Multi-Touch Attribution Models: Why They Need Clean UTMs to Work

Published by Terminus, the marketing taxonomy governance platform.

Multi-touch attribution models distribute conversion credit across the touchpoints a buyer encountered before they converted. There are five core touch-based models in standard use (first-touch, last-touch, linear, U-shaped or position-based, and W-shaped), plus a sixth class of data-driven attribution that uses machine learning to infer credit from observed conversion paths. Every one of these models is a mathematical assumption stacked on top of campaign-level data. If that data is dirty (split sources, missing UTMs, broken redirects, leaked direct traffic), the model output is noise, no matter how sophisticated the algorithm above it. This article walks through what each model assumes, what it needs from your tagging layer, and how each fails when the inputs are not clean.

TL;DR

There are five widely accepted touch-based models (first-touch, last-touch, linear, U-shaped, W-shaped) and a sixth class, data-driven attribution, which is GA4’s default for paid and organic conversions in newly created properties.

Every model reads from the same input layer: a sequence of touches identified by source, medium, and campaign. Garbage in, garbage out applies to all of them.

GA4 removed the four rules-based models other than last-click for paid and organic conversions in 2023, leaving last-click and data-driven attribution in the Advertising workspace.

Dirty data breaks specific models in specific ways: source fragmentation poisons linear and DDA, lost first-touch destroys U-shaped, leaked last-touch sinks last-non-direct, and channel mislabelling biases every model’s output.

A clean-tagged last-non-direct model beats a dirty data-driven model on almost every operational decision a CMO actually makes.

Audit the inputs (taxonomy, channel grouping, direct-traffic rate, campaign cardinality) before debating the model.

Why we still talk about multi-touch attribution in 2026

Half the analytics field has spent the last few years declaring multi-touch attribution dead. Apple’s Link Tracking Protection strips parameters from links in Mail and Messages. Safari’s ITP caps JavaScript-set first-party cookies at seven days, and at 24 hours when set after a cross-site tracker redirect. Third-party cookies in Chrome have been on a multi-year deprecation arc. And the AI-era surfaces (ChatGPT, Claude, Perplexity, Google’s AI Overviews) send referral traffic whose channel classification is genuinely ambiguous.

Despite all that, every team with a meaningful budget is still trying to decide where the next dollar goes. Some model of credit assignment is unavoidable. You can run deterministic incrementality tests on the biggest channels, but for the long tail (a newsletter sponsorship, a podcast read, a retargeting cohort, an LLM citation) you are stuck deciding how much credit each touchpoint deserves. Multi-touch attribution is the tool people still reach for.

The catch: the conversation about which model to use is almost always premature. Before “should we be on linear or U-shaped” matters, you have to know your touchpoint data is correctly labelled. That part determines whether any model output is trustworthy.

The five core models and what each assumes

Every touch-based attribution model is a rule for distributing one unit of conversion credit across the touches in a buyer’s journey. The five core models in common use are consistent across Google, Adobe, HubSpot, and most analytics textbooks. (Google deprecated the four rules-based models other than last-click inside GA4 in 2023 for paid and organic conversions, but the model definitions themselves remain standard reference vocabulary.)

First-touch

Assigns 100 percent of credit to the first known touchpoint.

Assumption: the first interaction is the one that mattered. Everything after was inevitable follow-through.

Useful for top-of-funnel diagnostics. Which channels introduce new prospects? It is honest about its bias (it ignores everything that happened after the first click), which is a virtue.

Needs from your tagging: a reliable record of the first known touchpoint. Most platforms (HubSpot Original Source, Salesforce custom first-touch fields, GA4’s first_user_* dimensions) capture this once and never overwrite it. If the first session is not properly tagged, credit falls into whatever default bucket the session lands in.

Last-touch (and last-non-direct)

Last-touch assigns 100 percent to the most recent touch before conversion. Last-non-direct is a refinement: if the very last touch is Direct (no referrer, no UTM), skip backward to the most recent non-direct touch. GA4’s default last-click model is a last-non-direct variant, as was Universal Analytics’ default.

