Multi-Channel Attribution: A CMO's Guide to Models, MMM and Lift Tests

Customers rarely buy after a single ad. A B2B buyer might see a LinkedIn post, search for your company, read a comparison page and join a webinar before asking for a demo. A shopper might find a product in a social feed, compare prices on a search engine and buy after an email offer. When the budget review comes, you need to know which of those channels earned more money and which can be cut.

Multi-channel attribution helps by assigning the credit for each conversion to the tracked touchpoints that came before it. The credit depends on the model, though, and on its own it does not show how many extra sales a channel caused. The tools have changed too: Google Ads and Google Analytics retired four of the classic models in 2023.

This guide explains how attribution works, which models you can still use and where, and how it compares with marketing mix modeling and lift tests. It ends with a set-up plan you can follow and the published prices of a few tools.

DisclosureJake Jorgovan is a partner in Fieldtrip, a group of marketing agencies, which is mentioned in this article.

What is multi-channel attribution?

Multi-channel attribution assigns the credit for a conversion (a sale, a sign-up, a booked demo) to the marketing touchpoints a customer interacted with before it. Google's Analytics help (opens in a new tab) defines attribution as "the act of assigning credit for important user actions to different ads, clicks, and factors along the user's path." The attribution model is the rule, set of rules or data-driven algorithm that decides how that credit is split. When the credit is shared across several touches, this is also called multi-touch attribution.

The simplest approach gives all the credit to one touch, the first or the last. Google's own example shows the difference. A customer searches "restaurant tuscany", "restaurant florence", "3 star restaurant florence" and finally "3 star restaurant abigaille florence", clicking an ad each time, and then books a table. Last click gives the final keyword 100% of the credit. A data-driven model gives each keyword a share, depending on how much it contributed.

For a CMO the point is practical. You want to see which combination of channels leads to revenue, so the budget follows contribution and not just whichever channel happens to sit at the end of the path.

The attribution models, and where you can still use them

The table describes the common models as Google and HubSpot define them. The last column matters more than it used to.

Image

Zoom

Original hypothetical four-touch example based on attribution definitions in HubSpot and Google documentation. Credit assignment is not incremental sales lift. GA4 no longer offers all these rule-based models; platform availability varies.knowledge.hubspot.com (opens in a new tab) · support.google.com (opens in a new tab) · knowledge.hubspot.com (opens in a new tab)

What Google changed in 2023

In April 2023 Google announced (opens in a new tab) that the first click, linear, time decay and position-based models were going away across Google Ads and Google Analytics 4. By then, it said, fewer than 3% of Google Ads web conversions were attributed with them. Conversion actions still using them were switched to data-driven attribution. When this guide was checked in October 2026, Google Ads (opens in a new tab) offered last click and data-driven, and Google Analytics offered data-driven, paid and organic last click, and Google paid channels last click. Google Analytics says the four rule-based models have not been available since November 2023.

If an agency presents linear or U-shaped attribution as a built-in Google Ads or Google Analytics model, ask where that calculation comes from: those models were retired in 2023. The rule-based models are still useful ways to think about a journey, and B2B teams can still run them in a CRM. HubSpot's attribution reports (opens in a new tab), in Marketing Hub or Content Hub Professional and Enterprise, offer U-shaped, J-shaped, linear and other models. The W-shaped model works only in deal create and revenue attribution reports, and full path (22.5% each to the first interaction, lead creation, deal creation and last interaction, with 10% across the rest) only in revenue reports. Those two report types, the ones that tie touches to pipeline and closed revenue, are Marketing Hub Enterprise only. Dedicated tools also keep rule-based models: Ruler Analytics and Attribution both offer first and last touch, linear, time decay and position-based models, and Attribution's Custom plan adds custom weights for teams that want to set their own.

How data-driven attribution works

Google's data-driven attribution (opens in a new tab) compares the paths of customers who convert with the paths of those who don't, and gives more credit to the interactions that make a conversion more likely. Google Analytics adds that its model weighs factors such as the time before the conversion, the device, the number of ad interactions, their order and the type of creative. It estimates each touch's effect by comparing exposed users with similar users in a holdback group.

In Google Ads, every conversion action is eligible for data-driven attribution, whatever its volume. Google still recommends at least 200 conversions and 2,000 ad interactions in supported networks within 30 days for the model to work well, and notes that with little data, last click and data-driven can give the same results.

Two limits matter for budget decisions. Google Ads' data-driven model looks at interactions with Google's ads: Search (including Shopping), YouTube, Display and Demand Gen. Which channels can receive credit is set for each conversion action, so check it before you compare channels. Google Analytics can also credit organic channels for web conversions, but it gives direct visits no credit unless the whole path was direct. Either way, the model assigns credit within the recorded journey; it cannot credit a trade show or a sales conversation that was never recorded or imported.

