TL;DR
Meta has had no native tool for comparing multi-touch attribution models since it retired Facebook Attribution in August 2021. Cross-channel budget decisions now rest on incrementality experiments (Conversion Lift, GeoLift) and marketing mix modeling such as Meta's open-source Robyn, calibrated with those experiments. Ads Manager stays useful, but read it as a direction, not as the truth about what your ads caused.
Quick answer: what happened to multi-touch attribution in Meta ads?
Multi-touch attribution (MTA) splits credit for a sale across the ad touchpoints a person saw before buying. Meta used to offer this natively through Facebook Attribution, a standalone tool with last-click, linear, positional and data-driven models. That tool was retired in August 2021. Today Meta offers three attribution models in Ads Manager (standard, incremental and custom) plus randomized experiments. Custom attribution can bring in a third-party tool's own multi-touch logic, but no native Meta model splits credit across touchpoints the way Facebook Attribution did.
This guide is for media buyers, DTC founders and marketing leads who still get asked "which channel really drove this?" and want an answer that holds up.
Does Meta still have a multi-touch attribution tool?
No. Meta has no native tool that compares multi-touch attribution models. What Meta offers now answers narrower questions.
Facebook Attribution let advertisers compare last-click, linear, positional and data-driven models and see cross-channel and cross-device conversion paths. Meta retired the Facebook Attribution tool's interface and API in August 2021, and the Facebook Attribution line of business was deprecated across Meta Business Tools. Contemporary accounts link the move to industry developments, including mobile operating system and browser policy changes; Meta's original Help Center notice is no longer publicly accessible (Portent). Meta's current measurement guidance points advertisers to Ads Manager reporting and to Experiments such as Conversion Lift (Meta Business Help Center).
Here is what remains, and why none of it is native MTA:
| Meta feature | What it does | Is it native multi-touch attribution? |
|---|---|---|
| Standard attribution (attribution settings) | Credits conversions inside click, engage-through and view windows chosen per ad set | No. Window-based credit, not touchpoint weighting |
| Compare attribution settings | Shows results across attribution models (standard, incremental, custom) and settings side by side | No. Compares Meta's own counting methods; it does not split credit along a path |
| Incremental Attribution | A Meta-modeled estimate of conversions the ads likely caused | No. A model of incrementality, not a path model |
| Custom attribution | Lets a supported partner or in-house system send its own attribution to Meta, which then optimizes to it | No. The multi-touch logic lives in the outside tool, not in Meta |
| Conversion Lift | Randomized user-level holdout experiment | No. Causal experiment |
| GeoLift | Open-source geo-level experiment method from Meta | No. Causal experiment |
| Robyn | Open-source marketing mix model from Meta Marketing Science | No. Aggregate time-series model |
Custom attribution is the closest Meta now gets to MTA, and the credit logic stays outside Meta. Meta's Help Center lists Adobe, Northbeam, Triple Whale and Rockerbox as supported partners, or an in-house system that sends AppendAttribution events through the Conversions API. It requires a Sales campaign optimizing for purchases with an auto-bid strategy, Meta says it needs about two weeks of high-quality signal before it is ready, and the feature may not be available to every account yet (Meta Business Help Center). Meta also warns against comparing ad sets with different attribution models in the campaign overview table and points to Compare attribution settings instead (Meta Business Help Center).
For how the windows themselves work, see what an attribution window is. For why Meta and GA4 report different numbers for the same purchase, see our pillar guide to Meta ads attribution. This article covers a different question: what to use for budget decisions now that Meta's native MTA tool is gone.
Drowning in Meta Ads?
Let Nova run your campaigns, approval-first.
Read moreWhy did cross-channel multi-touch attribution get harder?
User-level MTA needs to follow one person across ads, sites and devices. Privacy changes since 2021 removed much of that thread. The models did not get worse. The data they depend on got thinner.
An MTA model stitches touchpoints into a path ("saw a Reel Monday, clicked a Google ad Wednesday, bought Friday") and then splits credit along it. Apple's App Tracking Transparency, introduced with iOS 14.5 in April 2021, requires apps to ask permission before tracking people across other companies' apps and websites (Apple Developer). That made cross-publisher paths less complete, especially when journeys span apps, browsers and devices. A path model with missing links tends to over-credit the touchpoints it can see and under-credit the rest.