Assumption: the most recent interaction closed the deal. Everything earlier was warm-up. Useful for bottom-of-funnel optimisation, especially high-intent search and retargeting.

Needs from your tagging: a clean last touch. Every paid click, every email link, every social post, every QR code must land on a URL carrying an accurate utm_source and utm_medium. Untagged links and tools that strip UTMs in transit dump credit into Direct and break the model. Last-non-direct papers over some of this, but only as far as the next clean touch.

Linear

Assigns equal credit to every touchpoint. A four-touch path gives 25 percent to each.

Assumption: every touch matters equally. Useful for long-cycle B2B journeys where you genuinely believe the early-funnel content and the late-funnel sales touch contributed roughly equally.

Needs from your tagging: every touch in the journey captured and correctly attributed. Linear is the most punitive model for missing touches. Every untagged session either shrinks the credit pool or shows up as Direct and pulls credit toward a channel that did nothing.

U-shaped (position-based)

Assigns 40 percent to the first touch, 40 percent to the last, and splits the remaining 20 percent across the middle touches.

Assumption: the introduction and the close are the two highest-value moments. The middle is supporting work. Useful when you have a strong opinion that lead generation and opportunity creation deserve the bulk of the credit.

Needs from your tagging: both the first touch AND the last touch correctly identified. U-shaped has the highest sensitivity to first-touch loss, because it concentrates 40 percent of credit on a moment that often happens before any conversion-tracking system has stable identity on the user.

W-shaped

A B2B variant of U-shaped. Assigns 30 percent to the first touch, 30 percent to the lead-creation touch (when the contact became an MQL), 30 percent to the opportunity-creation touch (when sales accepted the lead), and splits the remaining 10 percent across other touches.

Assumption: B2B journeys have three milestone events, not two. Each deserves outsized credit. Useful for B2B SaaS, enterprise sales, any motion structured around discrete lifecycle stages.

Needs from your tagging: lifecycle-event tagging, not just session-level UTM tagging. Every touch around the MQL conversion and the opportunity creation must carry correct source attribution. This is usually the job of marketing automation (HubSpot workflows, Marketo programs, Pardot Engagement Studio) and CRM (Salesforce custom UTM fields), not the analytics platform.

Data-driven attribution and where its signal comes from

The sixth model class is data-driven attribution (DDA). Instead of a fixed rule, DDA uses machine learning over your observed conversion paths to learn which touchpoints are most predictive of conversion. Touches that consistently appear in converting paths but rarely in non-converting paths get more credit. Touches that show up in both equally get less.

In GA4, DDA is the default model for the Advertising workspace in newly created properties. It operates at the event-scoped conversion level, so each conversion event can have its own model and attribution is recalculated as new path data arrives. In plain language, DDA runs a counterfactual: for each touch in a path, it asks how predicted conversion probability changes if that touch is removed. Touches whose removal causes a large drop get more credit.

What this means for your tagging:

DDA’s input is the touchpoint stream identified by source / medium / campaign. The algorithm is more sophisticated than first-touch, but it reads from the same input layer. If two Facebook campaigns are tagged inconsistently (utm_source=facebook vs utm_source=fb), DDA sees two channels, learns separate weights, and produces split credit that looks like noise.
DDA needs enough data to train. GA4 has historically required a minimum number of conversions per path over a 28-day window. For low-volume conversions, DDA falls back to a rules-based model behind the scenes. Splitting campaign cardinality unnecessarily starves DDA of the volume it needs.
DDA is trained on your own paths. It cannot fix bad inputs. If 28 percent of your conversions land on Direct because your email links are untagged, DDA learns to associate Direct with conversions and assigns credit accordingly. The model is doing what it was asked to do. The data lied to it.

“Data-driven” does not mean “self-correcting.” It means the credit distribution rule is learned from your data instead of being a fixed formula. If the data is wrong, the learned rule will be wrong, and the answer will look more authoritative than first-touch or last-touch precisely because the algorithm is opaque.

