Attribution Shift After iOS Privacy

Business Context

You’re the analytics lead for DoorDash’s growth marketing org (US + Canada), which spends $45M/month across paid search, paid social, affiliate, and influencer channels to drive first-time orders. The core business goal this quarter is to grow new customer acquisition while keeping blended CAC under control.

After an iOS privacy-related change (SKAdNetwork updates + stricter consent prompts) and a parallel shift in media mix, the CMO notices that last-click attribution reports show paid search driving +18% more “new customer conversions” month-over-month, while paid social appears to drop -22%. However, finance is skeptical because overall new customers (from backend order data) are only up +3%, and total spend is up +12%. The team is debating whether to reallocate $8M/month from paid social to paid search based on the last-click dashboard.

At the same time, your data science team has rolled out a data-driven attribution (DDA) model (Markov/Shapley-style multi-touch) that assigns fractional credit across touchpoints. The DDA dashboard tells a different story: paid social’s contribution is flat, and paid search’s incremental contribution is much smaller than last-click suggests.

Stakeholders (CMO, finance, channel managers) ask you to explain the difference between last-click and DDA in a way that leads to a decision: Should we move budget? If so, how much and with what guardrails?

Data Available

Source	What it contains	Notes
`ad_clicks`	click_id, user_id, timestamp, channel, campaign_id, device_os, cost	Click-level; some iOS users missing user_id (consent)
`ad_impressions`	impression_id, (sometimes) user_id, timestamp, channel, campaign_id	View-through is partially observable
`attribution_last_click`	order_id, click_id, attributed_channel, attributed_campaign_id	Deterministic last-touch rules
`orders`	order_id, user_id, order_timestamp, is_first_order, subtotal, fees, promo_cost	Source of truth for conversions
`user_identity_map`	hashed_email, user_id, device_id, confidence_score	Cross-device stitching with uncertainty
`experiments`	experiment_id, user_id, variant, start_ts, end_ts	Some geo holdouts exist

Your Task (What you must produce)

Explain, precisely and practically, the difference between last-click attribution and data-driven attribution in this context (multi-channel acquisition with partial observability on iOS).
Define the primary KPI you would use to compare channels under each method (e.g., attributed conversions, attributed revenue, CAC) and call out what each method optimizes for.
Provide at least 4 concrete failure modes where last-click can mislead budget decisions here (e.g., demand capture vs demand creation, retargeting bias, missing identifiers, channel interaction effects).
Describe how a DDA model would typically assign credit (high-level approach is fine), including what assumptions it makes and what data quality requirements it has.
Recommend a budget decision framework for the next 30 days: what you would change (if anything), how you would validate incrementality (e.g., geo holdouts), and which guardrail metrics you would monitor.

Constraints:

You must deliver a recommendation within 7 days.
You cannot run a full randomized experiment across all channels immediately, but you can propose targeted holdouts and use existing geo experiments.
Compliance requires that any user-level stitching must respect consent flags.

Business Context

Data Available

Source	What it contains	Notes
`ad_clicks`	click_id, user_id, timestamp, channel, campaign_id, device_os, cost	Click-level; some iOS users missing user_id (consent)
`ad_impressions`	impression_id, (sometimes) user_id, timestamp, channel, campaign_id	View-through is partially observable
`attribution_last_click`	order_id, click_id, attributed_channel, attributed_campaign_id	Deterministic last-touch rules
`orders`	order_id, user_id, order_timestamp, is_first_order, subtotal, fees, promo_cost	Source of truth for conversions
`user_identity_map`	hashed_email, user_id, device_id, confidence_score	Cross-device stitching with uncertainty
`experiments`	experiment_id, user_id, variant, start_ts, end_ts	Some geo holdouts exist

Your Task (What you must produce)

Explain, precisely and practically, the difference between last-click attribution and data-driven attribution in this context (multi-channel acquisition with partial observability on iOS).
Define the primary KPI you would use to compare channels under each method (e.g., attributed conversions, attributed revenue, CAC) and call out what each method optimizes for.
Provide at least 4 concrete failure modes where last-click can mislead budget decisions here (e.g., demand capture vs demand creation, retargeting bias, missing identifiers, channel interaction effects).
Describe how a DDA model would typically assign credit (high-level approach is fine), including what assumptions it makes and what data quality requirements it has.
Recommend a budget decision framework for the next 30 days: what you would change (if anything), how you would validate incrementality (e.g., geo holdouts), and which guardrail metrics you would monitor.

