# Why Did GA4 Remove 4 Attribution Models? (And What DDA Needs to Replace Them) GA4 removed 4 attribution models in 2023, leaving only Data-Driven Attribution (which needs 400+ conversions to run). Here's what changed and what to do. - Canonical: https://mbuzz.co/articles/ga4-attribution-models-removed - Published: 2026-02-05 - Last updated: 2026-05-25 - Author: Holly Mehakovic, mbuzz (https://mbuzz.co) --- > **TL;DR:** In late 2023, Google removed first-touch, linear, time-decay, and position-based attribution from GA4—leaving only last-click and data-driven. Google's reasoning: rule-based models are 'outdated' and data-driven (DDA) is superior. The reality is more nuanced. DDA uses Shapley values (game theory) and requires significant conversion volume to work well. If you need the removed models, you'll need a third-party tool or build your own in BigQuery. ## What Google Removed—And What's Left In late 2023, Google quietly removed four attribution models from GA4: | Removed Models | What They Did | |---------------|---------------| | **First-touch** | 100% credit to first interaction | | **Linear** | Equal credit to all touchpoints | | **Time-decay** | More credit to recent touchpoints | | **Position-based** | 40% first, 40% last, 20% middle | What remains: | Remaining Models | How They Work | |-----------------|---------------| | **Last-click** | 100% credit to final click before conversion | | **Data-driven (DDA)** | Algorithmic credit using Shapley values | This wasn't a subtle change. Google eliminated the models that marketing teams relied on for neutral baseline measurement (linear), demand gen attribution (first-touch), and e-commerce optimization (time-decay). ### Google's Official Reasoning Google's explanation was straightforward: rule-based models are "outdated" and data-driven attribution provides more accurate, property-specific insights. The subtext: Google wants everyone using their machine learning-based approach. > **Warning:** > **The practical problem:** Data-driven attribution requires significant conversion volume. If you're under 400 monthly conversions, DDA falls back to last-click—which means you effectively only have last-click available. ## How Google's Data-Driven Attribution Actually Works Google's DDA isn't a black box of pure machine learning. It's built on a specific algorithm with a Nobel Prize lineage. ### The Shapley Value Foundation GA4's data-driven attribution uses **Shapley values**—a concept from cooperative game theory developed by economist Lloyd Shapley (Nobel Prize, 2012). The core idea: How do you fairly distribute the output of a team among its members? In attribution terms: - The "team" = marketing touchpoints in a journey - The "output" = the conversion - Shapley values = fair credit distribution

How Shapley Values Work

For each touchpoint, calculate its marginal contribution:

  1. 1. Consider every possible subset of touchpoints
  2. 2. For each subset, measure conversion rate with the touchpoint
  3. 3. For each subset, measure conversion rate without the touchpoint
  4. 4. The touchpoint's "marginal contribution" = the difference
  5. 5. Average across all possible subsets = Shapley value

This ensures "fair" credit distribution based on actual contribution — the same math used to split profits among team members in cooperative game theory.

### What Makes It Different from Markov Chains You might hear "data-driven attribution" and think of Markov chain models—but Google doesn't use Markov chains for GA4 DDA. | Aspect | Shapley Values (Google) | Markov Chains | |--------|------------------------|---------------| | **Core concept** | Marginal contribution across all permutations | State transition probabilities | | **How credit is calculated** | Average contribution when added to any subset | "Removal effect"—what happens if channel is gone | | **Data requirements** | Higher—needs many conversions | Lower—can work with less data | | **Computational cost** | O(2^n)—expensive | O(n)—much faster | | **Best for** | High-volume properties | Lower volume, keyword-level analysis | **Why this matters:** If you have fewer conversions, Markov chains might give you better data-driven insights than Shapley values. But Google chose Shapley—so if you need Markov-based attribution, you'll need to build it yourself or use a third-party tool. ### GA4's Implementation Details GA4's DDA adds some specific features on top of Shapley: 1. **Time decay component:** Recent interactions get more weight 2. **50-interaction window:** Considers up to 50 touchpoints (vs 4 in Universal Analytics) 3. **90-day