The industry moved away from manually chosen rule-based models rather than from multi-touch credit itself. In 2023 Google retired first-click, linear, time-decay and position-based attribution in Google Ads and GA4, noting that fewer than 3% of Google Ads web conversions used them (Google Ads Help). Google did not abandon multi-touch credit: GA4 and Google Ads still offer data-driven attribution, which distributes credit based on each account's conversion data (Google Analytics Help). The broader shift is toward modeled attribution, experiments and MMM.
Who is telling you MTA is dead?
Much of the most visible content questioning MTA and platform-reported numbers comes from measurement vendors, so weigh incentives alongside methodology. Two groups are worth separating. Third-party MTA vendors such as Northbeam, Triple Whale and Rockerbox sell attribution, often alongside MMM and testing, and three of them are also Meta's supported Custom attribution partners. MMM and incrementality vendors such as Recast, Haus, Measured, Lifesight and Prescient AI sell the alternatives. Many of their critiques are reasonable, and each group has a commercial interest in the answer. Meta and Google, which publish Robyn and Meridian, benefit when advertisers measure inside their tools too.
The defensible claim is narrower: user-level, cross-channel MTA became much harder after 2021, and experiments plus calibrated MMM became the more credible basis for cross-channel budget decisions. Third-party MTA still has a role, especially for brands with strong first-party data, which we compare in AdAdvisor vs Northbeam.
What replaced multi-touch attribution for Meta ads? The Question-to-Method Map
Nothing replaced MTA one for one. Four methods replaced it, and each answers a different budget question. Most wrong budget decisions come from asking one method a question it was not built to answer.
We call this the Question-to-Method Map:
| Budget question | Method | Evidence type | What it needs | Main limitation |
|---|---|---|---|---|
| What happened inside Meta this week? | Ads Manager attribution settings | Platform-reported, window-based | Pixel and Conversions API data | Directional only; counts are not proof of causation |
| What does Meta estimate its ads caused? | Incremental Attribution | Meta-modeled estimate | An eligible account | A model, not an experiment; worth validating with lift tests |
| Did this Meta campaign cause extra sales? | Conversion Lift | Randomized user-level holdout | Meta's guide: a campaign with $5,000+ spend and 500+ conversions | Measures Meta only, inside Meta's environment |
| Did Meta spend in these markets cause extra sales? | GeoLift | Geo-level experiment | Enough comparable markets and a power analysis | Usually less statistical power than user-level tests |
| Where should the next dollar go across channels? | MMM (Robyn or Google Meridian), calibrated with experiments | Aggregate statistical model | Two years of weekly data recommended | Slow to build; uncalibrated models can drift far from causal results |
The rule behind the map: answer each budget question with the method built for it, and use the causal methods to calibrate the slower ones.
Our view
A team with no MMM but two well-designed lift tests a year will likely make better Meta budget decisions than a team with an uncalibrated MMM and no tests. A model that has never been checked against an experiment can look precise and still point budget in the wrong direction.
Ads Manager attribution: the operational layer
Ads Manager attribution is the right tool for daily and weekly operations: catching a broken pixel, comparing two ad sets, spotting creative fatigue. It is the wrong tool for deciding whether Meta deserves more budget than Google, because it only sees Meta's touchpoints and credits any conversion that falls inside the window. Compare periods only when they use the same attribution setting.
Incremental Attribution: a modeled estimate, not a proven one
Incremental Attribution uses machine learning to estimate which conversions the ads caused rather than which ones followed an ad interaction. Meta first announced it in 2024 and widened access during 2025 (Social Media Today). Incremental results count only the conversions Meta considers incremental, so a campaign with 100 standard conversions might show 70 under incremental attribution, and these results are not available for date ranges before April 1, 2025 (Meta Business Help Center). A large gap between standard and incremental numbers on retargeting, for example, is a good reason to test retargeting first.
It is still a model, and Meta's own measurement page points advertisers to Conversion Lift to validate strategies built on it (Meta for Business). We cover the difference in AI incrementality testing for Meta ads.
