Conversion Attribution
“Which touchpoints drove this conversion?”
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
Which touchpoints drove this conversion?
Last-click attribution overvalues bottom-funnel channels and undervalues awareness campaigns, leading to systematic misallocation of ad spend. A brand spending $50M per year on digital ads misallocates 30-40% of budget when relying on last-click, representing $15-20M in wasted or suboptimal spend annually.
Quick answer
Graph-based attribution models the full user journey across channels by connecting touchpoints, conversions, campaigns, and user segments in a relational graph. Instead of over-crediting the last click, the GNN learns how different touchpoint sequences contribute to conversion, assigning credit proportional to each channel's true causal impact.
Approaches compared
4 ways to solve this problem
1. Last-click / first-click attribution
Assign 100% of conversion credit to the final (or first) touchpoint in the user journey. The default in most analytics platforms.
Best for
Simplicity. Works when you need a quick directional read and your conversion paths are short (1-2 touchpoints).
Watch out for
Systematically overvalues bottom-funnel channels and undervalues awareness. Brands that optimize on last-click consistently overspend on search and underspend on display and email.
2. Shapley value / data-driven attribution
Use game-theoretic methods to distribute credit across touchpoints based on their marginal contribution. Google Analytics 4 uses a version of this.
Best for
Better than rule-based models. Accounts for the contribution of each channel relative to all possible orderings.
Watch out for
Assumes touchpoint independence -- the order doesn't matter. In reality, seeing a display ad before an email changes the email's conversion impact. Also computationally expensive for long journeys.
3. Marketing mix modeling (MMM)
Regress aggregate sales on channel-level spend over time. Top-down approach that works with aggregated data, no user-level tracking needed.
Best for
Privacy-compliant environments where user-level tracking is limited. Good for budget allocation across broad channels.
Watch out for
Cannot attribute at the user or conversion level. Misses within-channel dynamics and individual journey patterns. Too coarse for tactical optimization.
4. KumoRFM (relational graph ML)
Connect users, touchpoints, conversions, campaigns, and channels into a single graph. The GNN learns how touchpoint sequences and channel interactions drive conversion at the individual level.
Best for
Highest accuracy for multi-touch attribution. Captures interaction effects between channels (display priming email, email nurturing search) that linear and Shapley models miss.
Watch out for
Requires user-level journey data with timestamps and clear conversion events. Not suited for environments with only aggregate channel data.
Key metric: Brands using graph-based multi-touch attribution recover 30-40% of previously misallocated ad spend compared to last-click models.
Why relational data changes the answer
Attribution is a path problem, not a point problem. The value of an email touchpoint depends on whether a display ad preceded it, which publisher served that display ad, what campaign objective drove the email, and how many days passed between them. Flat attribution models treat each touchpoint independently or assume fixed position-based weights, collapsing this rich sequential structure into a single credit-assignment rule.
Relational models preserve the full journey graph. They learn that for Enterprise users acquired through organic search, a display-then-email-then-search sequence converts at 3x the rate of search alone, and that the display ad on TechNews contributes more than the same creative on a generic news site. On the RelBench benchmark, relational approaches score 76.71 vs 62.44 for single-table baselines -- a gap that in attribution terms means the difference between correctly reallocating $15M and continuing to waste it on over-credited channels.
Last-click attribution is like giving the assist in basketball to whoever made the final pass before a layup. You miss the point guard who ran the play, the screen-setter who created the opening, and the ball handler who drove the defense to collapse. Graph-based attribution watches the entire play from the opening whistle and credits every player for the role they actually played in the score.
How KumoRFM solves this
Graph-powered intelligence for advertising
Kumo connects users, touchpoints, conversions, campaigns, and channels into a unified graph. The GNN learns how different touchpoint sequences contribute to conversion, capturing interaction effects between channels that linear models and even Shapley-based approaches miss. Each conversion gets a per-touchpoint attribution score grounded in the full user journey.
