Beyond last touch: what multi-touch attribution reveals about your real ROAS

Every growth team has one channel it argues about. The CPI looks expensive, someone proposes cutting it, and the people closest to that channel insist it does more work than the dashboard credits it for. That argument never resolves, because everyone in it is reading a report that only knows what happened last.

We opened our recent panel by asking where teams sit today with multi-touch attribution. 70% said they still rely primarily on last touch. A few minutes later, we asked the same room where last touch misses the most value, and the answer that came out on top was cross-channel assists, with discovery and view-through close behind. So 70% are running on a model they already suspect is misplacing credit.

Our panel was Eran Friedman, Co-founder and CTO at Singular, Sourabh Khandelwal, Manager of Marketing Science at Snapchat, Anthony Mkhize, Global Sr. Sales Engineering Director at Smadex, and Alexei Moltchan, VP of Product at Dataseat, hosted by our CMO Steph Pilon.

What is last-touch attribution, and why do most teams still run on it?

Last-touch attribution asks what the user engaged with immediately before converting, then hands that channel the full credit. The rule earned its place. It is fast, it is consistent, and it is the signal your media partners optimize against, which is why nobody on the panel argued for switching it off.

Don’t ignore last touch, because it’s what the industry has been relying on for so long. That’s what the networks are optimizing towards.

— Eran Friedman, Co-founder and CTO, Singular

Sourabh Khandelwal made the case for why it holds on so well, and inertia is only part of it. A model that is easy to explain is easy to get a room to agree on, which matters more in practice than most measurement debates admit. In his framing, measurement choices trade off between speed and robustness, and last touch is very good at speed. The cost shows up later, when the channel that closed gets read as the channel that worked.

Multi-touch attribution splits credit across the whole user journey

Multi-touch attribution, or MTA, evaluates the full sequence of touchpoints in a journey and splits credit across them using a weighting model you define. The weighting is yours to set. You can spread credit evenly, favor earlier exposures, or keep more weight near the conversion. Alexei Moltchan made the point that there is no single correct model, because the consideration time and journey shape differ by vertical, so the attribution model should reflect how your customers actually buy.

Eran used a shopping example most teams will recognize from their own funnel. A friend recommends a pair of Nike shoes, you see them on social, and that is where the product enters your head. Days later, you search the model to compare prices, and the purchase lands on search. Under last touch, search takes the entire conversion, and the logical conclusion from that report is to put everything into search. Do that for two quarters and the discovery layer feeding search disappears, along with the demand it was creating.

What you’re comparing	Last-touch attribution	Multi-touch attribution
Question it answers	Which channel closed the conversion?	Which channels contributed across the journey?
Credit	100% to the final touchpoint	Distributed across eligible touchpoints by your model
Best used for	Daily optimization, partner reconciliation, network signal	Budget allocation across channels
Cadence	Daily	Weekly or monthly
Main blind spot	Demand creation, view-through influence, cross-device journeys	Causality, and conversions with no observed touchpoint

The two do different jobs. Anthony Mkhize described last touch as the pulse of the business, the read you use to see what campaigns need attention today, with MTA layered on for the larger decisions about where channel budget goes.

CTV and upper-funnel channels pay the highest price under last touch

Any channel contributing early is structurally exposed, because the credit sits at the end of the journey while the influence sits at the start. That covers view-through interactions, upper-funnel social and video, and CTV above all.

Alexei described what makes CTV work and what makes it hard to measure as the same thing. A big screen in a living room leaves a stronger impression of a brand than almost any other format, and it almost never produces the click. The impression lands on the television, the conversion happens later on a phone, and the journey is cross-device before anyone makes a measurement decision about it. Anthony described that separation as something the industry cannot design away. The behavior is not going to change, so the measurement model has to absorb it.

Alexei added the second-order effect, which is that an undervalued channel is usually also amplifying the channels around it. A CTV campaign not only fails to get credit for its own conversions, but it also raises the performance of the search and social campaigns that do get credit, and cutting it shows up as a decline somewhere you were not watching.

The caveat the panel kept returning to came from Sourabh: a channel appearing early is likely to be undervalued, not proven to be. Timing, consistency, and data quality all matter before that earns a budget decision.

Three metrics reveal hidden channel value before you build any model

This is the part most teams skip, and it needs no setup. Before you choose a weighting model, three numbers already in your reporting will tell you whether your read on a channel is wrong.

Metric	What it counts	How to read it
Assists	Conversions a channel contributed to without taking the last touch	A high assist rate means the channel regularly sets up conversions that other channels then close
Overlap, or co-attribution	How often does a channel appear alongside others in the same journey	High overlap suggests the channel helps capture demand, and can also flag quality issues worth checking
Single-attributed conversions	Conversions where a channel was the only touchpoint	This is unique reach: users that no other channel brought you

Read together, they change how a CPI looks. A channel with an expensive last-click CPI may be carrying a heavy assist load. A channel with an attractive CPI may overlap so heavily with everything else that little of its volume is genuinely additional.

That second pattern is also a fraud signal. Eran’s example was a long-tail affiliate that shows a surprisingly high overlap rate with your major channels. Sometimes that is a genuine mid-funnel contributor. Sometimes it is a partner firing clicks in bulk to be standing there when a conversion arrives that another channel drove, which is what click spamming looks like in the data.

Two Singular analyses show how much these numbers move the picture. Our analysis of Snapchat found a 51% assist rate, so for roughly every ten installs Snapchat received credit for, it contributed to five more it did not. On Meta, across trillions of ad impressions, billions of clicks, and billions of installs, the shape was different: a 94% single-attributed install rate and a 6% co-attributed rate across mobile gaming advertisers, generating up to 29% additional assisted installs in several gaming datasets, and up to 50% higher ROAS than the same campaigns showed under last touch. Across incremental channels generally, Singular customers see 43% higher ROI on average once contribution is measured across the journey.