The clean-input requirement

Every model reads the same input layer: for each conversion, an ordered sequence of touches, each carrying enough metadata for the analytics platform to assign it to a channel. In GA4 the metadata is utm_source, utm_medium, utm_campaign, optional utm_term and utm_content, plus gclid and dclid for Google Ads / Display, and equivalent click IDs for other platforms. Here is what each model demands.

Model	Most demanding requirement	Failure when unmet
First-touch	First session correctly tagged and stably attributed	First-touch credit falls into Direct or Referral and is never reassigned
Last-touch (last-non-direct)	Last session correctly tagged	Untagged last touches collapse into Direct; last-non-direct reaches only as far as the next clean touch
Linear	Every touch in the path captured and labelled correctly	Missing touches shrink the credit pool; misattributed touches pull credit toward wrong channels
U-shaped	First touch AND last touch correct, with high confidence	40 percent of credit lands on whatever the first session was, even if Direct or a misclassified Referral
W-shaped	First, MQL, and opportunity touches all captured with lifecycle event tagging	Lifecycle events without source attribution dump 30 percent credit into Unknown or onto the user’s last session
Data-driven	Sufficient volume per consistent channel path, with a stable channel grouping	Source fragmentation splits channels into low-volume cohorts, starves the model, biases learned weights

The pattern is consistent. Every model has at least one point of high sensitivity to dirty data, and that point is somewhere in your tagging layer. The model is downstream of the taxonomy.

Concrete failure modes

Four patterns recur across teams. Each breaks specific models in specific ways. Numbers in the examples below are illustrative.

Source fragmentation

The same channel arrives under multiple names. utm_source values like facebook, Facebook, FB, fb, facebook.com, and meta all refer to the same source, but GA4 treats them as distinct rows.

Linear: a four-touch path with two Facebook touches and two LinkedIn touches should give Facebook 50 percent. With fragmentation, the report shows 25 percent each to facebook and FB, plus 50 percent to LinkedIn. The aggregate is technically correct but split across rows no marketer will mentally re-aggregate.
U-shaped: 40 percent of credit lands on the first touch’s source name. If the first touch is FB and your weekly report rolls up facebook only, the credit is invisible.
Data-driven: each variant is treated as a separate channel. Cohort volume falls below DDA’s threshold and the model under-credits the fragmented channel or falls back to a rule-based model.

The directional cost: the fragmented channel looks smaller than it is, and the channel that cannot be fragmented (Direct, Organic Search) looks larger.

Channel mislabelling

A campaign goes out with the wrong utm_medium. An email campaign tagged utm_medium=social because the link came from a template last edited for a social post. A paid LinkedIn ad tagged utm_medium=email because someone copy-pasted from an email link.

All five touch models: credit is assigned to the wrong channel. The model is doing what it was told. The report is wrong.
Data-driven: this is the failure DDA cannot recover from. DDA does not validate that the source / medium combination is sensible. It treats the field value as ground truth. Every mislabelled campaign is training data pointing the model the wrong direction.

Channel mislabelling is the silent killer. Nothing flags it. The numbers add up. The model runs cleanly. The output is wrong.

Lost first-touch

The buyer’s actual first touch is not the first touch in your record. Causes: a first visit on Safari with ITP capping cookie lifetime, a Mail.app link with stripped UTMs, a marketing subdomain that does not share cookies with the conversion domain, or conversion tracking implemented only after the buyer’s actual first touch.

First-touch: 100 percent of credit goes to whatever the platform considers the first touch, which is in fact the second or third actual touch.
U-shaped: 40 percent of credit is misassigned, often to a mid-funnel retargeting channel that was simply the first channel the platform managed to identify.
W-shaped: same problem, muted by the MQL and opportunity 30 percents. Still means a third of credit is on the wrong channel.

In a long-cycle B2B journey with a 90-day consideration window, the analytics platform often knows only the last 30 to 45 days of touches.

Leaked last-touch

The last touch before conversion is captured as Direct because the actual last link did not carry a UTM, or the UTM was stripped in transit. Common sources: untagged links in transactional emails, untagged chat-tool links (Slack, Teams), untagged QR codes, untagged in-product CTAs going out via push or in-app notification.

