Claude for Google Ads: Prompts for Audits, Negatives, RSAs and Pacing

Claude is good at the parts of Google Ads that are reading and writing: hundreds of search terms, a page of ad copy, a month of daily spend. This guide gives you the prompts for each job, shows what comes back, and is honest about where a chat window stops being enough. For Claude across every marketing job, start with Claude for marketing.

Three ways to give Claude your Google Ads data

Exports are fine for a one-off audit. If you do the same review every week, the export becomes the job: download, upload, read, then make every change yourself in the Google Ads interface. A connector removes the first two steps, and a writing connector removes the last one.

Search terms audit

This is the highest-value prompt in the guide. The search terms report shows the queries you actually paid for, and on broad and phrase match it drifts. Pull 30 to 90 days with cost, clicks, conversions and the matched keyword, then ask:

Here is my Google Ads search terms report for the last 60 days.
Our business: [one line on what you sell and to whom].
Our target CPA: $[X].

1. List every search term that spent more than 2x our target CPA with zero conversions.
2. Group them into themes (competitor names, jobs/careers, free/cheap, wrong product, informational).
3. For each theme, say whether it should be a negative and at which match type.
4. List search terms that converted below target CPA but are not yet keywords.
Output two tables, then a one-paragraph summary of total wasted spend.

What comes back is a waste table and an opportunity table. A worked example from a B2B software account:

Check one thing before you accept the list: Claude only sees the terms you gave it. If the export was filtered or truncated, the waste figure is a floor, not a total.

Negative keyword lists

The risk with negatives is blocking something you want. Give Claude your live keywords as well as the proposed negatives, and ask it to check for collisions:

Here are my active keywords (with match types) and a proposed negative list.
For each negative:
- Flag it if it would block any active keyword or any search term that converted.
- Recommend campaign-level or ad-group-level, and say why.
- Suggest the narrowest match type that still stops the waste.
Return the cleaned list ready to paste into a shared negative keyword list.

Build a starter negative keyword list for a [business type] that sells
[product] at [price point]. Cover: job seekers, students, DIY intent,
free-seekers, competitor brand terms we do not want, and wrong-industry
uses of our main keywords. 60 terms max, grouped, with match types.

The second prompt is for new accounts that have no search terms yet. Treat its output as a draft: Claude knows how searches usually drift, not how your account drifts.

Responsive search ad copy

Google's help center sets the limits: up to 15 headlines and 4 descriptions per responsive search ad, headlines up to 30 characters, descriptions up to 90 and each path field up to 15. Put those in the prompt, because Claude counts characters reliably only when told to check.

Write a responsive search ad for [product] targeting the keyword "[keyword]".
Landing page copy is pasted below.
- 15 headlines, each 30 characters or fewer. Show the character count after each.
- 4 descriptions, each 90 characters or fewer, with counts.
- 2 path fields, 15 characters or fewer.
- Mix: 5 headlines on the core benefit, 4 on proof (numbers from the page only),
  3 on the offer, 3 with the keyword.
- Do not invent statistics, prices or awards that are not on the page.

A sample of what that returns for a scheduling tool:

• Book Meetings Without Email (27)
• Scheduling Links in 2 Minutes (29)
• Works With Google Calendar (26)
• Stop the Back and Forth Today (29)

Then ask Claude to rewrite the weakest three against your current best performer. If you can export asset performance labels, include them: headlines rated Low are the ones to replace first.

Here are my current RSA assets with Google's performance labels.
Replace every asset rated Low with two alternatives in the style of the
assets rated Best. Keep the character limits. Explain each swap in one line.

Reading Performance Max

Performance Max shows less than Search, so the prompt has to ask Claude to separate what the data shows from what it suggests. Export asset group performance, the asset report and the campaign's conversion data, then ask:

This is a Performance Max campaign's asset group and asset data for 30 days.
1. Rank asset groups by conversions and by cost per conversion.
2. List assets with the lowest performance labels and what they have in common.
3. Say plainly which conclusions the data supports and which are guesses.
4. Propose two new asset groups, each with 5 headlines and 2 descriptions,
   built around the themes that already convert.