Conversion Lift and GeoLift: the causal calibration layer
A sufficiently powered randomized experiment is the closest practical causal benchmark available to Meta advertisers. Conversion Lift withholds ads from a random control group and measures the difference in conversions. GeoLift does the same at the level of regions or cities.
As a guide, Meta says the ad account needs a campaign started in the past year with $5,000 or more in spend and at least 500 conversions, prorated for campaigns longer than 90 days, plus a signal-quality requirement such as Conversions API with an event match quality score above 5 (Meta Business Help Center). For well-powered tests, Meta recommends campaigns spending at least $10,000, a 10% holdout and more than 500 total conversions during the test (Meta Business Help Center).
GeoLift is often a better fit when the decision is geographic, such as whether Meta spend in particular markets drove incremental sales, because it reads business outcomes by market rather than conversions Meta observed. It does not by itself separate the effects of other channels. Its documentation notes that Meta recommends people-based experiments where possible because they have higher statistical power than geo-based ones, so run a power analysis before committing to a geo test (GeoLift documentation). For test design and sizing, follow AI incrementality testing for Meta ads.
MMM with Robyn or Meridian: the planning layer
Marketing mix modeling estimates each channel's contribution from aggregate spend and sales over time, so it does not depend on user-level tracking. That is why it survived the privacy changes that weakened MTA.
Robyn is Meta Marketing Science's open-source, semi-automated MMM package, and Robyn MMM is a common starting point for Meta-heavy brands that build their own model. Its documentation says "an MMM will need a minimum of two years of historical weekly data" for robust results, suggests more than two years if only monthly data exists, and recommends about 10 observations per independent variable (Robyn documentation).
Treat two years of weekly data as the standard recommendation, not a universal cutoff. Robyn says daily data can be acceptable with extra validation, at the cost of more noise, and that a model on six recent months needs daily data because weekly would be too sparse. A brand with a short or recently changed history can try MMM, but it should validate the model harder and lean more on experiments.
Google's Meridian is the other main open-source option, generally available since January 29, 2025. Google suggests at least two years of weekly data for geo-level models and three years for national-level models, and treats two years as a starting point rather than a guarantee: its worked example calls two years of national weekly data (about four data points per parameter) too low to estimate reliably, and suggests more data, fewer parameters or a geo-level model (Google; Google for Developers).
The most citable line in Robyn's documentation is about calibration: "We strongly recommend using experimental and causal results that are considered to be the ground truth to calibrate MMM." The docs name Meta Conversion Lift and GeoLift for this, accept point estimates as calibration input, and advise against using studies whose confidence is not high enough (Robyn documentation). Robyn's feature page cites a study by the consultancy Analytic Edge in which calibrating MMMs with Meta lift experiments changed Meta's estimated ROI by 25% on average and changed it in two out of three models; the public summary does not give a sample size (Robyn features; Analytic Edge). Treat that as a reason to calibrate, not as a universal error rate.
MMM vs incrementality testing for Meta ads: which should you use?
Use both, in a fixed order: experiments first, MMM second, and experiments again to keep the model honest. If you can only do one, the choice depends mostly on your data history.
| Your situation | Start with | Why |
|---|---|---|
| Under two years of clean weekly data | Conversion Lift or GeoLift | MMM likely needs extra validation and may not be reliable yet |
| Two years or more, one or two channels | Lift tests on your biggest channel, then a simple MMM | Experiments give the model its anchor |
| Two years or more, five or more channels | Calibrated MMM, refreshed with recurring experiments | Only MMM can compare many channels at once |
| One specific question ("is retargeting incremental?") | A single lift test | MMM is too coarse for one ad set decision |
An experiment answers one question with high confidence at one point in time. MMM answers many questions with lower confidence across time. If your brand has strong repeat purchase, read results against lifetime value too; see optimizing Meta to LTV, not first-order ROAS.
How do you avoid wrong budget decisions without multi-touch attribution?
The most common mistake is adding up platform-reported conversions and treating the total as business results. Meta, Google and TikTok can each claim the same sale, so the summed platform total can exceed actual orders. Compare it against backend revenue and blended ROAS before moving budget.
Four rules tend to prevent the expensive errors:
- Never move budget on one platform's CPA alone. A rising Meta CPA may reflect a reporting change, a tracking gap or a real decline. Check Conversions API health and backend orders first.
