Multi Touch Attribution Skill

---
name: multi-touch-attribution
description: Use this skill when the user wants to do multi-touch attribution, marketing attribution modeling, or figure out which channels or campaigns drive purchases. Also trigger when the user uploads marketing data from multiple platforms (Google Ads, Meta, analytics tools, CRM exports) and wants to combine or compare them. Trigger for phrases like "attribution model", "what's driving conversions", "which campaigns work", "marketing mix", "channel performance", or any mention of combining ad platform data with purchase data. This skill prevents the common mistake of feeding messy, mismatched marketing data into an LLM and getting confident-sounding nonsense back.
---

# Multi-Touch Attribution

This skill helps users build reliable multi-touch attribution — or stops them from wasting time when their data isn't ready.

## Why This Skill Exists

LLMs are bad at attribution when the input data is messy. The typical failure looks like this:

1. User dumps exports from several ad platforms, dashboards, and analytics tools
2. Each source defines metrics differently — attribution windows, naming, event definitions
3. The LLM produces answers that sound right but aren't grounded in consistent data
4. Nobody catches the errors because the output looks polished

The core problem: fragmented, inconsistent data cannot produce reliable attribution, no matter how good the model is. The model will fill gaps with guesses and present them with confidence.

## Before You Start: Data Readiness Check

Run through these questions with the user before touching any data:

### 1. Do your sources agree on what a "conversion" means?

Different platforms count conversions differently. Google Ads may use a 30-day click window. Meta may use a 7-day click or 1-day view window. Your CRM tracks actual purchases.

Ask the user:
- What counts as a conversion in each data source?
- What attribution window does each platform use?
- Is there a single source of truth for actual purchases (e.g., CRM, order database)?

If there's no single source of truth for purchases, stop here. Help the user pick one before going further.

### 2. Can you join the data?

Attribution needs a way to connect a user's journey across channels. This means some shared key — a customer ID, email, cookie, or device ID.

Ask the user:
- Is there a common identifier across your data sources?
- Can you trace a single customer's path from ad click to purchase?

If there's no join key, the best you can do is aggregate-level analysis (media mix modeling), not user-level attribution. Tell the user this plainly.

### 3. Are naming conventions consistent?

Campaign names, UTM parameters, and channel labels often differ across platforms. "Brand Search" in one tool might be "brand_search_google" in another.

Ask the user:
- Do your campaign names follow a consistent pattern?
- Are UTM parameters set up and used across all channels?

If naming is inconsistent, build a mapping table first.

## Step-by-Step Process

Only proceed here once the data readiness check passes.

### Step 1: Establish a Single Conversion Source

Pick one system as the source of truth for purchases. Usually this is the CRM or order database. Everything else is a touchpoint source, not a conversion source.

### Step 2: Build a Touchpoint Log

Create a single table of all marketing touches. Each row should have:
- Customer ID (the join key)
- Timestamp
- Channel (mapped to a consistent naming scheme)
- Campaign name (mapped to a consistent naming scheme)
- Interaction type (click, view, email open, etc.)

Pull this from your ad platforms, analytics tools, and email system. Use the mapping table from the readiness check to make names consistent.

### Step 3: Join Touchpoints to Conversions

Match the touchpoint log to the conversion source using the shared customer ID. For each purchase, you now have an ordered list of touches that led to it.

### Step 4: Choose an Attribution Model

Now — and only now — apply a model. Common options:

- *Last touch*: All credit to the final touchpoint. Simple but misleading for long journeys.
- *First touch*: All credit to the first touchpoint. Good for understanding awareness.
- *Linear*: Equal credit to every touch. Fair but bland.
- *Time decay*: More credit to touches closer to conversion. Often the best starting point.
- *Position-based*: 40% to first, 40% to last, 20% split among the middle. A reasonable default.
- *Data-driven*: Uses your actual data patterns. Needs a large dataset (thousands of conversions minimum).

Help the user pick based on their data volume and their question. If they have fewer than 500 conversions, stick with rule-based models (first five above). Data-driven models need scale.

### Step 5: Compute and Validate

Run the chosen model. Then validate:
- Do the results match your gut sense of what works? If not, investigate — don't just accept.
- Compare two models side by side. If they tell wildly different stories, your data may still have problems.
- Check for channels with suspiciously high or low credit.

## What the LLM Can and Cannot Do Here

*Good uses of the LLM in this process:*
- Cleaning and mapping messy campaign names
- Writing code to join and transform data
- Explaining model tradeoffs in plain language
- Spotting oddities in the data (missing values, duplicates, date gaps)

*Bad uses of the LLM in this process:*
- Asking it to "figure out what drove purchases" from raw platform exports
- Trusting its narrative when the underlying data isn't joined or consistent
- Treating its confidence as evidence

## Common Pitfalls

1. *Platform double-counting*: Every ad platform takes credit for the same conversion. This is normal — it's why you need a single conversion source.
2. *Missing offline touches*: If sales calls, events, or in-store visits matter, your touchpoint log is incomplete. Acknowledge the gap.
3. *Too few conversions*: Small datasets make all models unreliable. Be honest about this.
4. *Confusing correlation with causation*