Before you switch a conversion action from last click, Google recommends that you test the new model and check how it changes your return on investment. Its Model comparison report shows two models side by side, and the setting also changes what automated bid strategies such as Target CPA and Target ROAS optimize toward.

How the choice of model changes the budget

The model decides which channels look productive, and so where the budget goes. These four hypothetical businesses show what each model would fund and what it can miss.

A real case shows the stakes. Walks of Italy, a Rome-based tour company, used search ads to reach travelers researching tours. According to Google's case study (opens in a new tab), 34% of its conversion paths involved two or more touchpoints, and last-click brand searches accounted for 48% of its ticket sales. Its agency, Merkle | Periscopix, set up a data-driven model in Google Search Ads 360, which showed that generic searches such as "Rome tour" were being undervalued. The agency then used the model with automated bidding to bid more on those keywords. Against summer 2016, its data-driven bidding campaigns raised revenue 33% (against 6% for its non-brand campaigns) and ROI 25%, and ticket sales rose 69%. The model and the bidding change worked together; the lesson for a CMO is how much last click had been hiding.

Attribution, marketing mix modeling or lift tests?

Attribution is one of three ways to measure marketing's contribution. Each answers a different question, and the budget decision usually needs more than one.

Marketing mix modeling (MMM) works from aggregate data instead of individual journeys. Meta's guide to MMM (opens in a new tab), written for its open-source Robyn package, describes it as an econometric model that estimates the incremental effect of marketing and non-marketing activities (price, promotion, seasonality) on a KPI such as sales. Because it needs no user-level data, it can include TV, out-of-home and other offline media, and it does not depend on cookies or pixels. Google's open-source Meridian (opens in a new tab), free for anyone to use, is built to answer three questions: each channel's historical ROI and contribution, how its impact changes with spend, and how to allocate the next budget.

MMM asks a lot of the data. The Robyn guide says a model needs a minimum of two years of weekly data, more if you only have monthly figures, and about 10 observations for each variable. Spend has to vary: if TV ran at the same level every week, the model has difficulty telling what TV did. And analysts, it says, still play "a very important role" in tuning the model and interpreting it. A first study also takes time. The guide's typical timings add up to about 12 to 22 weeks: one to two weeks to define the questions, four to six to collect the data, one to two to review it, four to eight to model and two to four for analysis and recommendations.

Lift tests (also called incrementality tests) measure cause directly. You split an audience, or a set of regions, into a group that can see your ads and a group that can't, and compare the conversions. Google's Conversion Lift (opens in a new tab) reports incremental conversions, incremental cost per action and incremental return on ad spend; its regional version supports offline data, and it is available through a Google account representative, not to every account. Meta's lift tests (opens in a new tab) use randomized test and holdout groups and report a confidence percentage with the result.

The methods work best together. The Robyn guide strongly recommends calibrating a mix model with experiments and running them regularly, and Google's Meridian offers geo experiments (Meridian GeoX) for the same purpose. Ruler Analytics, which sells attribution and mix modeling, puts the trade-off plainly on its homepage: "MMM provides a holistic but less granular view and incrementality testing is useful for determining causation but not substitute for ongoing reporting."

Image

Zoom

Choose the measurement method by the decision it must support. Experiments can calibrate mix models; each method retains its own blind spots.

Which to use when

• Mostly digital journeys and enough conversions: start with data-driven attribution in Google Analytics or your ad platforms, and use a lift test to check your largest channel where a valid test is available.
• Significant offline media, or budget spread across many channels: add a mix model for the split between channels, calibrated with lift tests, and keep attribution for campaign decisions.
• B2B with long cycles and few large deals: use the CRM's attribution to see which touches lead to pipeline and revenue, and treat the result as one input.

That last point is where I'd be most careful. In a Leaders of B2B episode (opens in a new tab) in December 2024, talking about brand and messaging, I described B2B companies that close 20 deals a year, each worth half a million dollars. In deals like those, a few interactions no one can track (a piece of content, a customer's testimonial) can move the decision. As I put it then, "there's intuition plus data as a guiding light, but it can't be the full piece there." With so few won deals, use the data to guide the conversation, and talk to the sales team and the customers before you cut a channel.

The same caution applies to reading ROI. The Robyn guide notes that a channel with a low ROI but high contribution and spend can still be a big driver of results, so moving money away from it may be a mistake. When a channel looks weak, consider a lift test before you cut it: hold it back in a few regions or for part of the audience for a set period, and compare. Have an analyst confirm first that the holdout, the duration and your conversion volume can detect an effect of the size that matters, and treat an inconclusive result as uncertainty, not as proof the channel does nothing.

How to set up multi-channel attribution

These seven steps take you from the decision you need to support to a model you can check.