- Compare like for like. The same attribution setting, the same revenue definition and periods of equal length. If any of these differ, the comparison likely tells you about the measurement, not the ads.
- Mark 2026 as a reporting break. On January 12, 2026 Meta stopped returning 7-day and 28-day view windows from its Ads Insights API (Meta for Developers). In March 2026 Meta narrowed click-through attribution to link clicks and moved other interactions into engage-through attribution (Meta for Business; Search Engine Land). Use backend revenue or another consistently defined series for year-over-year analysis, and do not assume Meta can recalculate older periods under the new definitions. The full list is in Meta ads updates 2026.
- Let an experiment break ties. When Ads Manager, GA4 and your MMM disagree on a large budget decision, run a lift test on that specific decision instead of arguing about which dashboard is right.
Where does an AI agent fit after multi-touch attribution?
An AI agent can run the operational layer of the map; it does not replace experiments or MMM. Monitoring delivery, flagging Conversions API outages, catching period comparisons that mix attribution settings, and moving budget between ad sets within guardrails you set are well suited to automation. Deciding whether a channel is incremental is not. That still takes a randomized test or a calibrated model.
AdAdvisor brings more than 8 years in media buying, over $60M in managed ad spend, and an ex-Meta engineer on the team. Its Nova AI media buyer works at the operational layer: it proposes Meta changes for your approval and uses your break-even ROAS as a guardrail, so a lift-adjusted target from your own experiments can inform what it recommends. It does not run Conversion Lift studies or build MMMs.
FAQ
Summary
Meta's native multi-touch attribution ended when it retired Facebook Attribution in August 2021, and Meta has not built a replacement. The Question-to-Method Map replaces one tool with four: Ads Manager for operations, Incremental Attribution as a modeled screen, Conversion Lift and GeoLift as the causal calibration layer, and an experiment-calibrated MMM such as Robyn for cross-channel planning. Use each for the question it was built for, treat 2026 as a reporting break, and let an experiment settle any large budget argument.
Sources
- Meta Business Help Center: Test and learn
- Meta Business Help Center: About Conversion Lift
- Meta Business Help Center: Best practices to set up a conversion lift test
- Meta for Business: Conversion Lift and Incremental Attribution
- Meta Business Help Center: About attribution models and attribution settings
- Meta Business Help Center: Set up custom attribution
- Meta Business Help Center: View results for incremental attribution
- Meta for Business: Simplifying ad measurement for a social-first world (March 2026)
- Meta for Developers: Ads Insights API metric availability updates (October 16, 2025)
- Meta Marketing Science: Robyn analyst's guide to MMM
- Meta Marketing Science: Robyn features and calibration
- Meta Open Source: GeoLift documentation
- Apple Developer: App Tracking Transparency
- Google Ads Help: Retirement of rule-based attribution models
- Google Analytics Help: Attribution models in GA4
- Google: Meridian is now available to everyone (January 2025)
- Google for Developers: How much data Meridian needs
- Search Engine Land: Meta introduces click and engage-through attribution updates (March 2026)
- Analytic Edge: The value of calibrating MMM with lift experiments (2022, vendor study)
- Social Media Today: Meta updates information on Incremental Attribution (2025)
- Portent: Facebook attribution settings, what's changed (account of the August 2021 retirement)
Related reading
Analytics & Reporting
Meta Ads Attribution: Windows, Why Meta and GA4 Never Agree, and Which Number to Trust
Facebook attribution assigns credit by click and view windows, not last touch. Here's why Meta and GA4 always disagree, and which number to use for each decision.
Read moreWhat is an attribution window?
Read moreWhat is blended ROAS?
Read more
Analytics & Reporting
AI Incrementality Testing for Meta Ads: How to Measure True Lift (2026)
Incrementality testing shows which Meta ad conversions your ads actually caused. Learn how to run Conversion Lift and geo tests, and turn lift into a break-even ROAS rule.
Read more
Performance Optimization
Optimize Meta Ads to LTV, Not First-Order ROAS
Most Meta ads accounts are still judged against a single first-order ROAS number. For any brand with real repeat purchase or subscription revenue, that number can be too conservative. Here is the calculation that fixes it.
Read more
