Marketing Attribution
“How much incremental revenue does each marketing touchpoint generate?”
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
By submitting, you accept the Terms and Privacy Policy.
Loved by data scientists, ML engineers & CXOs at
A real-world example
How much incremental revenue does each marketing touchpoint generate?
Marketing teams spend millions across channels but cannot measure true incremental impact. Last-touch attribution over-credits the final click; multi-touch models spread credit arbitrarily. Without causal uplift measurement, budget allocation is guesswork — high-performing channels get under-funded while vanity channels consume budget with no real conversion lift. CMOs need to know: 'If we turned off paid search, how many conversions would we actually lose?'
Quick answer
Marketing attribution measures how much incremental revenue each touchpoint actually generates. Counterfactual models compare predicted conversion probability with and without each channel, separating true causal impact from coincidental correlation. This reveals that paid search drives 40% incremental lift for new users but only 3% for existing customers who would have converted anyway.
Approaches compared
4 ways to solve this problem
1. Last-Touch Attribution
Credit 100% of the conversion to the last touchpoint before purchase. The default in Google Analytics, most ad platforms, and CRM reporting.
Best for
Quick reporting when you need a simple answer. Universally available, no setup required.
Watch out for
Massively over-credits bottom-funnel channels (branded search, retargeting) and under-credits awareness channels. A customer who saw 5 ads and then searched your brand name gets 100% credit to branded search. This leads to systematic over-investment in channels that take credit but do not create demand.
2. Multi-Touch Attribution (MTA)
Distribute credit across all touchpoints using rules (linear, time-decay, U-shaped, W-shaped). More nuanced than last-touch but still based on arbitrary weights.
Best for
Teams that want a fairer view of the full funnel without building custom models. A reasonable improvement over last-touch.
Watch out for
The weight distribution is arbitrary. A U-shaped model gives 40% to first touch and 40% to last touch because someone chose those numbers, not because the data supports them. No MTA model answers the causal question: 'Would this conversion have happened without this touchpoint?'
3. Marketing Mix Modeling (MMM)
Aggregate-level regression that correlates total channel spend with total conversions over time. Used by brands with large offline media budgets (TV, radio, out-of-home).
Best for
Measuring the impact of offline channels that cannot be tracked at the individual level. Good for portfolio-level budget decisions.
Watch out for
Operates at the macro level. Cannot measure individual touchpoint impact or distinguish between user segments. Requires 2+ years of data and is typically updated quarterly. Useless for tactical campaign optimization.
4. KumoRFM (Counterfactual Attribution via ASSUMING PQL)
Uses the ASSUMING clause to predict conversion probability with and without each touchpoint. The difference is the true causal uplift. Operates at the individual user level across the full relational graph of users, touchpoints, and conversions.
Best for
Marketing teams that need to measure true incremental impact at the campaign and channel level. Answers the causal question that MTA and MMM cannot.
Watch out for
Requires user-level touchpoint data with timestamps. If your data only has channel-level aggregates without user IDs, you need MMM instead. Also requires historical periods with some variation in channel exposure (not every user saw every ad).
Key metric: Counterfactual attribution reveals that 30-40% of attributed conversions are non-incremental. SAP SALT benchmark: 91% accuracy for relational counterfactual models vs 75% for single-table approaches.
Why relational data changes the answer
User U003 is a new user with no prior brand awareness. Paid search was their first-ever touchpoint. The counterfactual model predicts a 71% conversion probability with paid search and 31% without, a +40% uplift. This is a genuinely incremental conversion. Now consider User U004, a high-value existing customer. Their conversion probability is 91% with paid search and 88% without, only +3% uplift. Paid search gets last-touch credit for both conversions, but the true incremental value is 13x higher for U003.
A flat attribution model cannot distinguish these cases because it does not model the counterfactual scenario. The relational graph enables this by connecting users to their full touchpoint history, segment behavior, and conversion patterns of similar users who did and did not receive each touchpoint. The GNN learns that new users with no organic touchpoints have high paid search uplift, while high-value returning users have near-zero uplift because they would convert through direct or organic channels regardless. On the SAP SALT benchmark, multi-table relational models achieve 91% accuracy vs 75% for single-table models. For attribution specifically, the relational advantage is that the graph naturally encodes the heterogeneity of treatment effects across user segments.
Last-touch attribution is like giving the waiter credit for your restaurant experience because they were the last person you interacted with. Counterfactual attribution asks: what if there was no waiter? You would still eat, but the experience would be worse. What if there was no chef? You would not eat at all. The counterfactual reveals who actually created the outcome vs. who was just present when it happened.
How KumoRFM solves this
Relational intelligence for smarter acquisition
Kumo uses counterfactual PQL queries to answer causal questions. The ASSUMING clause simulates a world where a specific touchpoint exists, and the model compares this to the baseline (without that touchpoint). The difference in predicted conversion probability is the true incremental uplift for that channel. This is not correlation-based attribution — it is learned counterfactual reasoning over the full relational graph of users, touchpoints, and conversions.
