Prompt
How should I build attribution if I only have partial tracking data?
Latest observation
If you only have partial tracking data, the right approach is usually “probabilistic attribution with explicit uncertainty” rather than pretending you have full-funnel truth.
Here’s a practical way to build it:
1) Start with what you do know
Inventory your data sources and gaps:
- Known conversions: purchases, signups, leads
- Known touchpoints: ad clicks, email opens/clicks, site visits, app events
- Missing areas: cross-device identity, offline conversions, walled-garden platforms, cookie loss, view-throughs
Then classify data quality by channel/source:
- High confidence: deterministic IDs, logged-in users, direct clicks
- Medium confidence: session-based web data, UTM-tagged traffic
- Low confidence: view-through, cookie-limited, modeled platform data
2) Use a hybrid attribution model
A good partial-data setup often combines:
- Deterministic attribution where you have user-level paths
- Rule-based attribution for incomplete paths
- Modeled / probabilistic attribution for missing touches
Examples:
- If a user is logged in, use user-level paths.
- If not, use session-level or campaign-level paths.
- For channels with missing exposure data, use lift tests or modeled estimates.
3) Avoid overclaiming precision
With incomplete tracking:
- Don’t rely on last-click as “truth” unless that’s explicitly your business rule.
- Report attribution as a range or confidence interval when possible.
- Separate:
- Observed conversions
- Attributed conversions
- Modeled/estimated conversions
4) Fill gaps with incrementality testing
Partial tracking often makes attribution noisy. Use tests to calibrate:
- A/B or holdout experiments
- Geo tests
- Audience suppression tests
- Incrementality studies by channel
These help answer: “What drove lift?” rather than “What got credit?”
5) Build a simple model first
A robust first version:
- Primary credit: based on observed user journeys
- Adjustment factor: based on known tracking loss by channel
- Normalization: make sure total attributed conversions don’t exceed total measured conversions + modeled lift
For example:
- Email has near-complete tracking → small adjustment
- Paid social has limited user-level tracking → larger modeled adjustment
- Organic/direct often absorb missing attribution, so treat carefully
6) Account for bias from missing data
Missing data is rarely random. Common biases:
- Mobile Safari / iOS undercounts
- Cookie blockers undercount paid media
- Walled gardens hide exposure data
- Logged-in users are overrepresented in deterministic paths
You should estimate whether missingness is:
- Random
- Systematic by device/channel/audience
That determines whether simple weighting is enough or you need a more advanced model.
7) Use a hierarchy of evidence
When making decisions, prioritize:
- Experiment results
- Deterministic user-level attribution
- Modeled attribution
- Heuristic fallback rules
This keeps the model useful without overstating certainty.
8) Communicate attribution as decision support, not absolute truth
The goal is usually to guide budget allocation, not to reconstruct reality perfectly.
Best practice reporting:
- Channel credit
- Data completeness score
- Confidence level
- Incrementality evidence
- Sensitivity analysis
9) If you want a concrete framework
A simple implementation could look like:
- Step 1: Build known journey paths from available IDs
- Step 2: Assign rule-based credit for observed paths
- Step 3: Estimate missing exposure using historical loss rates or test-based lift
- Step 4: Calibrate with incrementality experiments
- Step 5: Report both attributed and modeled values
If you want, I can also give you:
- a recommended attribution architecture,
- a mathematical model for partial-data attribution, or
- a practical dashboard/reporting template.