Claude will usually flag the honest limit itself: Performance Max reports do not break cost down the way a Search campaign does, so a claim like "this asset drove those sales" is a guess. If you are weighing whether a campaign belongs in Performance Max at all, read Performance Max vs Search.

Budget pacing

Pacing is arithmetic that people get wrong on the 18th of the month. Give Claude the monthly budget, the days elapsed and spend to date by campaign:

Monthly budget: $3,000. Today is day 18 of 30.
Spend to date by campaign is below, with daily budgets.
For each campaign and the account total:
- Expected spend to date at even pacing, actual spend, and pace as a %.
- The daily budget needed to land exactly on the monthly figure.
- Flag anything pacing above 110% or below 90%.

The account-level answer for that example: even pacing says $1,800 by day 18. If you have spent $2,100, you are at 116.7% of pace. The remaining $900 over 12 days means about $75 a day, against the roughly $117 a day you have been averaging. Claude does that per campaign in one pass and tells you which ones to cut.

Impression share and Quality Score

Here is campaign data with Search impression share, IS lost to budget
and IS lost to rank, plus keyword Quality Score components.
Tell me which campaigns are limited by money and which by quality.
For quality-limited ones, name the component dragging the score
(expected CTR, ad relevance or landing page experience) and the fix.

This separates two problems that look the same in a weekly report. Budget-limited campaigns need money or tighter targeting; rank-limited campaigns need better ads or pages. The Quality Score guide covers what each component means.

Where Claude stops

• It cannot see what you did not give it. On exports, every answer is bounded by the file. Claude will not know about the campaign you forgot to include.
• It can be confidently wrong about your account. Ask it to quote the rows behind every number, and spot-check two.
• Read-only access does not change anything. Google's own server and every export end with you making the edits by hand.
• It does not know your margins. A target CPA or ROAS has to come from you. Without one, "wasted spend" is a guess.

For a structured full-account review, use the checklist in the Google Ads audit guide and run the prompts above against each section. The same approach works on Facebook and Instagram, covered in Claude for Meta ads, and a longer prompt library lives in Claude marketing prompts.

Where Sprites fits

Sprites is an official MCP connector you add to Claude by URL. Once connected, the prompts above run against the live account instead of an export: Sprites reads search terms, assets, impression share and change history itself. When the answer is a change, such as campaign-level negatives, a budget rebalance across several campaigns, a new Search campaign or new Performance Max asset groups, it arrives as an approval card in the chat. Negatives are checked against live keywords before you see the card, and budget changes are re-read after they apply to confirm they took.

Setup takes a minute and is described on the Google Ads and Claude setup page. Plans start at $99 a month, with a 7-day trial for $1.

Claude for Google Ads questions

Can Claude manage Google Ads?

Claude can analyze Google Ads data and write copy, negatives and plans on its own. To change the account it needs a connector with write access. Google's own Google Ads MCP server is read-only, so changes need a connector like Sprites, which shows each change as an approval card first.

Is there an official Google Ads MCP server for Claude?

Yes. Google publishes a Google Ads MCP server, documented on its Google Ads API developer site. It is read-only, runs Google Ads Query Language queries, and you host it yourself with a developer token and Google Cloud credentials.

What are the best Claude prompts for Google Ads?

The highest-value prompts are a search terms audit against your target CPA, a negative keyword collision check, responsive search ad copy with character counts, and a budget pacing calculation. Each works best with a real export or a live connection and a stated target CPA or ROAS.

Can Claude write responsive search ads within the character limits?

Yes, if the prompt states the limits and asks for a character count after each line. Google allows up to 15 headlines of 30 characters and 4 descriptions of 90 characters per responsive search ad. Ask Claude to show counts so you can check them before pasting.

Can Claude find negative keywords?

Yes. Give it a search terms report and your target CPA and it will list terms that spent with no conversions, grouped by theme, with suggested match types. Ask it to check the list against your active keywords so a negative does not block traffic you want.

Is Claude good for PPC?

Claude is strong at the reading and writing parts of PPC: audits, copy, pacing math and reporting. It needs your data and your targets to be useful, and it needs a writing connector to act on what it finds instead of handing you a to-do list.