How the UTM Audit Engine Works: From CSV to A–F Grade

You paste a GA4 source/medium export, and about a minute later you have an A–F grade and a ranked list of what is splitting your channels. This page opens that minute up: every stage the utm audit engine runs, what the Claude API does and does not touch, and why a person approves the result before anything is treated as canonical.

The five stages of the audit engine#

The engine is a short pipeline. Each stage hands a cleaner, more structured view to the next, and only one stage uses AI. Everything runs synchronously for up to roughly 500 links, so you watch it happen rather than wait on a queue. The first public scan has no signup wall.

1. 1

   1. Parse the export

   Each row is split into its utm_source, utm_medium, and utm_campaign values (plus term and content when present). The engine trims stray whitespace, decodes encoded characters, and flags rows with no UTM at all — the ones GA4 drops into Unassigned. See exporting the right GA4 report for the exact columns to include.

2. 2

   2. Cluster the duplicates with Claude

   This is the one AI stage. The Anthropic Claude API reads the distinct values in each dimension and groups semantic duplicates — fb, Facebook, facebook, and meta — along with case variants and typos into a single proposed canonical token. GA4 is case-sensitive, so email, Email, and EMAIL are three different mediums until something collapses them; the engine proposes the collapse, it does not assume it.

3. 3

   3. Grade it A–F

   The engine scores how fragmented the values are and how many sessions are sitting in Unassigned, then returns a letter grade with a one-sentence reason. The grade is deliberately blunt — a number you can put on a slide — and it gives you a sense of how deep the cleanup goes before you spend time on it.

4. 4

   4. Rank issues by traffic impact

   Every issue cluster is sorted by the share of sessions it touches, biggest first. A case duplicate splitting 30% of your paid social outranks a one-off typo on a dead campaign. You read down the list and stop when the curve flattens.

5. 5

   5. Break it down by dimension

   Finally the engine separates the damage across utm_source, utm_medium, utm_campaign, and the Unassigned bucket, so you can see whether the drift is a few mis-cased sources or a campaign-naming free-for-all. Each dimension links straight to its clusters.

Where the AI stops and you decide#

The clustering step is bounded on purpose. The Claude API returns a proposal: here are the values it believes are the same channel, and here is the canonical token it suggests for each group. Nothing about that proposal is applied until you approve it, and approval is one click per group. Override anything that does not match how your team actually talks — if you say paid-social and the engine guessed cpc, you keep paid-social.

Why ranking by traffic impact matters#

A raw list of every inconsistency is noise. The engine ranks clusters by the sessions they affect because that is what changes your numbers. When UTM data is fragmented, as much as 26% of conversions can be credited to the wrong channel, and 10–20% of GA4 sessions commonly land in Unassigned — but those losses concentrate in a few high-traffic clusters, not evenly across the export.

Once you approve the canonical tokens, that map is what you carry into your reporting layer and into the governed taxonomy you lock for new links. Because the audit cannot rewrite GA4 history — past hits are immutable — the map is how Facebook and facebook finally sum to one channel where you report. The deeper mechanics of why a single capital letter forks a channel are covered in UTM case sensitivity in GA4.