lookback:** Attribution window extends 90 days back 4. **Property-specific:** Models are trained on YOUR data, not global averages ```ruby # Simplified conceptual model of Shapley calculation class ShapleyAttribution def calculate_credit(journey) touchpoints = journey.touchpoints touchpoints.map do |touchpoint| # For each touchpoint, calculate marginal contribution # across all possible subsets subsets = all_possible_subsets(touchpoints) contributions = subsets.map do |subset| with_touchpoint = conversion_probability(subset + [touchpoint]) without_touchpoint = conversion_probability(subset) with_touchpoint - without_touchpoint end { channel: touchpoint.channel, credit: contributions.sum / contributions.size # Average marginal contribution } end end end ``` ## Why the Removal Matters ### The Volume Threshold Problem GA4's DDA requires minimum thresholds to function: | Requirement | Threshold | |-------------|-----------| | **Minimum conversions** | 400 per conversion type | | **Minimum path length** | 2+ touchpoints | | **Time period** | 28 days | | **Optimal performance** | 15,000+ clicks, 600+ conversions | If you don't meet these thresholds, **GA4 falls back to last-click**. For many businesses, this means: - Startups and SMBs → effectively stuck on last-click - New products → no data-driven insights until volume builds - Multiple conversion types → each needs its own volume ### What You Lose Without the Removed Models | Model | Use Case You Lose | |-------|-------------------| | **First-touch** | Demand gen reporting, pipeline sourcing, awareness measurement | | **Linear** | Neutral baseline, long consideration journeys, unbiased comparison | | **Time-decay** | E-commerce optimization, recency-weighted credit | | **Position-based** | B2B lead gen, balanced first/last crediting | If you used linear attribution as your neutral baseline—and many sophisticated teams did—you now have no simple way to replicate it in GA4. ### The Google Ecosystem Lock-In Concern There's an elephant in the room: **Google's DDA is trained on Google's data, measured in Google's tool, and may favor Google channels.** While Google states DDA is unbiased, the algorithm is a black box. You can't: - Inspect the model weights - Verify the Shapley calculations - Check for channel bias - Customize the algorithm The Dropbox IEEE Access paper (Chivukula et al., 2026) measured the gap between click-based attribution — DDA included — and causal reality directly: click attribution overstates causal impact by 2 to 10x compared to geo-incrementality experiments. HockeyStack publicly exited the attribution category in Q1 2026 because they concluded a single hidden model wasn't a defensible answer. Avinash Kaushik — formerly Google's Digital Marketing Evangelist, now strategy advisor to Measured — has been clear about the same gap: MTA (including DDA) is the "smart" rung of measurement; incrementality is the "super smart" rung. DDA doesn't get you to the top rung. It gets you off the bottom one. > **Get Back the Models GA4 Removed.** mbuzz supports first-touch, linear, time-decay, position-based, Markov, and Shapley. Free up to 30K records/month. [Start Free](https://mbuzz.co/signup) ## Your Alternatives ### Option 1: Build It Yourself in BigQuery GA4 exports raw event data to BigQuery. You can rebuild any attribution model from scratch. **First-touch in BigQuery:** ```sql -- First-touch attribution from GA4 BigQuery export WITH journey_first_touch AS ( SELECT user_pseudo_id, traffic_source.source AS first_source, traffic_source.medium AS first_medium, traffic_source.name AS first_campaign, MIN(event_timestamp) AS first_touch_time FROM `your_project.analytics_12345678.events_*` WHERE traffic_source.source IS NOT NULL GROUP BY user_pseudo_id, traffic_source.source, traffic_source.medium, traffic_source.name ) SELECT j.first_source, j.first_medium, COUNT(DISTINCT c.user_pseudo_id) AS conversions, SUM(c.conversion_value) AS revenue FROM journey_first_touch j JOIN conversions c ON j.user_pseudo_id = c.user_pseudo_id GROUP BY j.first_source, j.first_medium ORDER BY conversions DESC; ``` **Linear attribution in BigQuery:** ```sql -- Linear attribution: equal credit to all touchpoints WITH touchpoints AS ( SELECT user_pseudo_id, traffic_source.source, traffic_source.medium, event_timestamp, ROW_NUMBER() OVER (PARTITION BY user_pseudo_id ORDER BY event_timestamp) AS touch_order FROM `your_project.analytics_12345678.events_*` WHERE traffic_source.source IS NOT NULL ), journey_length AS ( SELECT user_pseudo_id, MAX(touch_order) AS total_touches FROM touchpoints GROUP BY user_pseudo_id ) SELECT