Multi-touch attribution: how to use it correctly in 2026

December 5, 2021

Tin Le

Updated on October 19, 2026: Adtek rewrote this article from scratch with the latest data, platform features and regulations from 2025 to 2026.

Multi-touch attribution (MTA) splits the credit for a sale, an app install or a lead across every touchpoint the customer went through on the way: a Facebook ad, a Google search, an email, a KOL video, a sales call. Last click hands 100% of the credit to the final click. Multi-touch attribution answers a different question: how much did each channel contribute?

In 2021, the main question was whether to pick a linear, U-shaped or W-shaped model. The situation in 2026 is different. Apple has tightened user-level tracking on iOS, GA4 dropped four rule-based models in November 2023, and marketing mix modeling (MMM) and incrementality testing have become far easier to access. MTA is still useful, but it is now one of three measurement tools, not the only one.

What is multi-touch attribution?

Picture someone buying an air purifier. Last week they saw a TikTok video, then clicked a Facebook ad twice, read a comparison post on a blog, and finally typed the brand name into Google and placed the order. A first click model gives the entire order to TikTok. A last click model gives it to branded search. Both are partly right and mostly wrong.

MTA treats the whole path as the unit of analysis. Each touchpoint receives a share of the credit, either by a fixed rule or by an algorithm that learns from the data, and the shares always add up to one conversion. For MTA to work at all, you have to be able to tie the touchpoints back to the same person. That is exactly where privacy changes hit hardest.

Multi-touch attribution models and where they stand in 2026

Model	How credit is split	Good for	In GA4 and Google Ads
Last click	100% to the final touchpoint	Billing partners	Still available (Paid and organic last click, Google paid channels last click)
First click	100% to the first touchpoint	Seeing which channels bring in new customers	Removed
Linear	Split evenly across all touchpoints	Listing which channels took part	Removed
Time decay	The closer a touchpoint is to the conversion, the more credit it gets	Short promotional campaigns	Removed
U-shaped (position-based)	40% to the first touch, 40% to the last, 20% shared across the middle	Short purchase cycles	Removed
W-shaped	30% to each milestone: first touch, lead creation, opportunity creation; the remaining 10% goes to the middle	B2B, real estate, education	Not built in, has to be set up in your CRM
Full path	22.5% to each milestone, adding a deal-closed milestone; 10% to the middle	High-value deals	Set up in your CRM
Data-driven	An algorithm learns from paths that did and did not convert	Accounts with plenty of conversions	Default

Rule-based models are easy to understand, but people set the percentages. No data proves that the first touch is worth 40% rather than 25%. A Markov chain, the algorithmic approach teams most often use when they build their own model, measures how much the conversion rate drops if you remove one channel from every path. It is more objective, but it needs a lot of user-level path data.

W-shaped and full path models are still worth using in slow-selling industries. In a digital marketing strategy for industrial park real estate, milestones such as lead creation, site visits and signing a memorandum of understanding all sit in the CRM with clear dates, so splitting credit by milestone is practical.

Why old-style MTA has become harder to use since 2021

iOS: ATT, SKAdNetwork and AdAttributionKit

Apple launched App Tracking Transparency (ATT) in April 2021. Apps now have to ask users for permission before tracking them across other companies’ apps and websites. According to Adjust’s Mobile App Trends 2026 report (published February 2026), the average opt-in rate in Q1 2026 was 38%, up from 35% in the same period of 2025, counted among users who actually saw the permission prompt. Put another way, most iOS users are still missing from user-level path data.

According to StatCounter, iOS accounted for 40.19% of mobile web traffic in Vietnam in September 2026. An MTA model that misses most iPhone users can end up skewed toward channels that perform well on Android.

To measure app advertising without identifiers, Apple offers SKAdNetwork and its newer framework, AdAttributionKit, and the two are interoperable. According to Apple’s WWDC 2025 session (June 2025), AdAttributionKit launched with iOS 17.4 and added re-engagement measurement from iOS 18. Starting with iOS 18.4, an app can have several re-engagement windows running in parallel. Advertisers can set their own attribution windows per ad network and per interaction type (view or click), set a “cooldown” so that a re-engagement right after an install does not steal the credit, and receive the country code in postbacks once the anonymity threshold is met.

Even so, the data still arrives as aggregated, delayed postbacks with no list of touchpoints per person, so it cannot be fed straight into a traditional MTA model. Mobile measurement partners (MMPs) have to combine data from users who opted in to ATT with these postbacks and model the rest.

GA4 now only offers data-driven and last click

In April 2023, Google announced that it would retire first click, linear, time decay and position-based models in both Google Ads and GA4, since fewer than 3% of web conversions in Google Ads still used those four models. According to the GA4 help page, since November 2023 GA4 offers only three options: data-driven, Paid and organic last click, and Google paid channels last click.

GA4’s data-driven model learns from paths that did and did not convert, estimates how much the probability of converting changes when a touchpoint is added, and assigns credit accordingly. The limit is the input. GA4 can only give credit to clicks it records (plus engaged views on YouTube). Facebook ad views without a click, KOL videos, billboards and hotline calls are all absent. Direct traffic is excluded from credit unless the entire path is direct.