1. 1

   Write down the decisions the measurement must support

   Start with the questions: which channels to grow, which to cut, how to split next year's budget. Meta's guide suggests collecting these questions first and mapping each to the method that answers it, since some are better answered by a lift test or a mix model than by attribution. Name the KPIs too: pipeline, revenue, cost per acquisition.

2. 2

   Choose the conversions and the model

   Decide which actions count as conversions and which model credits them. In Google Analytics you also choose the channels that can receive credit and the lookback window: 30 days by default for first visits and first opens (or 7), and 90 days for other key events (or 30 or 60). In a B2B cycle longer than 90 days, the early touches fall outside that window, and the CRM is where they can still be tied to the deal.

3. 3

   Tag every channel the same way

   Use UTM parameters on the campaign links you control that bring people to your site, such as email, social and partner links, with one naming convention across teams and agencies, so the same campaign is not reported under three names. Check that conversion tags fire: when conversions go unrecorded, Google's help says to verify the tags on every relevant page, especially thank-you and confirmation pages.

4. 4

   Connect offline and sales touchpoints

   Bring in events, sales calls, meetings and offline conversions where the data exists. Ruler Analytics lists sales calls, trade shows and webinars among the offline touches it tracks, and Attribution offers offline channel tracking on its Custom plan. The CRM holds the deal stages that rule-based B2B models need.

5. 5

   Keep the data in one place

   Consolidate ad, analytics and CRM data in one platform: the CRM, a customer data platform or a data warehouse. In the MMA's 2024 attribution benchmark (opens in a new tab), 80% of marketers said they were dissatisfied with their ability to reconcile results from different tools.

6. 6

   Check the model before you move budget

   Compare models side by side. Before a material change to a channel's budget, check whether a lift test has enough volume, time and a sound holdout design to give a useful answer, and run it where it does.

7. 7

   Review on a fixed schedule

   Set a cadence, monthly or quarterly, and keep the settings stable between reviews. In Google Analytics, changing the reporting model also changes historical credit in reports that use event-scoped traffic dimensions such as source and campaign, while a new lookback window applies only going forward, so note every change in the report.

Image

Zoom

The approach described in this article, shown as a visual guide.

If you work with an agency, ask it to show the model and settings behind its reports before you compare its numbers with your own. The guide to performance marketing agencies compares firms that run paid acquisition.

What attribution can't see

Every attribution model works from tracked interactions, so its blind spots are the interactions no tool records.

• Offline touches. A store visit, a phone call or a conversation at an event is invisible unless you import it. If much of your spend is offline, a mix model or a regional lift test (Google's supports offline data) can account for it.
• People who decline tracking. On Apple devices running iOS 14.5 or later, Apple requires (opens in a new tab) apps to get the user's permission before tracking them across other companies' apps and websites, including for advertising measurement. Without it, the device's advertising identifier is all zeros.
• Cookies. Google decided in April 2025 to keep its current approach (opens in a new tab) of letting Chrome users choose whether to allow third-party cookies in their settings, without a new standalone choice screen. Chrome's Incognito mode blocks them by default. Measurement that depends on them misses the people who block them. Google Analytics lists cookieless measurement and modeling of behavior and conversions among its privacy controls.
• Several devices. A buyer who starts on a phone and finishes on a laptop can look like two people. Adobe's Cross-Device Analytics (opens in a new tab) stitches devices into one person using deterministic matching, but Adobe's documentation says it is available only to existing customers who already have a license and is no longer sold as an add-on.
• Views without clicks. An ad someone sees but doesn't click leaves no click path. Google Analytics' data-driven model counts some engaged YouTube views (30 seconds watched, for example), and some dedicated tools offer impression attribution, but a click-based model misses the rest.

What attribution tools cost

The costs below come from the tools' own pages. Prices change, so check the current page before you budget.

Both attribution vendors publish named customer results. In a case study published in 2026 (opens in a new tab), Ruler says the outsourced communications provider Moneypenny had been losing visibility on around 20% of the leads that came in by phone, until it tracked them back to their Google Ads source. In a 2025 case study (opens in a new tab), Attribution says the webinar platform Livestorm tripled its ROAS in a year. These are the vendors' own figures. With the two open-source mix models, the cost is the data preparation and the analysts' time, not a license.

Can attribution measure brand awareness?

Not directly. Attribution can connect the tracked interactions of an awareness campaign to later conversions, and some tools credit qualifying views without a click, but it does not measure whether people recognize or remember your brand. For that, use a brand lift study: Meta's lift tests can measure awareness by asking exposed and holdout groups questions about the brand. A mix model can estimate the contribution of brand media, offline included, to sales over time.

Where to start

Pick the one decision you have to make this year, such as whether to grow a channel or cut it, and measure for that decision first. Use attribution to manage campaigns, test where a valid test is possible before you move a large share of the budget, and add a mix model when your spend and history are big enough to support one. Which channel in your plan would you cut today, and what evidence would you want before you did?