How to Find Wasted Ad Spend with Claude (Step by Step)

Why this is the first thing to try

Search-term cleanup is tedious by hand and trivial for a model. A month of data on an active account can run to thousands of unique queries; skimming them for junk is exactly the sort of work people postpone. A model reads the whole table in one pass and ranks it by cost.

It is also the safest place to start. Negative keywords are reversible, cheap to test, and the downside of a mistake is small — unlike pausing a campaign or moving a budget.

Setup (about two minutes)

1. Connect Google Ads and Meta with each platform's own sign-in — no passwords shared.
2. In claude.ai: Settings → Connectors → Add custom connector, paste the connector URL, and approve the Connect prompt.

Full walkthrough: how to have Claude run your Google Ads.

The prompts that work

Start broad, then narrow:

• "Which of my campaigns spent money last month with zero conversions?"
• "Show me the 20 most expensive search terms with no conversions in the last 30 days."
• "Which search terms look like the wrong intent for what I sell?"
• "Compare cost per conversion across my campaigns and flag the outliers."
• "Draft campaign-level negative keywords from the junk you found, grouped by why."

That last one matters: asking for the reasoning grouped alongside the terms makes the list far easier to review, and it exposes any term the model has misjudged.

What the junk usually looks like

• Competitor and unrelated brand names. Broad match picks these up constantly. Some conquest traffic is deliberate — check before cutting.
• Wrong intent. Someone searching for a cheaper adjacent product, or a job, or a how-to guide, when you sell software.
• Free-seekers. Queries containing "free", "template", "no sign up" against a paid product.
• Category drift. Terms adjacent to your category but for a different job entirely — the most expensive kind, because they look relevant.

The mistake to avoid

Check which conversion column the numbers use before you cut anything. Many accounts have a generic "Conversions" figure that counts actions you do not actually optimise against — and a "zero conversions" verdict built on the wrong column will tell you to pause something that is quietly producing your best leads.

This bites hardest on campaigns whose real payoff is offline or delayed — demo requests, sales calls, trials that convert weeks later. Ask the model to state which metric it used, and sanity-check one campaign in the platform UI before acting on the rest.

Then let it make the change

With write actions enabled, Claude can add the negatives directly. Insist that it runs as a dry run you approve, that budgets are capped, and that everything is audit-logged. Start with the no-regrets terms — obvious wrong-intent junk — and leave judgement calls, like conquest campaigns, for a human decision.