Last-touch: credit lands on Direct. Last-non-direct walks back to the previous non-direct touch, which may or may not be the right channel.
Linear: Direct picks up an outsized share because it appears at the end of nearly every converting path.
Data-driven: DDA learns that Direct has strong predictive power. Over time, the model shifts credit toward Direct, effectively under-crediting every paid channel.

If Direct is more than about 20 percent of converting sessions and you are not a household-brand consumer business, the leak is almost certainly in your last-touch tagging.

How to audit MTA inputs before trusting MTA outputs

Before debating which attribution model to adopt, run this audit. The point is not to be exhaustive: it is to surface whether your input layer is healthy enough that the model debate is worth having.

Step 1: channel distribution for converting sessions

In GA4, Acquisition > Traffic acquisition grouped by Default channel group, filtered to your conversion event. If Direct is over 20 percent for B2B SaaS or over 35 percent for consumer brands, you have a leaked last-touch problem. If Unassigned is over 2 percent, you have either a channel grouping that needs updating (likely for AI-era referrers) or campaigns going out with malformed UTM values.

Step 2: source / medium cardinality

Same report, ungrouped, broken down by Session source and Session medium. Sort by sessions descending and scan the top 50 rows. How many are duplicates of the same source under different spellings? google / Google, facebook / Facebook / fb, linkedin / LinkedIn / li, newsletter / Newsletter / NEWSLETTER. Each duplicate is fragmentation, and each duplicate breaks linear and DDA.

Step 3: campaign cardinality

Same report, broken down by Campaign. Count distinct values. For most mid-market teams running under 20 active campaigns, the count should be under 50 (including historical campaigns rotating out). If it is in the hundreds or thousands and you are not a programmatic display advertiser, you have campaign explosion. Common cause: campaign names including dates, timestamps, or ad IDs that should have been in utm_content or utm_term. Campaign explosion poisons DDA because each campaign-as-channel has too few conversions to train against.

Step 4: trace converting sessions backward

Pick five recent conversions. For each, walk the User explorer (or BigQuery export) backward. For each touch ask: does the source make sense for the time it happened, is the medium tagged correctly (social, email, cpc, organic, referral), is the campaign interpretable or a date string / internal ID / untitled-1, and if there’s a Direct touch, what was the touch immediately before? Five hand-walked paths surface every common failure mode within an hour.

Step 5: check the upstream taxonomy

The audit so far tells you whether the analytics platform is receiving clean data. It does not tell you whether the taxonomy is enforceable going forward. If every marketer is hand-typing UTMs into ad platform UI fields, the audit will produce the same failure modes next quarter no matter how much you clean up today.

This is the layer where Terminus, the marketing taxonomy governance platform, lives in the stack. The job of a governance tool at this layer is not to generate prettier UTMs. It is to constrain the input space (controlled vocabularies for source and medium, validation on campaign naming convention, audit of who created what) so the analytics platform never receives a malformed UTM in the first place. Once you have a clean input layer, the question of which attribution model to use becomes a real question with a useful answer.

The pragmatic stance

If you have a clean input layer, the choice between models becomes a thoughtful tradeoff. Most B2B teams end up on W-shaped or DDA. Most ecommerce teams end up on last-non-direct or DDA. Reasonable people disagree.

If you do not have a clean input layer, the pragmatic call is this: a clean-tagged last-non-direct model produces better operational decisions than a dirty data-driven model. Three reasons.

First, last-non-direct is interpretable. When a marketer asks “why did Facebook get less credit this week?” you can give an honest answer and trace credit to specific sessions, campaigns, and creative. When DDA shifts credit, almost nobody can explain why except in vague language about “the model learned something.” Interpretable is a feature when the stakes are budget decisions.

Second, last-non-direct fails loudly. When tagging breaks, the report immediately shows credit flowing to Direct, and someone notices. When DDA’s training data is poisoned, the report looks fine: numbers add up, the chart trends sensibly, the bias is buried in the learned weights. Loud failures get fixed. Quiet failures compound.

