Do this once campaigns are running
Attribution needs data: at least a month of campaigns with GCLIDs and UTMs saved on leads, and some leads that have moved through your sales stages. If you're in your first week, go back to the Quick start. Come back here in month two or three.
| Model | Gives credit to | Good for | Misleads when |
|---|---|---|---|
| Last click | The last ad click (or last non-direct visit in GA4) before converting | Simple, easy to explain, matches what sales sees | Brand search and retargeting look great; whatever created the demand looks worthless |
| First touch | The first recorded visit or click | Seeing which channels bring new people in | Ignores everything that moved the deal forward |
| Position-based (U-shaped) | 40% first, 40% lead-creating touch, 20% spread across the middle | A balanced default for B2B reporting you build yourself | The weights are arbitrary; treat it as a convention, not a measurement |
| Data-driven (Google Ads, GA4) | Touches, based on comparing paths that converted with paths that didn't | Bidding inside Google Ads; it's the default | It only sees Google's touchpoints and the conversions you send it; it can't see sales calls, events or referrals |
| Self-reported | Whatever the buyer says brought them | Catching word of mouth, podcasts, communities, AI assistants | People forget and simplify; use it alongside tracked data |
Google Ads now offers only data-driven and last click; the older rule-based models (first click, linear, time decay, position-based) were removed in 2023. GA4 works the same way.
If you want first touch or position-based, you build it yourself from exported data, which this lesson shows. As of October 2026
## Why B2B makes this hardA typical path for a mid-size deal: someone hears about you in a community thread, searches your category weeks later and clicks an ad, reads two pages, leaves.
A colleague later searches your brand name, clicks the brand ad and requests a demo. Sales runs a six-week cycle with three more people. Last click credits the brand ad. First touch credits the category ad.
Neither sees the community thread or the sales work.
Three practical consequences:
It takes ten minutes and catches what tracking misses. Make it a free-text field (not a dropdown, which biases answers), optional, on the demo or sign-up form. Have your LLM categorise the answers monthly.
Expect a large share of "Google" answers; that's useful in itself when you compare it with tracked paid search leads.
## Build your own attribution analysis with an LLMYou don't need an attribution tool to start. Three exports and an LLM that can run code are enough.
### 1. Export the data| Source | Export | Key columns |
|---|---|---|
| CRM | All leads created in the last 6 to 12 months, with current stage | Lead ID, email or domain, created date, GCLID, UTM fields, "how did you hear" answer, stage, opportunity amount, close date, region |
| Google Ads | Campaign report by day (or click report for the last 90 days if you want GCLID-level detail) | Date, campaign, ad group, cost, clicks, conversions; GCLID in the click performance report |
| GA4 | Exploration with first user source/medium/campaign and session source/medium/campaign for converting users, by date | Date, first-user source/medium/campaign, session source/medium/campaign, key events |
The join key is the GCLID and UTMs saved on each lead (that's why stage-one tracking saves them). GA4's free exports don't include user-level IDs, so you join GA4 at the campaign and date level, not per lead.
### 2. Ask the LLM to build the analysis ### 3. Check the result before you trust itAn illustrative summary for a six-month window (not a benchmark):
| Channel (first touch) | Leads | SQLs | Pipeline | Ad cost | Pipeline ÷ cost |
|---|---|---|---|---|---|
| Paid Search Non-brand | 410 | 62 | $1,240,000 | $96,000 | 12.9× |
| Paid Search Brand | 180 | 48 | $1,010,000 | $9,000 | 112× |
| Organic | 520 | 71 | $1,420,000 | – | – |
| Referral / word of mouth | 90 | 30 | $780,000 | – | – |
Reading it: brand search looks extraordinary, but it mostly captures people who already know you. Non-brand search at 12.9× pipeline to cost is the number that tells you whether to put more money into Search.
Compare it with your target return (see Budget decisions).
Starter budget: keep it simple
With a few dozen leads a month, a monthly spreadsheet is enough: leads, SQLs and pipeline from paid search (GCLID present) compared with cost.
Run the prompt above quarterly rather than monthly; with small numbers, monthly changes are mostly noise.
Intermediate budget
Run the analysis monthly and keep the script in a shared folder or repository. Add a position-based view once you capture both first and last touch in the CRM (most CRMs can store "original source" and "latest source").
Advanced budget
Move the joins into a warehouse (BigQuery has free Google Ads and GA4 export connectors; GA4's BigQuery export includes event-level data).
Ask your LLM to write the SQL views instead of a Python script, and schedule them.
Consider incrementality tests (geo holdouts or campaign experiments) for big budget questions; attribution shows correlation, a holdout shows what the ads caused.
Data-driven attribution (DDA) is the default for new conversion actions. It spreads credit across Google ad interactions (Search, YouTube, Display) based on patterns in your account.
It matters mainly for bidding: with DDA, Smart Bidding values upper-funnel clicks a bit more than with last click.
Keep DDA on your conversion actions unless you have a specific reason to change it. Just don't read Google Ads' conversion columns as a cross-channel truth. They only see Google touchpoints and only what you upload.
Limits of LLM-built analysis
LLMs write good join and aggregation code, but they will also happily join on the wrong key, silently drop rows with missing values, or describe a pattern in 4 data points as a trend.
Always ask for row counts before and after each step, read the code's join conditions, and spot-check results against the source systems.
Don't paste personal data you aren't allowed to share; remove names and emails (keep a hashed ID) if your policy requires it.