Multi-touch attribution promised to show the exact worth of every ad, email and push. It rests on one condition: the ability to follow the same person from touch to touch. Browsers and phones have made that harder for ad networks. A loyalty program still meets the condition on its own channels, because members sign in.
What is multi-touch attribution?
Multi-touch attribution, or MTA, splits the credit for each sale across all the tracked touches that came before it. It works from paths at the level of the person. A path can read: saw a display ad, clicked a search ad, opened an email, bought. Single-touch models give all the credit to the first or last step. MTA gives each step a share.
The share comes from a rule or an algorithm. Rule-based MTA uses a fixed split, such as equal credit or 40% to the first and last touch. Algorithmic MTA learns the split from data, including paths that never ended in a sale (Google Analytics Help).
The two terms overlap. An attribution model is the rule that splits the credit. Multi-touch attribution is the whole practice around it: tracking each person’s touches, joining their devices and reporting credit by channel. A team can run MTA with a rule-based model or a data-driven one.
How does multi-touch attribution work?
First, collect every touch with a time stamp and a person ID. Then stitch the IDs together, so a phone visit and a laptop visit count as one person. Next, line up the touches before each sale inside a lookback window. Last, apply the model and add up each channel’s credit.
Two algorithms are common. Google’s earlier data-driven model applied the Shapley value from game theory, a way to split a team’s output fairly among its members. It compared the odds of a sale for similar users with and without each touch (Google Analytics Help, archived). Markov models treat the journey as a chain of steps and learn how each step leads to the next. The open-source ChannelAttribution package uses a Markov model to find these links in journey data (CRAN).
For illustration, a program sees four kinds of member paths. Members with no email and no push buy at 2%. Email alone gives 5% and push alone 4%. Members who get both buy at 8%. The Shapley value averages each channel’s gain over both orders. Email adds 3 points before push and 4 points after it, for 3.5. Push adds 2 and 3, for 2.5. The two shares sum to the 6-point gain, so email earns 58% of the credit.
What are the limits of multi-touch attribution?
MTA reads patterns in tracking data, and patterns are not proof. People who see more ads are often people who were already going to buy. Gordon and colleagues tested this with 15 US ad experiments at Facebook, covering 1.6 billion ad impressions. They found that observational methods often fail to match the randomized results, even with rich data on each user (Gordon et al.). In the illustration above, if the members who got both messages were the most active to begin with, part of that 8% came from who they were.
Click-based credit also flatters channels that catch buyers already on their way. In eBay’s field tests, brand keyword ads showed no measurable short-term benefit. Because clicks and purchases go together, true returns were a fraction of the usual estimates (Blake, Nosko and Tadelis). An MTA model that reads those clicks gives brand search a large share of credit it did not earn.
MTA also sees only what it can track. TV, radio, print and most store visits leave no click. A model fed only digital paths hands all the credit to digital channels.
| Method | Data it uses | Question it answers | Main weakness |
|---|---|---|---|
| Multi-touch attribution | Each person’s tracked touches and sales | How should credit for each sale be split? | Shows patterns, not cause; misses what it cannot track |
| Marketing mix modeling | Weekly totals of sales, spend, price and season | What did each channel add to total sales? | Needs years of data; little detail below channel level |
| Incrementality test | A random treated group and a random holdout | What did this campaign cause? | One campaign at a time; needs a holdout |
Is multi-touch attribution dead without third-party cookies?
Cross-site tracking has shrunk, but it has not ended everywhere. Safari has blocked third-party cookies by default since March 2020 (WebKit). On iOS 14.5 and later, an app needs the user’s permission to track them across other companies’ apps and websites. Without it, the device’s ad ID reads as all zeros (Apple).
Chrome took a different path. In July 2024, Google said it would offer users a choice in place of removing third-party cookies (Privacy Sandbox, 2024). In April 2025, it kept its current approach and dropped plans for a new standalone prompt (Privacy Sandbox, 2025). Chrome’s Incognito mode already blocks third-party cookies by default (Privacy Sandbox, 2025). So the gaps vary by browser and device. Any ad network’s view of a full journey has holes, and the holes are uneven.
How do loyalty programs use multi-touch attribution?
A loyalty program owns the identity that ad networks are losing. A member signs in to the app, enters a number at checkout or scans a card. Each of those ties a touch or a sale to one person in the program’s own first-party data. Email opens, push taps, offer views and store purchases can all join one path.
Apple’s rule covers linking app data with other companies’ apps, websites or offline properties for ads or ad measurement (Apple). Paths built from a program’s own app, site and stores sit on its own data. Retail media extends the same link to brands. Kroger’s ad arm matches ad exposure to real household purchases in stores and online. To find the true effect, it compares groups that saw the ad with groups that did not (Kroger Precision Marketing).
Log the touches that matter to the program: offer views, email opens, push taps, app sessions and card scans. Log offers that were seen and not taken, since those are the paths that did not convert. Set the lookback window to the member’s buying cycle, which for a grocery member is days and for a hotel member is months.
The member path still shows patterns and leaves cause unproven. Use MTA to split credit among program messages day to day. Then check the big calls with a random holdout, as our incrementality entry explains. A channel that earns credit in MTA but shows no lift in a test is reaching members who would have bought anyway. A standing holdout of members who get no promotional messages gives every MTA readout a baseline.
How GRAVTY handles multi-touch attribution
MTA in a program needs one member ID on every touch. GRAVTY’s Member 360 pulls each member’s behaviors and sales into one view, household included. Real-time events keep that record current as members earn and redeem. On lji.io, AgenticTest measures incremental lift by segment while a campaign is still running. That gives the credit split a check on cause.
Frequently asked questions
What is the difference between multi-touch and single-touch attribution?
Single-touch attribution gives all the credit for a sale to one touch, either the first or the last. Multi-touch attribution splits it across every tracked touch on the path. Google Analytics dropped its rule-based multi-touch models in November 2023. It now offers a data-driven model and last click models (Google Analytics Help).
Is multi-touch attribution accurate?
It is accurate about who saw what, and weak on what caused the sale. In 15 ad experiments at Facebook, observational methods often failed to match the randomized results (Gordon et al.). Treat MTA credit as a hypothesis and test the largest channels with a holdout.
Does Apple’s App Tracking Transparency affect multi-touch attribution?
Yes, for paths that cross other companies’ apps and sites. On iOS 14.5 and later, an app must get permission to track a user across other companies’ apps and websites. Without it, the device’s ad ID returns all zeros (Apple). Paths inside a brand’s own app and stores rest on its own data.
Is Chrome getting rid of third-party cookies?
No. In April 2025, Google said it would keep its current approach to third-party cookie choice in Chrome. It will not roll out a new standalone prompt. Incognito mode already blocks third-party cookies by default (Privacy Sandbox, 2025). Safari has blocked them by default since 2020 (WebKit).
What data does multi-touch attribution need?
It needs every touch and every sale tied to a person, with a time stamp. It also needs the paths that did not end in a sale. Google’s data-driven model learns from both converting and non-converting paths (Google Analytics Help). Without the losing paths, a model cannot tell which touches made a difference.
