Google Ads Optimization With Claude: A Step-by-Step Tutorial
The exact prompts, exports, and actions I use to optimize a multi-geo Google Ads account with Claude — illustrated with a fictional SaaS scenario you can adapt to your own campaigns.
I wrote The AI Blindspot after an AI nearly convinced me to pause keywords that were the literal definition of my product. That article was the warning. This one is the playbook.
Below are the seven core optimization actions I run with Claude, each with the exact data you need to export, the prompts that work, what to watch out for, and examples from a fictional SaaS called NovaCRM, a small-business CRM tool running Search, PMax, and Brand campaigns across the US, EU, MENA, and Asia-Pacific. The scenarios are made up. The methods are exactly what I use on real accounts.
If you manage Google Ads and have access to Claude, you can start using this today.
Before You Start: The Context Prompt
Every session starts with the same message. Skip this, and the AI will give you generic advice.
“I’m managing Google Ads for NovaCRM, a CRM SaaS for small businesses. Target CPA: $40. Customer LTV: $600. We run Search, PMax, and Brand campaigns across US, EU, MENA, and APAC. I’ll upload data campaign by campaign. For every recommendation, explain your reasoning. Do not recommend pausing keywords that describe our core product category, even if they show zero conversions in this window.”
That last sentence is the single most important guardrail. Without it, Claude will flag your category keywords as waste, because it doesn’t know what your product is.
1. Search Terms Analysis & Negative Keywords
This is the highest-ROI action you can do with Claude. A single session can save thousands in wasted spend.
What to export
Google Ads → Campaigns → select your Search campaign → Search terms. Download the report as CSV. Include: Search term, Match type, Clicks, Impressions, Cost, Conversions, Conv. value.
The prompt
“Here’s the search terms report for [campaign name]. Go through every search term and categorize them:
1. NEGATIVE — clearly irrelevant (wrong intent, wrong language, competitor names, template/education seekers, unrelated industries)
2. REVIEW — ambiguous, could go either way
3. KEEP — relevant to our product
What you’ll get
Claude will process hundreds of rows and return a structured list. For NovaCRM’s MENA Search campaign, imagine a single pass producing 100+ negative keywords grouped into clusters like: student/homework queries, free spreadsheet seekers (“free CRM template Excel”), competitor brand names (“HubSpot login,” “Salesforce pricing”), and queries in languages you’re not targeting (French, Urdu queries triggered by broad match in a region with multiple languages).
What to watch out for
Claude will sometimes flag category keywords as negatives. For NovaCRM, it might recommend pausing “small business CRM” or “customer management software” because they had zero conversions in a 14-day window. Those are literally what your product does. A 14-day window isn’t enough to judge category keywords; they need time and volume. This is why the guardrail in the context prompt matters.
After it gives you the list, always run this follow-up or manually handle them (I suggest you do it manually):
“Look at the NEGATIVE list again. Which of these could potentially block a real customer searching for our product? Flag anything you’re less than 90% confident about.”
How to apply
Go to Google Ads → Campaign → Negative keywords → add the list. I recommend creating a shared negative keyword list if you’re running the same product across multiple campaigns, a “CRM competitors” list, a “free template seekers” list, etc.
2. Keyword Cleanup & Pause Decisions
Different from negatives. This is about your own keywords, the ones you’re bidding on.
What to export
Google Ads → Campaign → Ad groups → Keywords. Download with: Keyword, Match type, Status, Clicks, Impressions, Cost, Conversions, Conv. value, Quality Score, Search impressions. share.
The prompt
“Here’s the keyword report for [campaign name]. Analyze each keyword and flag:
1. PAUSE candidates — keywords with significant spend and zero conversions that don’t describe our core product (e.g., “CRM spreadsheet template” attracts template seekers, not SaaS buyers)
2. LOW QS — keywords with Quality Score below 5, with a note on whether the issue is ad relevance, landing page, or expected CTR
3. MATCH TYPE issues — keywords where the match type might be too broad or too narrow for the campaign strategy
Do NOT flag category-defining keywords (like ‘CRM software’, ‘customer management tool’, ‘sales pipeline app’) as pause candidates regardless of their short-term performance.”
Example
Say Claude flags “CRM spreadsheet template” as a pause candidate. You check the search terms, and 90%+ of queries triggered by this keyword are “free CRM template download,” “customer list spreadsheet free,” etc. The intent is wrong; these people want a free spreadsheet, not a SaaS subscription. Pause it.
But then Claude also suggests pausing “customer management software” due to high spend and low conversion rates. That’s a core category keyword; it is your product. Push back. Ask Claude to explain why. If it immediately reverses course, the recommendation wasn’t grounded. That’s your signal that it was pattern-matching on surface data, not reasoning from business logic.
Keyword cleanup checklist
“For the keywords flagged as PAUSE, confirm:
(1) Is the intent behind this keyword misaligned with our product?
(2) Can we fix this with negatives instead of pausing?
(3) Would pausing this keyword remove us from searches where our product is a legitimate answer?”
If the answer to #3 is yes, don’t pause; add negatives to filter the bad traffic instead.
3. Campaign Structure: Language & Geo Segmentation
This is where Claude’s processing speed matters most. Restructuring campaigns manually across languages and geographies is tedious. With Claude, you can plan and validate the structure in minutes.
When to restructure
Your campaign has mixed-language keywords with different landing pages. Performance varies wildly by country within a single campaign. PMax asset groups are serving in languages that don’t match the creative.
The prompt
“Here’s the performance breakdown by country and language for [campaign name]. Current structure: [describe it]. I’m seeing [specific problem — e.g., ‘Country X gets 40% of spend but 0% of conversions’]. Recommend a restructure. For each new campaign you suggest, specify: countries included, language, daily budget, which assets/keywords move there, and which countries to exclude from the original.”