Constraints:

You must deliver a recommendation within 7 days.
You cannot run a full randomized experiment across all channels immediately, but you can propose targeted holdouts and use existing geo experiments.
Compliance requires that any user-level stitching must respect consent flags.

Business Context

Data Available

Source	What it contains	Notes
`ad_clicks`	click_id, user_id, timestamp, channel, campaign_id, device_os, cost	Click-level; some iOS users missing user_id (consent)
`ad_impressions`	impression_id, (sometimes) user_id, timestamp, channel, campaign_id	View-through is partially observable
`attribution_last_click`	order_id, click_id, attributed_channel, attributed_campaign_id	Deterministic last-touch rules
`orders`	order_id, user_id, order_timestamp, is_first_order, subtotal, fees, promo_cost	Source of truth for conversions
`user_identity_map`	hashed_email, user_id, device_id, confidence_score	Cross-device stitching with uncertainty
`experiments`	experiment_id, user_id, variant, start_ts, end_ts	Some geo holdouts exist

Your Task (What you must produce)

Explain, precisely and practically, the difference between last-click attribution and data-driven attribution in this context (multi-channel acquisition with partial observability on iOS).
Define the primary KPI you would use to compare channels under each method (e.g., attributed conversions, attributed revenue, CAC) and call out what each method optimizes for.
Provide at least 4 concrete failure modes where last-click can mislead budget decisions here (e.g., demand capture vs demand creation, retargeting bias, missing identifiers, channel interaction effects).
Describe how a DDA model would typically assign credit (high-level approach is fine), including what assumptions it makes and what data quality requirements it has.
Recommend a budget decision framework for the next 30 days: what you would change (if anything), how you would validate incrementality (e.g., geo holdouts), and which guardrail metrics you would monitor.

Constraints:

You must deliver a recommendation within 7 days.
You cannot run a full randomized experiment across all channels immediately, but you can propose targeted holdouts and use existing geo experiments.
Compliance requires that any user-level stitching must respect consent flags.

Business Context

Data Available

Source	What it contains	Notes
`ad_clicks`	click_id, user_id, timestamp, channel, campaign_id, device_os, cost	Click-level; some iOS users missing user_id (consent)
`ad_impressions`	impression_id, (sometimes) user_id, timestamp, channel, campaign_id	View-through is partially observable
`attribution_last_click`	order_id, click_id, attributed_channel, attributed_campaign_id	Deterministic last-touch rules
`orders`	order_id, user_id, order_timestamp, is_first_order, subtotal, fees, promo_cost	Source of truth for conversions
`user_identity_map`	hashed_email, user_id, device_id, confidence_score	Cross-device stitching with uncertainty
`experiments`	experiment_id, user_id, variant, start_ts, end_ts	Some geo holdouts exist

Your Task (What you must produce)

Explain, precisely and practically, the difference between last-click attribution and data-driven attribution in this context (multi-channel acquisition with partial observability on iOS).
Define the primary KPI you would use to compare channels under each method (e.g., attributed conversions, attributed revenue, CAC) and call out what each method optimizes for.
Provide at least 4 concrete failure modes where last-click can mislead budget decisions here (e.g., demand capture vs demand creation, retargeting bias, missing identifiers, channel interaction effects).
Describe how a DDA model would typically assign credit (high-level approach is fine), including what assumptions it makes and what data quality requirements it has.
Recommend a budget decision framework for the next 30 days: what you would change (if anything), how you would validate incrementality (e.g., geo holdouts), and which guardrail metrics you would monitor.

Constraints:

You must deliver a recommendation within 7 days.
You cannot run a full randomized experiment across all channels immediately, but you can propose targeted holdouts and use existing geo experiments.
Compliance requires that any user-level stitching must respect consent flags.

Interview Guides

Business Context

Data Available

Your Task (What you must produce)

Attribution Shift After iOS Privacy

Business Context

Data Available

Your Task (What you must produce)

Your Answer

Attribution Shift After iOS Privacy

Business Context

Data Available

Your Task (What you must produce)

Attribution Shift After iOS Privacy

Business Context

Data Available

Your Task (What you must produce)

Your Answer