In Vietnam, many orders happen on marketplaces

Metric data cited by Đại biểu Nhân dân newspaper (January 2026) puts 2025 sales on the four marketplaces Shopee, TikTok Shop, Lazada and Tiki at about VND 429,700 billion, up 34.75%. TikTok Shop grew its market share from 29% to 41.31%. When a customer sees an ad on Facebook, searches on Google and then buys on Shopee, the website pixel never sees the order. Each platform also claims credit under its own attribution window, so adding up the conversions reported by Meta, Google and TikTok usually produces a number larger than the real order count.

On the legal side, Vietnam’s Personal Data Protection Law No. 91/2025/QH15 took effect on January 1, 2026. It gives users the right to consent, withdraw consent and request deletion of their data. Matching phone numbers and emails in your CRM with advertising data to rebuild customer paths now needs a fresh review of the legal basis for consent. For more on first-party data, see our article on marketing analytics in a cookieless world.

Multi-touch attribution, MMM and incrementality testing answer three different questions

Method	Question it answers	Data needed	How often	Blind spot
MTA, data-driven attribution	Within the paths we can measure, which touchpoints are linked to conversions?	User-level paths	Weekly	Channels it cannot measure; it shows correlation, not causation
Marketing mix modeling (MMM)	How much does each channel contribute to total sales?	Weekly spend and sales, plus pricing, promotions and seasonality	Quarterly	Needs a long data series and spend that actually varies
Incrementality testing	If we turn this channel on or off, how much do sales change?	A test group and a control group (by person or by region)	Per test round	One question at a time, and it takes time

Two changes have made this combination realistic for mid-sized businesses. On January 29, 2025, Google opened Meridian, its open-source MMM, to every marketer, along with a program of more than 20 certified measurement partners. Meridian uses Bayesian causal inference, measures video by reach and frequency, and accepts incrementality test results as priors to calibrate the model.

The second change is the cost of testing. According to Think with Google (February 2026), in 2025 Google Ads lowered the minimum budget for an incrementality test from a previous high of USD 100,000 to USD 5,000. Treat small-budget tests as directional signals. Our article on incrementality testing covers how to design control groups.

Here is how the three tools fit together. Incrementality tests produce causal numbers for a few large channels and are used to calibrate the MMM. The MMM decides how budget is split each quarter. MTA handles weekly optimization inside each channel, at the level of keywords and creatives.

Which multi-touch attribution model should a Vietnamese business choose?

Adtek’s view: choose based on how much you spend and where your orders actually happen.

Small spend, one or two main channels

Take a shop that mainly runs Facebook ads and sells on Shopee. It does not need to build MTA yet. Optimize within each channel using the platform reports, but judge overall performance by total revenue divided by total marketing cost, week by week. To find out whether a channel really brings in extra orders, pause it in a few provinces and compare. It is crude, but more honest than adding up the numbers from three ad managers.

A website or app with many conversions and several Google channels

Keep GA4 and Google Ads on data-driven. Once a month, open the model comparison report. A channel that gains noticeably more credit when you switch from last click to data-driven is one that last click has been undervaluing. For apps, use an MMP, design a conversion value schema for SKAdNetwork and AdAttributionKit, and set the attribution windows and cooldown to match how customers really use the app.

B2B, real estate, education, automotive

Sales cycles are long and many milestones happen offline. A W-shaped or full path multi-touch attribution model built in the CRM is a sensible choice, as long as the sales team logs every milestone (lead, meeting, site visit, deposit) and the lead source is captured right at the form or call. The hardest part here is usually data entry discipline, not the formula.

Large spend, including TV, out-of-home and KOLs

This is the point to invest in MMM. Meridian’s code is free, but you need someone who works with data. Every quarter, run at least one incrementality test on your biggest channel or on a channel suspected of “riding on” organic sales, such as branded search and retargeting, then use the result to calibrate the model.

Checklist before using attribution data to split your budget

Write down the decision you need to make: splitting budget across channels, choosing creatives, or proving that a channel generates additional sales.
Agree on one source of real revenue (CRM, ERP or marketplace reports), and do not add up conversions from the platforms.
Standardize UTMs and campaign naming, shared by the agency and the in-house team.
Review consent collection on your website, app and lead forms under the Personal Data Protection Law.
GA4: correct key events, test transactions filtered out, Google Ads linked, data-driven selected.
App: when the ATT prompt appears, the conversion value schema, the attribution windows.
One incrementality test per quarter, starting with the most expensive channel.
Move to MMM once you have three or more large paid channels, any offline channels, or attribution reports that contradict each other so much that nobody trusts them.

Two questions belong in every budget review: “If we turned this channel off for two weeks, how many orders would we lose?” and “When did we last verify this channel against a control group?” If the answer to the second is never, your multi-touch attribution model is getting more trust than it has earned.

If your business has attribution reports that disagree across Meta, Google, TikTok and real revenue, talk to the Adtek team about reviewing your measurement setup and planning tests.

This article was rewritten and updated by Adtek in 2026, based on original content from AppsFlyer (2021).