How to get Claude to actually understand a Google Ads strategy

A screenshot of the campaigns tab is where bad advice starts. Our audits pull 34 datasets for every account, plus more for lead generation or ecommerce. The ones people leave out are the ones that matter:

• Search terms, not just keywords. Keywords are what you bid on. Search terms are what people typed.
• Conversion actions and goals, so it knows what counts as a conversion before it reads one.
• Negatives at all three levels: ad group, campaign and shared lists. Read only one and it cannot see a term blocked in one place and targeted in another. We have seen exactly that.
• Budgets, including whether they are shared. Google will not run Maximize Conversions, Target CPA or Target ROAS on a shared budget. A bidding recommendation written without that flag may be impossible to carry out.
• Change history. Google keeps 30 days, and every change says who or what made it. That is how you find the change nobody on the team made.
• Policy status. A disapproved keyword looks exactly like a keyword nobody searches.

Claude cannot see your strategy in the account. It will guess it from campaign names, and it will guess wrong. Write it down:

• Is this business lead generation, ecommerce, or both?
• What is each campaign for, and the one number that judges it?
• What demand must be protected, such as brand searches or a top service?
• What is a lead or a sale worth?
• What have you already decided? A campaign you paused on purpose is a decision, not a problem to fix.

That last one saves the most time. Without it, Claude finds the paused campaign every week and tells you to turn it back on.

Most wrong conclusions come from reading a number correctly and meaning it wrongly. We keep a metric dictionary, one card per metric, in this shape:

Others worth a card: impression share lost to budget versus rank, conversion lag, Performance Max’s channel split, and why platform ROAS is not profit.

“If CTR is low, rewrite the ads” is a rule, and rules are how you get confident wrong answers. A pattern names the signal, the causes in order, and the one read that tells them apart:

One pattern saves more money than any other: the last few days always look bad. Conversions are filed under the day of the click and keep arriving for weeks. A drop in the last three days is usually not a drop. Tell Claude never to change bids on it.

We learned this one the hard way. One account had three audits in eight days. One checked coverage, one found a $40 search term shift, and one wrote two thousand words on Quality Score. None of them asked where the leads went.

Now every audit starts with five moves, in order:

1. Leads, spend and cost per lead by month, for twelve months. When did it work?
2. The same table by campaign. Which one thing moved?
3. Inside that, what actually converted, and what everything else spent.
4. What changed: change history, statuses, the landing page today, the tracking.
5. Read the ads that converted next to the ones that did not. In that account, the ad that led with the buyer’s outcome converted at 10%. The rest did not.

An audit that hides what it skipped reads as complete when it is not. Ask for a coverage line in every answer: each dataset marked complete, partial, blocked, empty or not pulled. “Not pulled” and “empty” are different findings, and missing data is never a clean result.

Every readout comes back in the same order, even a short one:

1. The verdict, in one line.
2. What the account looks like, with the coverage line.
3. Where it is wasting money.
4. Where it can grow.
5. What to do next, and how you would know it worked.

The last part is the one people drop. A recommendation nobody can check later is not finished.

Let Claude read freely and change nothing on its own. When it wants a change, have it write the before-state, the exact new values, the spend at risk, what would prove it wrong, and when to roll it back. A person approves it. It makes one attempt and reads the result back.

Never let it retry blindly. If the result is unclear, check the account. A blind retry applies the change twice.

Claude writes better ads with a structure than with a brief. Ours for responsive search ads, with two or three pinned headlines per position:

• Headline 1: the customer’s outcome or feeling. Never the company name.
• Headline 2: the offer, and what makes it different.
• Headline 3: one plain call to action.
• Description 1 expands headline 1. Description 2 expands the offer and the ask.

Yours can be different. The point is that it is written down.

Put the above in a Claude skill so you never paste it again. Save this as SKILL.md in a folder under .claude/skills/ (Claude Code) or zip it and upload it in Claude’s settings. Then fill in your own strategy and dictionary.

---
name: google-ads-analyst
description: Investigate one Google Ads account and recommend changes. Use for any
  question about Google Ads performance, search terms, bidding, budgets, tracking
  or ads. Read-only: never makes a live change.
---

# Google Ads analyst

## Before judging anything
- Read which conversion actions are primary. Say what "conversions" contains.
- Start with the culprit walk: 12 months by month, which campaign moved, what
  converted inside it, what changed, then read the ads side by side.

## The strategy (fill this in)
- Model: lead generation | ecommerce | both
- Each campaign, its job, and the one number that judges it:
- Protected demand:
- What a lead or sale is worth:
- Decisions already made (do not "fix" these):

## Never
- Change bids or budgets because the last few days look bad.
- Infer the strategy from campaign names.
- Treat platform ROAS as profit.

## Every answer, in this order
1. Verdict. 2. What it looks like + coverage line. 3. Where it wastes.
4. Where it can grow. 5. What next, and how we would know it worked.

## References
- metrics.md: one card per metric (means / adds up? / watch for)
- patterns.md: signal / causes in order / the read that separates them

One lesson on the reference files: write one idea per heading, in the words you would type when asking about it. A vague heading loses to whatever the file’s introduction says.