t.source, t.medium, SUM(1.0 / jl.total_touches) AS linear_credit, -- Equal share per touchpoint COUNT(DISTINCT t.user_pseudo_id) AS touched_users FROM touchpoints t JOIN journey_length jl ON t.user_pseudo_id = jl.user_pseudo_id GROUP BY t.source, t.medium ORDER BY linear_credit DESC; ``` **Pros:** - Full control over methodology - Access to all models - Can customize for your business **Cons:** - Requires SQL expertise - Must maintain the queries - No real-time reporting - Need BigQuery skills ### Option 2: Third-Party Attribution Tools Several tools provide the attribution models GA4 removed: | Tool | Models Available | Pricing | |------|-----------------|---------| | **mbuzz** | All standard + custom Shapley/Markov | Usage-based | | **Segment** | All standard + custom | Enterprise | | **Supermetrics** | Rule-based reconstruction | Mid-market | | **Triple Whale** | DTC-focused, all models | E-commerce | *[Chart omitted from the text version: Embeddable model comparison for articles]* *[Chart omitted from the text version: mbuzz product showcase — renders the canonical product mock (embed variant). See marketing_site_review_spec.md P0-1; one partial, every surface.]* ### Option 3: Run Multiple Tools in Parallel Many teams now run: - **GA4** for traffic analysis and last-click baseline - **Third-party tool** for multi-touch attribution - **Incrementality tests** to validate both This is more work but gives you the complete picture. ## Making the Best of GA4's DDA If you're going to use GA4's data-driven attribution, here's how to get the most from it: ### 1. Ensure You Meet Volume Thresholds Each conversion type needs its own volume. Here's how it plays out in practice: | Business Type | Monthly Conversions | DDA Status | |--------------|-------------------|------------| | Early-stage SaaS (50 signups/mo) | 50 | Falls back to last-click | | Mid-market e-comm (300 purchases/mo) | 300 | Falls back to last-click | | Scaling DTC (800 purchases/mo) | 800 | DDA active | | Enterprise lead gen (150 MQLs/mo) | 150 | Falls back to last-click | Most startups and mid-market companies don't hit the threshold. If you're under 400 monthly conversions per type, GA4's "data-driven" attribution is actually just last-click with a fancier name. ### 2. Use DDA for Direction, Not Precision DDA tells you which channels are likely over/undervalued relative to last-click. Use it for: - Identifying channels that might deserve more investment - Flagging channels that may be over-credited - General budget allocation guidance Don't use it for: - Precise ROAS calculation - Exact channel comparisons - Replacing incrementality testing ### 3. Compare DDA to Last-Click Regularly Run this comparison monthly: | Channel | DDA Credit | Last-Click Credit | Difference | |---------|-----------|-------------------|------------| | Paid Social | 28% | 8% | +250% (undervalued by LC) | | Email | 15% | 35% | -57% (overvalued by LC) | | Organic | 22% | 20% | +10% (fairly valued) | Large differences reveal channel role insights. ### 4. Validate with Holdout Tests DDA is still correlational. Validate with incrementality: 1. Pick a channel DDA says is valuable 2. Run a geo-holdout test 3. Compare attributed lift to incremental lift 4. Calibrate your trust in DDA accordingly ## What This Means for the Future Google's move signals a broader industry shift: ### 1. ML-Based Attribution Becomes Default Expect all major platforms to push algorithmic attribution. The question isn't whether to use data-driven—it's which algorithm and whose implementation. ### 2. First-Party Data Becomes More Critical As third-party cookies die, the properties with more first-party data will have better attribution. GA4's DDA will work better for larger sites. ### 3. Incrementality Testing Rises As attribution becomes more opaque, incrementality testing becomes the ground truth. Teams that can run proper holdout tests will have unfair advantages. ### 4. Tool Fragmentation Increases No single tool will do everything. Expect tech stacks to include: - Analytics tool (GA4, etc.) - Attribution tool (specialized) - MMM tool (for long-term planning) - Experimentation platform (for validation) ## Summary Google removed first-touch, linear, time-decay, and position-based attribution from GA4—leaving only last-click and data-driven. **Key facts:** - DDA uses Shapley values (game theory), not Markov chains - Requires 400+ conversions per conversion type to function - Falls back to last-click below threshold - Is a black box you can't inspect or customize **What to do:** 1. Check if you meet DDA volume thresholds (many don't) 2. For removed