Third, the gap between best and good-enough attribution model is much smaller than the gap between clean and dirty inputs. The right sequence is: clean the inputs first, pick the model second, debate refinements third. The teams that get attribution right are not the ones who picked the most sophisticated model. They are the ones who got the input layer under control, picked a defensible model, and spent the rest of their attention on tracking real incrementality on the biggest channels.

FAQ

What are the five core multi-touch attribution models?

First-touch (all credit to the first interaction), last-touch (all credit to the last interaction, often refined to last-non-direct), linear (equal credit to every touch), U-shaped or position-based (40 percent to the first touch, 40 percent to the last, 20 percent to the middle), and W-shaped (30 percent each to first touch, lead-creation, and opportunity-creation; 10 percent to the rest). A sixth class, data-driven attribution, uses machine learning over observed paths instead of a fixed rule.

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

GA4’s default for newly created properties’ Advertising workspace conversions is cross-channel data-driven attribution. The reporting attribution model used by most acquisition reports remains a last-non-direct variant for session-scoped traffic. Google has been steadily shifting more reports onto DDA since 2023, when the rules-based models other than last-click were removed from the Advertising workspace for paid and organic conversions.

Why do all attribution models fail on dirty UTM data?

Every model reads from the same input: an ordered sequence of touches identified by source, medium, and campaign from the UTM tagging layer. If the inputs are wrong, the output is wrong, regardless of the sophistication of the model on top. Linear is sensitive to missing touches, U-shaped to lost first-touch, last-non-direct to leaked last-touch, and DDA to channel mislabelling and source fragmentation that splits its training cohorts.

What is leaked last-touch and how do I detect it?

Leaked last-touch is when the actual last channel before conversion appears as Direct because the link was not tagged, or the UTM was stripped in transit (by iOS Link Tracking Protection, by an ESP’s redirect, by a chat tool, or by a tracking-prevention browser). Detect it by checking the share of conversions attributed to Direct: above 20 percent for B2B SaaS, or above 35 percent for consumer brands, is a strong signal. The fix is upstream tagging of every link source: transactional email, chat, QR codes, in-product CTAs.

What is source fragmentation?

Source fragmentation is when the same channel arrives under multiple utm_source values: facebook, Facebook, FB, fb, facebook.com. Each spelling is treated as a distinct source. Reports under-count the actual aggregate. Data-driven attribution treats the fragments as separate channels, learns separate weights, and often produces nonsensical credit splits.

Does first-touch attribution still work with iOS Link Tracking Protection and ITP?

It works, but with reduced reliability. ITP caps JavaScript-set first-party cookies at seven days (and at 24 hours when set after a cross-site tracker redirect), so any first-touch attribution depending on a long-lived client-side cookie can lose the original value if consideration exceeds the cap. iOS Link Tracking Protection strips known click identifiers (gclid, fbclid, and others) from links in Mail, Messages, and Safari Private Browsing; UTMs generally survive but may be missing on traffic that relied on a click ID alone. Server-side first-touch capture (writing the value into a CRM record at form submission) is much more reliable than relying on a client-side cookie.

Should we use U-shaped or W-shaped attribution for B2B SaaS?

W-shaped is the more B2B-aligned model because it recognises the three discrete lifecycle moments (first touch, MQL conversion, opportunity creation) that structure a B2B sales motion. It requires lifecycle-event tagging in your marketing automation and CRM, not just session-level UTM tagging. If your team does not yet have reliable MQL and opportunity timestamps with source attribution attached, U-shaped is a reasonable starting point.

What is the single highest-ROI investment in attribution accuracy?

The taxonomy and tagging layer. Concretely: a controlled vocabulary for utm_source and utm_medium enforced before the link is created, a validated utm_campaign naming convention, and an audit log of who tagged what. The choice of attribution model is a second-order decision. The first-order decision is whether the touchpoint data feeding any model is correctly labelled and consistently structured.

Last updated: 2026-06-30.