# Monitoring With an LLM
Course: [Google Ads for B2B Teams](https://www.devtutorials.ai/learn/google-ads) · Section: Run and Grow
Source: https://www.devtutorials.ai/learn/google-ads/run-and-grow/monitoring-with-an-llm
> Weekly review workflows with Claude or ChatGPT: exported reports, anomaly checks and search-term cleanup, and where an LLM falls short.
## How often to look
Starter: every two days in the first week after launch (search terms only), then weekly. A monthly budget decision is enough. Looking daily at small numbers mostly produces worry and unnecessary changes.
Intermediate: a weekly review of every campaign, a quick daily glance at spend and conversions for anomalies (5 minutes, or an automated alert), and a monthly budget and region review.
Advanced:daily anomaly checks (automated rules or scripts emailing you when spend or conversions move more than a set percentage), a weekly review per campaign group, and a monthly review of budgets, regions and pipeline from the CRM.
Set up two simple safety nets in Google Ads regardless of budget: an automated rule or email alert when daily cost exceeds about 1.5× the daily budget average, and one when conversions are zero for two days on a campaign that normally converts daily.
Broken tracking is the most expensive silent failure.
## The weekly review
Use the weekly review sheet (CSV) to record each week’s numbers and decisions. Open it in Google Sheets or Excel and add a row per week.
Export (last 7 days and the 7 before, or last 28 days on a small budget):
Search terms report: search term, match type, campaign, ad group, cost, clicks, conversions.
Ads and assets report (monthly is enough): ad, headlines, CTR, conversions.
Matched locations report (monthly, see Geo targeting).
From the CRM: leads from paid search with campaign, stage and date.
Run the prompts below in your Project.
Decide on at most three changes. Write them in the review sheet with the date.
Apply them in Google Ads, and add a note in the account (Notes are available on the performance graph) so later you know why a line moved.
Download the reports from the Google Ads interface: open the table, use the download icon, choose CSV. If you run the same export every week, save it as a report in Reports and schedule it to email you. As of October 2026
## Set up your LLM workspace once
Create a Project in Claude called “Google Ads”. In the project knowledge, upload a short account facts document (below) and your negative keyword list. Paste the standing instructions into the Project instructions.
Then each week, start a new chat inside the Project and attach that week’s exports. Claude reads the CSV files with code execution, which is more reliable than reading numbers from pasted text.
Ask for charts or tables as an Artifact when you want something to share.
Create a Project in ChatGPT called “Google Ads”, upload a short account facts document and your negative keyword list as project files, and paste the standing instructions into the Project instructions.
Each week, start a new chat in the Project and attach the exports; ChatGPT’s data analysis reads CSVs with code, which is more reliable than pasting text.
If several colleagues run the same review, a Custom GPT with the same instructions and files lets everyone use one setup. Use Canvas when you want to edit a summary or ad copy together.
## Review prompts
### Spot anomalies
### Find negatives in the search terms report
Watch the word “free”
“Free” is the classic trap. As a negative it blocks “free trial”, which may be your best search on a trial-goal campaign.
Use more specific negatives such as "free download", "free version" or "open source" instead, and only if they don’t describe your offer.
### Ad copy performance
### Lead quality from the CRM
## Where LLMs go wrong
Arithmetic and reading tables. Models make mistakes adding up columns or match the wrong row, especially from pasted text. Attach CSVs, ask it to calculate with code, and spot-check two numbers per review.
Made-up metrics and causes. If a column isn’t in the file, the model may still produce a number or a confident explanation (“competitor X raised bids”). The standing instructions tell it not to; still, ask “which column did that come from?” when something surprises you.
No live access. It sees only what you upload. It doesn’t know about the change your colleague made yesterday, a policy disapproval, or Google’s latest interface. Check settings in the account itself.
Outdated product knowledge. Google Ads changes often. If the model describes a menu or feature, confirm it in Google’s help pages before acting.
Data privacy. Use a business plan your company has approved, and check its data retention and training settings. Remove personal data from CRM exports. Don’t paste credentials or customer lists.
Over-eager changes. Models like to recommend lots of changes. Your rule is three per week, and you decide.
Go deeper Going further: scripts and scheduled exports
Once the weekly routine is steady, you can automate the export step. Google Ads scheduled reports can email CSVs on a set day.
Google Ads Scripts (JavaScript that runs inside your account) can write daily data into a Google Sheet or email you when something breaks.
Your LLM can write these scripts:
Describe exactly what you want (“every morning, email me campaigns whose cost yesterday was more than 150% of their 14-day average”).
Paste the result into Tools → Bulk actions → Scripts.
Preview it.
Read what it will do before scheduling.
Start with read-only scripts that only report; don’t let a generated script change bids or pause campaigns until you’ve run it in preview for a couple of weeks.
Do this now
Create the LLM Project with standing instructions and the account facts document.
Download the weekly review sheet and set a weekly calendar slot.
Set up the two safety alerts (spend spike, zero conversions).
Run the first review: anomalies and search terms prompts; add the negatives you verified.
Strip personal data from a CRM export and run the lead-quality join once.