Attribution Thresholds Worth Writing Down

Why Attribution Thresholds Matter

Attribution output drives budget allocation, channel scaling decisions, and creative optimization. Without written thresholds, teams chase statistical noise, over-invest in low-signal channels, and misallocate spend.

The core problem: attribution is never ground truth. iOS privacy changes, cross-device journeys, and offline touchpoints create blind spots. Operators need decision rules that acknowledge uncertainty rather than pretend it doesn't exist.

Written thresholds force clarity. They separate signal from noise, define when to trust a model and when to revert to simpler logic, and create a shared language across marketing, analytics, and finance.

Model Selection Thresholds

Choose your attribution model based on average order value, customer journey length, and data completeness. The wrong model creates systematic bias.

First-touch: Use when AOV < $50, repeat purchase rate < 15%, or conversion paths are short (< 3 touchpoints median). Fast, defensible, low data overhead.
Last-touch: Use when AOV $50 - $300, repeat rate 15% - 40%, and most conversions happen within 7 days of final click. Standard for performance marketing.
Linear: Use when AOV > $300, repeat rate > 40%, or median path length > 5 touchpoints. Assumes equal contribution across all touches.
Time-decay: Use when you have clear seasonal patterns, long consideration windows (> 30 days), and want to weight recent touchpoints higher. Requires 6+ months of clean data.
Algorithmic (data-driven): Use only if you have > 10,000 conversions/month, < 5% data loss on key channels, and can remodel monthly. Otherwise, the model overfits to noise.

Data Quality Gates

Before trusting any attribution output, audit data completeness. Set hard thresholds for when to pause reliance on the model.

UTM coverage: > 85% of paid traffic must have valid utm_source, utm_medium, utm_campaign. Below 85%, revert to channel-level last-touch.
Cross-device tracking: If > 30% of users are logged out or use private browsing, algorithmic models lose 15% - 25% accuracy. Document this loss.
Conversion lag: If > 20% of conversions occur > 30 days after last touch, your window is too short. Extend lookback or switch to first-touch.
Channel data loss: If any paid channel has > 10% missing data (API outages, pixel fires), exclude it from multi-touch models for that period.
Organic/direct baseline: If organic + direct > 40% of conversions, your paid attribution is likely overstating channel contribution. Cap paid channel credit at 60% of total revenue.

Threshold for Reversion to Simple Models

Complex attribution models fail silently. Set explicit conditions for downgrading to first-touch or last-touch.

Revert to last-touch if: (1) month-over-month model output changes > 20% without campaign changes, (2) any channel shows > 40% attribution variance across cohorts, or (3) predicted channel ROI contradicts actual spend data by > 15%.

Revert to first-touch if: (1) you cannot explain the model's weighting logic to finance in < 5 minutes, (2) data loss exceeds thresholds above, or (3) the model has not been retrained in > 90 days.

Budget Allocation Thresholds

Attribution output should inform budget moves, not dictate them. Set guardrails.

Do not shift > 15% of monthly budget based on a single month of attribution data. Use 3-month rolling average.
Do not kill a channel if attributed ROI drops below target for 1 month. Require 2 consecutive months + qualitative evidence (creative fatigue, audience saturation).
Do not scale a channel > 30% month-over-month based on attribution alone. Require incrementality testing or holdout validation first.
Do not trust attributed ROAS below 1.5x. Below this threshold, the channel is likely receiving credit for organic/direct conversions.
Do not allocate budget to a channel with < 100 attributed conversions in the period. Sample size is too small for reliable ROI calculation.

Common Attribution Failure Modes

Recognize when attribution is lying to you.

Halo effect: Paid channels get credit for conversions they influenced but did not drive. Symptom: attributed ROAS is 2x - 3x higher than incrementality test results.
Cannibalization blindness: Adding a new channel steals conversions from existing channels, but attribution credits both. Symptom: total attributed revenue grows > 10% while actual revenue grows < 5%.
Seasonal drift: Attribution model trained on Q4 data performs poorly in Q1. Symptom: model output diverges from actual spend performance by > 20% in new season.
Cohort bias: Attribution varies wildly by customer acquisition cohort (new vs. repeat). Symptom: same channel shows 2x - 3x ROI variance across cohorts. Solution: segment attribution by cohort.
Platform data lag: Attribution platform reports data 24 - 48 hours late. Symptom: real-time budget decisions are made on stale data. Threshold: do not use attribution for daily bid adjustments.

Audit Cadence and Documentation

Attribution models drift. Schedule audits and document decisions.

Monthly: Compare attributed revenue by channel to actual spend. If variance > 10%, investigate.

Quarterly: Remodel or retrain. If model has not been updated in 90 days, it is stale.

Annually: Audit data quality gates. Confirm UTM coverage, conversion lag, and cross-device tracking are still within thresholds.

Document every model change, threshold adjustment, and reversion decision. This creates institutional memory and prevents repeated mistakes.

Questions

FAQ

Should we use Shopify's native attribution or a third-party platform?

Shopify's native attribution uses last-click and first-click models. Use it if your data quality gates pass and you want simplicity. Switch to a third-party platform (Littledata, Triple Whale, Northbeam) only if you need multi-touch modeling and have > $50k/month paid spend to justify the cost. The platform matters less than the model selection and data quality.

What if our repeat purchase rate is 50%? Which model should we use?

At 50% repeat rate, linear or time-decay models are justified. Start with linear (equal weight across all touchpoints) for 3 months. If you can document that recent touchpoints drive more conversions than early ones, switch to time-decay with a 7-day half-life. Revert to last-touch if the model output becomes unstable (> 20% month-over-month variance).

How do we handle iOS privacy and attribution loss?

iOS privacy creates 15% - 30% data loss depending on your audience. Document this loss explicitly. Do not use algorithmic models if loss exceeds 30%. Instead, use last-touch on tracked conversions and assume organic/direct is underreported. For budget allocation, cap paid channel credit at 60% of total revenue and treat the remainder as unattributed. Run incrementality tests quarterly to validate paid channel contribution.

When should we stop trusting attribution and just use incrementality testing?

Incrementality testing (holdout groups, geo-experiments) is ground truth but expensive and slow. Use it to validate attribution quarterly, not as a daily decision tool. If attributed ROAS diverges from incrementality test results by > 30%, your attribution model is broken. Revert to simple last-touch and remodel. For channels with < $10k/month spend, skip attribution entirely and use incrementality testing instead.

Want this on your account?

Thirty minutes. Bring the number that keeps you up.

Book a demo