Multi-Touch Attribution Models and the Data They Need

By the end of this page you will be able to tell which of your attribution models is lying to you, and why.

Every model is an assumption stacked on the same input: an ordered list of touches, each labelled with a source, a medium and a campaign. Split those labels across four spellings and the most sophisticated model available produces confident nonsense.

So the model debate is almost always premature. Audit the inputs first.

TL;DR

• Five touch-based models plus data-driven attribution. Each fails differently on dirty data.
• The platforms have been retiring models: Google removed four rules-based models from its advertising reporting in 2023, and HubSpot folded its position-based family into a single Empirical model.
• Source fragmentation starves data-driven models. Lost first touch breaks position-based models. Leaked last touch sinks last-non-direct.
• A clean last-non-direct model beats a dirty data-driven one on every decision a CMO actually makes.
• Audit thresholds in this piece are practitioner rules of thumb, not measured benchmarks.

Why This Still Matters

Plenty of people have declared multi-touch attribution dead. Cookie lifetimes are capped, click identifiers get stripped, and assistant traffic often arrives with no referrer at all.

And yet every team with a budget still has to decide where the next dollar goes. Incrementality testing answers that for your biggest channels. For the long tail, a newsletter sponsorship or a podcast read or a retargeting cohort, you are left assigning credit somehow.

The Five Models

Each is a rule for splitting one unit of credit across a journey.

First touch gives everything to the first known interaction. Honest about its bias, useful for asking which channels introduce people. It needs the first session tagged and stably recorded.

Last touch, usually last non-direct, gives everything to the final interaction, skipping back past Direct to the most recent identifiable source. It needs every link tagged, because untagged links dump credit into Direct and the walk-back only reaches the next clean touch.

Linear splits credit evenly. It is the most punitive model for missing touches, since every gap either shrinks the pool or hands credit to a channel that did nothing.

Position-based (U-shaped) weights the first and last touches at 40 percent each. Its sensitivity is the first touch, which frequently happens before any system has stable identity on the person.

W-shaped adds the lead-creation and opportunity-creation moments at 30 percent each. It needs lifecycle events carrying source attribution, which is marketing automation and CRM work rather than analytics work.

One caveat on all five: they survive as reference vocabulary more reliably than as product features. Google removed the rules-based models other than last click from its advertising reporting in 2023. HubSpot consolidated U-shaped, W-shaped and the J-shaped variants into a single Empirical model. Check what your platform actually offers before writing a model into a playbook.

Data-Driven Attribution

Instead of a fixed rule, data-driven attribution learns from your own conversion paths, giving more credit to touches whose presence predicts conversion. It runs a counterfactual: remove this touch, and how much does predicted conversion probability fall?

Three things follow for your tagging.

It reads the same input layer as everything else. Two spellings of one source become two channels with separately learned weights.

It needs volume. Google publishes thresholds for its Ads data-driven model, requiring roughly 300 conversions and 3,000 ad interactions within 30 days to qualify, with lower numbers to stay eligible, and falls back to another model when a campaign drops below them. GA4 does not publish an equivalent per-path figure. Either way, splitting campaign cardinality unnecessarily starves the model.

And it cannot fix bad inputs. If a quarter of your conversions land on Direct because transactional emails are untagged, the model dutifully learns that Direct predicts conversion.

Data-driven does not mean self-correcting. It means the rule is learned rather than fixed, and a learned rule from poisoned data looks more authoritative than a simple one precisely because nobody can inspect it.

What Each Model Demands

Model	Most demanding requirement	Failure when unmet
First touch	First session tagged and stably attributed	Credit lands on Direct or Referral and is never reassigned
Last non-direct	Every link tagged	Credit collapses into Direct; the walk-back reaches only the next clean touch
Linear	Every touch captured	Missing touches shrink the pool; mislabelled ones move credit to the wrong channel
Position-based	First and last both correct	40 percent lands on whatever the first recorded session happened to be
W-shaped	Lifecycle events carrying source	30 percent of credit lands on Unknown
Data-driven	Volume per consistent channel	Fragmentation starves the model and biases the weights

Every model has one point of high sensitivity, and every one of those points sits in the tagging layer.

Four Failure Modes

Numbers below are illustrative.

Source fragmentation. One channel arrives under several names. Linear splits its credit across rows nobody re-aggregates. Position-based lands its 40 percent on a spelling your weekly report filters out. Data-driven treats each spelling as a separate channel and under-credits all of them. The channels that cannot fragment, Direct and Organic Search, quietly look bigger.

Channel mislabelling. An email tagged utm_medium=social because the template came from a social post. Every model assigns the credit exactly where you told it to. Nothing flags it, the numbers add up, and the output is wrong. This is the one data-driven attribution cannot recover from, because it treats your labels as ground truth.