Snapchat’s value showed up in assists, Meta’s in reach that no other channel duplicated. In both, the last-touch view was describing something smaller than what the channel was doing.

You do not need to be running at that scale to see it. Steph put it plainly on the call: looking at Singular’s own marketing data, the gap between what gets last touch and what the MTA view shows was startling enough that she now makes her budget decisions on MTA and no longer on last touch.

What should you do when MTA changes a channel’s CPI or ROAS?

Expect the numbers to move in favor of upper-funnel channels. CTV CPIs typically fall and ROAS rises once credit is distributed across the journey, and the instinct that follows is to move money immediately. Anthony’s advice is to treat it as a prompt to investigate instead. Find which channels the one in question commonly assists, check whether exposed users show stronger downstream behavior on retention or revenue, and confirm the pattern holds across geos and time periods rather than in a single slice.

He also put a fair challenge to anyone adopting MTA halfway. If you believe the model gives you a more complete view of the journey, you have to be willing to back what it tells you. Believing in a better measurement model while continuing to allocate every dollar on last touch is a position that does not survive contact with a budget meeting.

Sometimes backing it means doing nothing to the budget at all. Anthony’s point was that reinterpreting performance counts as a legitimate outcome: you stop judging a high-assist channel purely on whether it happened to win the last touch, and you leave the spend where it is.

When you do move, Alexei puts a number on the first step: shift roughly 10% to 20% of the budget toward the channel MTA favors, then watch total conversions rather than that channel’s own metrics. If conversions rise, the model is telling you something real. If they do not, look at how the channel is being optimized, and then at whether the weights in your model reflect the journey you think you have.

Our third poll found that this is where the difficulty concentrates. Asked what makes it hardest to act on a different MTA performance story, the most common answer was validating that the change drives incremental growth, ahead of knowing which signals to trust and separating contribution from exposure.

That is incrementality work, and the panel’s cleanest formulation of how the two fit together:

MTA can reallocate the pie, but it cannot make it bigger.

— Sourabh Khandelwal, Manager of Marketing Science, Snapchat

MTA redistributes credit for conversions you already recorded. Incrementality testing estimates whether a channel produced outcomes that would not have happened anyway. Sourabh described the loop: use MTA to find the question worth asking, design a test around the decision that matters, then use the result to recalibrate how you read the model. Run it more than once before you trust the number.

A 30-day MTA rollout you can run alongside last touch

1. Keep last touch as your benchmark. Singular preserves the attributed winner and exposes the other contributing touchpoints alongside it, so both views stay on the same screen.
2. Clean the data foundation: attribution settings, impression and click tracking, partner integrations, attribution windows, campaign taxonomy. As Anthony put it, a messy foundation gives you a more sophisticated version of messy.
3. Read the metrics you already have. Assists, overlap and single-attributed conversions are available before you build anything.
4. Find the strongest assist patterns, meaning which channels repeatedly appear together and in what order.
5. Pick one hypothesis and test it with a controlled budget shift, then validate with incrementality before scaling.

What stalls teams is usually the decision that comes before all of this, which is choosing a weighting model. Eran’s recommendation was to defer it: look at the metrics, run a default weighted model, and revisit the weights once you can see what your journeys look like. Singular customers get both out of the box, which removes the setup work that usually stops a first attempt.

The bottom line

Last touch tells you who closed the deal. Multi-touch attribution tells you who made the deal possible.

— Omri Gal, Head of Data, Singular

70% of the teams in our audience are still making budget decisions on the first of those two. You do not need to change your attribution model to find out what the second one would tell you. Put assists, overlap, and single-attributed conversions next to your last-click CPI for one reporting cycle and see whether the story holds.

Watch the full webinar

The panel went further on CTV measurement, attribution windows, choosing an MTA model, and what to do when MTA, last touch, and the ad platform each tell you something different. You can watch the session on demand here. If you already registered, the recording link is in your inbox.

Already running campaigns with Singular? See how attribution works across web, app, and CTV in one platform.

Frequently asked questions

What is multi-touch attribution?

Multi-touch attribution measures each touchpoint in a user’s journey and assigns fractional credit to the channels involved, instead of giving all credit to the final interaction. The advertiser defines the weighting, which can spread credit evenly, favor earlier touchpoints, or favor touchpoints closer to conversion.

How is MTA different from last-touch attribution?

Last-touch attribution awards 100% of the credit to the last channel a user engaged with before converting, while MTA distributes that credit across the eligible touchpoints in the journey. Last touch suits daily optimization and partner reconciliation. MTA suits decisions about how budget is distributed across channels.

Do I have to stop using last-touch attribution to use MTA?

No. Singular preserves your last-touch attributed winner and exposes the other contributing touchpoints alongside it, so MTA metrics sit next to the numbers your team already reports on.

What is an assist rate?

An assist rate measures how often a channel contributed to a conversion without receiving the last touch. Singular’s analysis of Snapchat found a 51% assist rate, meaning roughly five additional installs credited elsewhere for every ten Snapchat was credited with. Assists can be counted across multiple channels for the same conversion, so they do not sum to your total conversions.

Why is CTV usually undervalued by last-touch attribution?

CTV exposure happens on a television while the conversion usually happens on a mobile device, so the journey is cross-device, and the later mobile interaction receives the credit. That reflects the measurement model rather than the contribution of the channel.

Does MTA replace incrementality testing?

No. MTA redistributes credit for conversions that already occurred, while incrementality testing estimates whether a channel generated outcomes that would not have happened otherwise. Use MTA to identify which channel to question, and incrementality to validate the decision before a significant budget change.