models, use BigQuery exports or third-party tools 3. Treat DDA as directional guidance, not precise measurement 4. Validate with incrementality testing **Bottom line:** The models Google removed were useful for specific purposes. If you need them, you now need to build them yourself or buy them elsewhere. ## Further Reading **On GA4's Data-Driven Attribution:** - [GA4 DDA Complete Guide](https://www.webstarresearch.com/blog/ga4-data-driven-attribution-complete-guide-2024/) — Technical walkthrough of how DDA works - [Adswerve: Google's DDA Explained](https://adswerve.com/blog/googles-ga4-data-driven-attribution-model-explained) — Implementation details - [Arcalea: Understanding GA4 DDA](https://arcalea.com/blog/ga4-data-driven-attribution-0) — Practitioner perspective **On Shapley Values in Attribution:** - [Shapley, L.S. (1953)](https://www.rand.org/pubs/papers/P295.html) — Original paper on the Shapley value - [OWOX: Data-Driven Attribution](https://www.owox.com/blog/articles/data-driven-attribution) — Comparison of DDA approaches **On Building Your Own:** - [Stacktonic: Build DDA with BigQuery and Python](https://stacktonic.com/article/build-a-data-driven-attribution-model-using-google-analytics-4-big-query-and-python) — Technical tutorial **Related Articles:** - [What is Multi-Touch Attribution?](/articles/what-is-multi-touch-attribution) — The complete MTA guide - [Multi-Touch Attribution Tools Compared](/articles/multi-touch-attribution-tools-compared) — Find the right tool for your needs ## Key takeaways - GA4 removed 4 models: first-touch, linear, time-decay, position-based - Only last-click and data-driven attribution remain - Google's DDA uses Shapley values with time decay—not Markov chains - DDA requires 400+ conversions per conversion type to function - For removed models, use third-party tools or build in BigQuery ## FAQ **When did GA4 remove the attribution models?** Google removed first-touch, linear, time-decay, and position-based models in late 2023 (around October-November). The change applied to all GA4 properties, with no option to keep the old models. **Can I still use first-touch attribution in GA4?** No, not natively in GA4. You can export raw data to BigQuery and build first-touch attribution yourself, or use a third-party attribution tool like mbuzz that supports all models. **What algorithm does GA4 data-driven attribution use?** GA4's DDA uses Shapley values—a game theory algorithm developed by Nobel laureate Lloyd Shapley. It calculates each channel's marginal contribution across all possible touchpoint combinations, with an added time-decay factor that weights recent interactions more heavily. **Why is GA4 data-driven attribution different from Markov chain attribution?** Shapley values (GA4) calculate credit by measuring marginal contribution across all permutations. Markov chains calculate credit by modeling state transitions and removal effects. Markov works with less data and is computationally faster—Shapley is more theoretically 'fair' but needs more conversions. **How many conversions does GA4 DDA need?** GA4 requires at least 400 conversions per conversion type over 28 days, with a path length of 2+ interactions. Below this threshold, GA4 falls back to last-click. Some sources cite 15,000 clicks and 600 conversions for optimal results. **Does 'data-driven attribution' mean causal attribution?** No, and this is the most common misunderstanding. GA4's DDA is correlational — it learns which touchpoint patterns correlate with conversions in your historical data. Causal attribution requires controlled experiments. The Dropbox IEEE Access paper (Chivukula et al., 2026) measured the gap: click-based attribution (including DDA) overstates causal impact by 2 to 10x compared to geo-incrementality experiments. DDA is better than last-click for distributing credit; it is not a substitute for incrementality testing. **Where does GA4 DDA sit on the measurement maturity ladder?** Avinash Kaushik — formerly Google's Digital Marketing Evangelist, now strategy advisor to Measured — frames marketing measurement as a three-rung ladder: last-click ('dead'), multi-touch attribution ('smart'), and incrementality/MMM ('super smart'). GA4 DDA sits in the middle — it's MTA rather than last-click, but it's a black-box single-model implementation of MTA. Independent MTA tools that run multiple models side-by-side sit on the same rung but expose more of the underlying math. The IAB Australia State of Data 2026 found 60-75% of ad buyers say their measurement falls short of where they want it to be — most are stuck on rung 1 or low rung 2.