Lost first touch. Cookie caps, stripped click identifiers, a marketing subdomain that does not share storage with the conversion domain, or tracking that started after the buyer’s real first visit. First-touch and position-based models then credit whatever the platform managed to see first, often a mid-funnel retargeting channel. In a 90-day B2B cycle, your platform may only know the last month of it.

Note one wrinkle here. Platforms that store a first-touch value are supposed to keep it forever, and in practice it can be overwritten: HubSpot’s Original Source can be clobbered by imports, connector syncs and contact merges. If first-touch credit matters to you, copy the value into a property nothing else writes to. The HubSpot guide covers that in detail, and the Salesforce guide covers the equivalent at conversion.

Leaked last touch. The real final click carried no parameters, so Direct takes the credit. Linear inflates Direct because it appears at the end of nearly every path. Data-driven learns that Direct converts and shifts budget away from the channels that actually worked.

Audit Before You Choose

These thresholds are rules of thumb rather than benchmarks. Use them to decide where to look, not to decide whether you pass.

Channel distribution on converting sessions. Filter to your conversion and group by channel. A Direct share past roughly a fifth for B2B, or a third for consumer brands, suggests a last-touch leak. More than a couple of percent in Unassigned means values arriving that match no rule, though note GA4 now ships an AI Assistant channel, so assistant referrals no longer need a custom group to be classified.

Source and medium cardinality. Scan the top rows for the same source under several spellings. Each duplicate breaks linear and data-driven models.

Campaign cardinality. Count distinct campaign values. For a team running twenty campaigns, a couple of hundred historical values is reasonable and several thousand is not. The usual cause is dates, timestamps or ad IDs in the campaign field that belonged in content or term.

Trace five conversions by hand. Walk each path backward and ask whether each source makes sense, whether the medium is right, whether the campaign is interpretable, and what preceded any Direct touch. An hour of this surfaces every common failure.

Check whether it will stay clean. The four audits above tell you what arrived. They say nothing about next quarter. If marketers hand-type parameters into ad platform fields, you will run this same audit again with the same findings.

That last part is what Terminus, the marketing taxonomy governance platform, is for: controlled vocabularies for source and medium, validation on the campaign naming convention, and approval workflows on paid plans, so malformed values never reach the analytics platform at all.

The Pragmatic Stance

With a clean input layer, the model choice is a real tradeoff and reasonable people disagree. B2B teams often prefer a lifecycle-weighted model, ecommerce teams often prefer last non-direct or data-driven.

Without one, the call is easier than it looks: a clean last-non-direct model beats a dirty data-driven model.

It is interpretable. When someone asks why a channel lost credit this week, you can trace it to sessions and campaigns instead of gesturing at a model.

It fails loudly. Broken tagging shows up immediately as credit flowing to Direct, and somebody notices. Poisoned training data produces a chart that still looks fine.

And the gap between a good model and the best model is far smaller than the gap between clean and dirty inputs. Clean the inputs, pick a defensible model, and spend the remaining attention on measuring incrementality where the money is.

For the aggregate counterpart to all of this, see marketing taxonomy and MMM.

FAQ

What are the core attribution models?

First touch, last touch (usually last non-direct), linear, position-based and W-shaped, plus data-driven attribution, which learns the rule from your own paths rather than applying a fixed one.

Is GA4’s default last click or data-driven?

Data-driven for advertising conversions in new properties, while session-scoped acquisition reports still use a last-non-direct model. Two reports disagreeing is expected.

Why do all models fail on dirty data?

They read the same input: a sequence of touches labelled by source, medium and campaign. Each model has a different point of maximum sensitivity, but none of them can repair a label that was wrong on arrival.

What is leaked last touch?

The final click carried no parameters, so Direct gets the credit. Transactional email, chat links, QR codes and in-product calls to action are the usual culprits.

What is source fragmentation?

One channel arriving under several spellings, each treated as a separate source. Reports under-count the real total and data-driven models learn separate weights for each fragment.

Does first-touch attribution still work?

With reduced reliability. Cookie caps and stripped identifiers mean a client-side first-touch value can vanish before a long consideration cycle ends. Writing the value into a CRM record at form submission is far more durable.

Should B2B use position-based or W-shaped?

W-shaped maps better onto a B2B motion, but it needs lifecycle events carrying source attribution, and your platform may have retired the model. Check what is available before designing around it.

What is the highest-value investment in attribution accuracy?

The tagging layer: a controlled vocabulary enforced before links are created and a validated campaign naming convention. The model is a second-order decision.