Examples
MENA Search → Three-way split. Imagine NovaCRM’s MENA campaign has Arabic and English keywords mixed, targeting the entire region. Session 1 separates English into its own campaign. Session 2 goes further: three underperforming countries get excluded from the main campaign. One of them (say UAE) has enough volume to justify its own dedicated campaign with its own budget. The other two are simply cut — not enough data to justify standalone campaigns.
Regional Search → Language split. A campaign targeting Saudi Arabia has Arabic and English queries coming in. Split into two ad groups: Arabic Ads and English Ads. Each gets its own asset set, its own negative keyword list, and its own landing page URL. This is critical; mixed-language campaigns confuse Google’s algorithm because it can’t optimize bidding when one campaign serves two languages with two different landing pages.
Get Alara Akcasiz’s stories in your inbox
Join Medium for free to get updates from this writer.
EU PMax → Country extraction. One country is consuming a disproportionate share of the EU PMax budget. Extract it into its own campaign so the remaining EU countries get fair budget distribution. If another country is showing geography-mismatch symptoms (high spend, low relevance), exclude it entirely.
The rule of thumb
One language per campaign. One region logic per campaign. If a country is big enough to eat the budget, give it its own campaign. If a country is too small to generate meaningful data, exclude it and re-evaluate later.
4. Auction Insights & Impression Share Analysis
This is where most marketers stop using AI. They shouldn’t.
What to export
Google Ads → Campaign → Auction insights. Also: Columns → Modify columns → Competitive metrics → add Search impr. share, Search top IS, Search abs. top IS, Search lost IS (rank), Search lost IS (budget).
The prompt
“Here’s the auction insights and impression share data for [campaign name]. I’m sharing Search impr. , lost IS (rank) and, lost IS (budget) reports. Top competitors are [list from auction insights]. Diagnose the problem. Is this a bid issue, a quality issue, or a budget issue? What’s the most efficient lever to pull?”
How to interpret (and what Claude tends to get wrong)
Lost IS (budget) means your daily budget is exhausted before the day ends. The fix is either raising the budget or narrowing targeting, not raising bids.
Lost IS (rank) means your Ad Rank is too low. This could be a bid issue, but it could also be a Quality Score issue (landing page, ad relevance, expected CTR). Claude will often jump to “raise your bids.” Push back (but I never had to push back):
“Before recommending bid increases, check the Quality Score data. If QS is below 6, the issue is likely ad relevance or landing page, not bid level. What else could improve rank without increasing spend?”
5. Asset & Headline Optimization
What to export
Google Ads → Campaign → Assets → Asset details. Look at the Performance column (Low, Good, Best) and the combinations report.
The prompt
“Here are the asset performance results for [campaign name], and asset performance report. Our product positioning is: [one sentence]. Our target audience is: [describe].
Recommend:
(1) which headlines to remove and why,
(2) replacement headlines that match our positioning,
(3) any description changes.”
Example: PMax headline cleanup
Say NovaCRM’s PMax campaign has these headlines:
- “First Month Free” — attracts freebie seekers who churn immediately
- “Start Your Free Trial Today” — too generic, every SaaS says this
You’d remove all three and replace them with benefit-driven, positioning-aligned headlines:
- Close Deals 2x Faster
- Built for Small Sales Teams
- From Lead to Deal in Minutes
- No Training Needed
- Your Pipeline, Your Way
- Revenue-Ready in One Day
- Smart CRM for Growing Teams
Why does the human make this call, not the AI
Claude can generate 50 headlines in 30 seconds. But it doesn’t know which positioning attracts the wrong user. “First Month Free” looks fine to an AI — it’s a clear value proposition. To a marketer who knows that freebie-seekers never convert to paid, it’s a budget drain.
The prompt I use for headline generation:
“Generate 15 headline options (max 30 characters each) for a Google Ads. The product is [X]. Target user: [Y]. Positioning: [Z]. Avoid: free trial language, competitor comparisons, generic CTAs. Focus on outcomes and use cases.”
Then I cherry-pick. Claude drafts, the human selects.
The Prompts That Saved Me
These are the non-obvious prompts — the ones that prevented expensive mistakes.
The guardrail prompt (use at the start of every session):
“Do not recommend pausing keywords that describe our core product category, even if they have zero conversions.”
The uncertainty prompt (use after any recommendation):
“Which of these recommendations are you least confident about, and why?”
The reversal test (use when something feels off):
Push back on the recommendation. If Claude immediately reverses course, the recommendation wasn’t strong.
The structural check (use before any campaign change):
“Is this campaign currently in a learning phase? If I make this change, will it reset the algorithm’s learning?”
The documentation prompt (use at the end of every session):
“Give me a structured log of every change we made, organized by campaign. For each change: what we did, why, what metric to watch, and when to re-evaluate.”
What Claude Handles vs. What You Handle
Claude handles: Scanning hundreds of search terms and categorizing them. Spotting URL mismatches between keywords and landing pages. Generating headline options from a positioning brief. Calculating CPA/ROAS scenarios. Structuring and documenting your decisions. Comparing performance across segments.
You handle: Deciding whether a keyword is “core to your product” or a waste. Knowing if the account is in a learning phase. Understanding why a specific audience segment doesn’t convert (language mismatch? wrong positioning? market readiness?). Making the kill/keep call. Choosing the right bid strategy. Setting the budget. Reviewing the AI’s work before applying anything.
Claude processes, you decide.
This is Part 2 of a series. Part 1: The AI Blindspot: What Happens When No One Is Driving — where I learned these lessons the hard way.