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 | acquisition_source | ltv_tier |
|---|---|---|---|
| U201 | Enterprise | Organic | High |
| U202 | SMB | Paid Search | Medium |
| U203 | Enterprise | Referral | High |
TOUCHPOINTS
| touchpoint_id | user_id | channel | campaign_id | timestamp |
|---|---|---|---|---|
| TP401 | U201 | Display | CMP10 | 2025-02-15 09:00 |
| TP402 | U201 | CMP11 | 2025-02-18 14:30 | |
| TP403 | U201 | Paid Search | CMP12 | 2025-02-20 11:00 |
CONVERSIONS
| conversion_id | user_id | value | timestamp |
|---|---|---|---|
| CVR101 | U201 | $12,500 | 2025-02-20 11:15 |
| CVR102 | U203 | $8,200 | 2025-02-22 16:00 |
CAMPAIGNS
| campaign_id | channel | spend | objective |
|---|---|---|---|
| CMP10 | Display | $120K | Awareness |
| CMP11 | $15K | Nurture | |
| CMP12 | Paid Search | $80K | Conversion |
CHANNELS
| channel | avg_cpa | avg_roas | attribution_window |
|---|---|---|---|
| Display | $45 | 3.2x | 30 days |
| $12 | 8.1x | 7 days | |
| Paid Search | $28 | 5.5x | 14 days |
Write your PQL query
Describe what to predict in 2–3 lines — Kumo handles the rest
PREDICT LIST_DISTINCT(TOUCHPOINTS.channel, -30, 0, days) FOR EACH CONVERSIONS.conversion_id
Prediction output
Every entity gets a score, updated continuously
| CONVERSION_ID | USER_ID | VALUE | DISPLAY_ATTR | EMAIL_ATTR | SEARCH_ATTR |
|---|---|---|---|---|---|
| CVR101 | U201 | $12,500 | 0.28 | 0.35 | 0.37 |
| CVR102 | U203 | $8,200 | 0.15 | 0.52 | 0.33 |
Understand why
Every prediction includes feature attributions — no black boxes
Conversion CVR101 -- User U201
Predicted: Multi-touch: Display 28%, Email 35%, Search 37%
Top contributing features
Email open-to-conversion time
2 days
35% attribution
Search keyword intent score
High
25% attribution
Display ad first-touch awareness
5 days prior
20% attribution
Cross-channel journey length
3 touchpoints
12% attribution
Similar user conversion paths
68% match
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 conversion attribution
What is the best multi-touch attribution model?
Graph-based attribution models outperform both rule-based (last-click, linear) and statistical (Shapley, Markov chain) approaches because they capture interaction effects between channels and temporal sequencing. The key advantage is learning that a display-then-email sequence has a different conversion impact than email-then-display, something position-based and Shapley models cannot express.
How do you replace last-click attribution?
Start by collecting user-level journey data with timestamps across all channels. Then move to a model that respects the sequential, multi-entity nature of the data. Graph-based models are the most accurate option because they connect users, touchpoints, campaigns, and channels into a single structure. Expect to recover 30-40% of previously misallocated ad spend.
What data do you need for multi-touch attribution?
User-level touchpoint data with channel, campaign, and timestamp for each interaction, plus clear conversion events with value. For best results, add channel metadata (average CPA, ROAS), campaign objectives, and user segments. More connected tables means more signal for the model to learn causal contribution patterns.
How does attribution work without third-party cookies?
First-party data becomes critical. Graph models excel here because they extract more signal from the data you do have -- connecting CRM records, email engagement, on-site behavior, and conversion events through natural foreign keys. You lose some cross-site tracking, but the relational structure of your first-party data compensates significantly.
What is the ROI of better attribution modeling?
Brands spending $50M+ on digital ads typically misallocate 30-40% of budget under last-click. Switching to graph-based attribution recovers $15-20M in wasted spend by shifting budget from over-credited bottom-funnel channels to under-credited awareness and nurture campaigns that actually drive incremental conversions.
Bottom line: A brand spending $50M on digital ads recovers $15-20M in misallocated budget by replacing last-click with Kumo's graph-based multi-touch attribution. Every channel gets credit proportional to its true causal contribution.
Related use cases
Explore more ad tech 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.