From data to predictions
See the full pipeline in action
Connect your tables, write a PQL query, and get predictions with built-in explainability — all in minutes, not months.
Your data
The relational tables Kumo learns from
USERS
| user_id | segment | signup_date |
|---|---|---|
| U001 | high_value | 2025-09-01 |
| U002 | mid_value | 2025-09-15 |
| U003 | new_user | 2025-10-01 |
| U004 | high_value | 2025-10-10 |
TOUCHPOINTS
| touch_id | user_id | channel | campaign | timestamp |
|---|---|---|---|---|
| T01 | U001 | paid_search | brand_q4 | 2025-10-20 |
| T02 | U001 | nurture_series | 2025-10-22 | |
| T03 | U002 | organic | — | 2025-10-18 |
| T04 | U003 | paid_search | prospecting | 2025-10-25 |
| T05 | U004 | webinar | demo_day | 2025-10-28 |
CONVERSIONS
| conversion_id | user_id | revenue | timestamp |
|---|---|---|---|
| CV01 | U001 | $8,400 | 2025-11-01 |
| CV02 | U004 | $12,200 | 2025-11-05 |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT COUNT(CONVERSIONS.*, 0, 14, days) > 0 FOR EACH USERS.USER_ID ASSUMING COUNT(TOUCHPOINTS.* WHERE TOUCHPOINTS.CHANNEL = 'paid_search', 0, 1, days) > 0
Prediction output
Every entity gets a score, updated continuously
| USER_ID | True_PROB (with) | True_PROB (without) | UPLIFT |
|---|---|---|---|
| U001 | 0.87 | 0.62 | +0.25 |
| U002 | 0.34 | 0.29 | +0.05 |
| U003 | 0.71 | 0.31 | +0.40 |
| U004 | 0.91 | 0.88 | +0.03 |
Understand why
Every prediction includes feature attributions — no black boxes
User U003 — new_user segment
Predicted: +40% uplift from paid_search
Top contributing features
New user with no prior brand awareness
new_user
35% attribution
Paid search was first-ever touchpoint
first touch
28% attribution
Campaign — prospecting (top-of-funnel)
prospecting
17% attribution
No organic or referral touchpoints present
0 organic
12% attribution
Similar new users converted 3x more with paid search
3x lift
8% attribution
Feature attributions are computed automatically for every prediction. No separate tooling required. Learn more about Kumo explainability
PQL Documentation
Learn the Predictive Query Language — SQL-like syntax for defining any prediction task in 2–3 lines.
Python SDK
Integrate Kumo predictions into your pipelines. Train, evaluate, and deploy models programmatically.
Explainability Docs
Understand feature attributions, model evaluation metrics, and how to build trust with stakeholders.
Frequently asked questions
Common questions about marketing attribution
What is the difference between attribution and incrementality testing?
Attribution assigns credit for conversions that already happened. Incrementality testing (via A/B holdout experiments) measures whether a channel actually causes conversions. Counterfactual models bridge the gap: they estimate incrementality from observational data without requiring costly holdout experiments, updating continuously as new data arrives.
How much budget is wasted due to bad attribution?
Research consistently shows 30-40% of digital ad spend goes to channels and campaigns that would have converted anyway. Branded search and retargeting are the worst offenders. Counterfactual attribution reveals this waste by measuring the actual causal lift, enabling reallocation to truly incremental channels.
Can counterfactual attribution measure offline channels?
Yes, as long as you have user-level exposure data. For channels like direct mail, events, or sales outreach where you know which users were treated, the ASSUMING clause can estimate the counterfactual. For true mass-media channels (TV, billboard) where individual exposure is unknown, MMM is the better approach.
How does counterfactual attribution handle multi-touch journeys?
The model evaluates each touchpoint's incremental contribution by simulating removal. For a user who saw email, then display, then paid search, the model predicts conversion probability after removing each touchpoint one at a time. Touchpoints that increase conversion probability significantly get high attribution; those that do not change the outcome get low attribution.
Bottom line: Counterfactual attribution reveals that paid search drives 40% incremental lift for new users but only 3% for existing high-value customers — enabling precise budget reallocation that can save $2M+ annually in wasted ad spend.
Related use cases
Explore more acquisition use cases
Topics covered
One Platform. One Model. Infinite Predictions.
KumoRFM
Relational Foundation Model
Turn structured relational data into predictions in seconds. KumoRFM delivers zero-shot predictions that rival months of traditional data science. No training, feature engineering, or infrastructure required. Just connect your data and start predicting.
For critical use cases, fine-tune KumoRFM on your data using the Kumo platform and Research Agent for 30%+ higher accuracy than traditional models.
Book a demo and get a free trial of the full platform: research agent, fine-tune capabilities, and forward-deployed engineer support.
