a rules-based model for directional channel hygiene, but promote incrementality testing and pipeline-stage analysis to primary measurement. The comparison of multi-touch attribution and media mix modeling covers why long cycles push you toward aggregate methods.
Branch 3: How much of your spend can attribution actually see?
Here's the branch most definitional content skips entirely. Attribution, as Adjust puts it, identifies "which marketing touchpoints or channels led a user" to act. Note the assumption baked in: the touchpoint must be observable at the user level. TV isn't. Podcast ads aren't. A large share of iOS traffic isn't. Retail and wholesale revenue definitely isn't.
If more than ~90% of your spend is trackable digital (search, social, affiliate, email), touch-based attribution can carry the load. Verdict: whatever model Branches 1 and 2 produced, run it with confidence, and pressure-test setup issues with an audit tool like the Attribution Doctor.
If 10 to 25% of spend or revenue is untrackable, attribution still works but it's lying to you at the margins. Every dollar of untracked influence gets silently reassigned to trackable channels, usually branded search, which then looks artificially efficient. Verdict: keep your attribution model, but discount the trackable-channel numbers and start building toward MMM so the blind spots get valued.
If more than 25% of spend or revenue lives outside user-level tracking, no touch-based model, however clever, describes your business. This is the MMM handoff point. Media mix modeling works from aggregate spend and outcome data, so it doesn't care whether a channel drops a cookie. Verdict: MMM as the budget-allocation layer, attribution demoted to in-platform optimization, and a media mix calculator to sanity-check the reallocation before you commit it.
The chooser, on one table
| Your situation | Model verdict | Validate with |
|---|---|---|
| <300 conversions/mo, any cycle | Rules-based (last non-direct or position-based) | Quarterly geo-holdout tests |
| 300–3,000 conversions/mo, cycle <2 weeks | Platform data-driven for bidding; rules-based for reporting | Spot incrementality checks |
| 3,000+ conversions/mo, cycle 2 weeks–3 months, 90%+ trackable | Cross-channel data-driven attribution | Annual holdout on top channel |
| Cycle >3 months (B2B) | Rules-based for hygiene; pipeline analysis primary | Cohort and stage-conversion review |
| 10–25% of spend untrackable | Data-driven, with discounted branded/retargeting credit | MMM build in progress |
| >25% of spend untrackable (TV, iOS, retail) | MMM for allocation; attribution for in-platform only | Geo-holdouts to calibrate MMM |
The honest answer: you will end up with more than one
Every branch above ends in a single verdict because single verdicts are actionable. But notice what the validation column keeps saying: test it. That's because no model, rules-based, algorithmic, or econometric, is ground truth. Attribution infers. MMM estimates. Only a controlled experiment measures.
Mature measurement stacks therefore run three layers on a schedule: attribution for daily optimization, MMM for quarterly allocation, and incrementality tests to calibrate both. We've laid out how to sequence that without drowning the team in a measurement cadence that survives a CFO review. The chooser above tells you where to start; the cadence tells you where you'll end up.
One last framing. Definitions of attribution describe what happened. A model choice decides what you'll fund next quarter. Treat it with the seriousness that implies: revisit the branches whenever conversion volume doubles, a new channel crosses 15% of spend, or your sales cycle materially shifts. The right model two years ago is frequently the wrong model today.