The short version: I put my Google Ads operating process on GitHub. It covers measurement, feeds, Search, Shopping, PMax, YouTube, and the weekly decisions that connect them. You can use it yourself or give the relevant sections to your AI agent. Start with the weekly review.
I've spent 15 years running Google Ads, and one of the most useful rules in my public playbook is to finish Monday with zero to three account changes.
Zero is a perfectly good answer. Sometimes the account needs time. Sometimes the problem is a broken product feed. Sometimes you're looking at yesterday's spend before the purchases have caught up. You can make a lot of expensive changes by skipping those questions.
I've been putting the way I work into a Google Ads agent playbook. It's public. There are campaign guides, review templates, measurement checks, and instructions for an AI agent to follow.
This is how I'd use it on an account today.
Start with what the business can afford
Before asking AI to optimize anything, give it the business context. What are you selling? What's left after the product, fulfillment, discounts, and refunds? What can you afford to pay for a new customer? Which products can you actually keep in stock?
A $100 order doesn't mean you have $100 to spend acquiring it. And a repeat purchase from someone who already knows you isn't the same acquisition result as a first order.
That context should change the recommendations. A product with strong margins and plenty of inventory might deserve more room to grow. A product that's nearly out of stock probably doesn't need an aggressive budget increase, even if its ROAS looks great.
The strategic growth principles explain how I think about product economics and campaign roles. The full management playbook puts those pieces in order.
Make the numbers explainable
I want the agent to tell me where a number came from before it tells me what to do about it.
Check which conversions the campaigns optimize toward. Check the purchase values, duplicates, reporting windows, and conversion lag. Compare the platform's story with actual orders and revenue.
If reported ROAS drops after a tracking change, lowering the budget could be a very confident response to the wrong problem. That's an example in the playbook because it's exactly the kind of reasoning I want the agent to work through.
For acquisition, I also want first-time customer data and post-purchase surveys where we have them. Ask people where they first heard about you. Look at that alongside total revenue and total ad spend. Each view helps explain something different; adding their attributed purchases together will double-count people.
The measurement guide is one of the first files I'd load. Getting this wrong makes everything after it harder.
Fix the stuff that stops a good product from selling
I've spent a lot of time on product feeds recently. We built our own feed tool, and Google, our affiliate platform, and Meta Ads now use it.
It's made me even more interested in the boring details. Is the product approved? Is the price right? Does the ad show the product the person lands on? Is a good product sitting in the catalog with no traffic?
You can spend an afternoon debating bidding strategies while your best products can't serve.
The Merchant Center and feed guide gives you a fix order. Start with blocked products, then work through classification, labels, titles, images, and page alignment. The Shopping guide covers how those products get grouped and reviewed inside campaigns.
I wouldn't split every product into its own campaign. Group products when they need different business decisions, and make sure you have enough data to support the extra complexity.
Give each campaign a clear job
A brand campaign reaching people who searched for your name should have different expectations from a video introducing you to someone for the first time.
But the campaign name doesn't prove what happened. Search can bring in new customers. Demand Gen can reach people who already know you. You still have to check the traffic and customer mix.
I use separate playbooks because the questions change:
- Search: what did people search for, and did the ad and page answer it?
- Performance Max: which products and traffic are driving the result, and how much existing demand might be in there?
- Demand Gen and YouTube: does the creative give a new buyer a reason to care?
- AI Max for Search: what are we testing, and which settings and destinations need review?
If you're doing video, there's also a more detailed creative testing workflow and a creator partnerships guide.
Decide what would make you stop before you launch
It's easy to keep a test running because you like the idea. I've found it much more useful to write down what we're trying to learn, what we can spend, and when we're making the decision.
Give the test enough delivery to have a chance. Account for the time it takes people to buy. Then respect the loss limit you agreed to. Reaching that limit might leave you with an inconclusive result. That's still a reason to stop spending.
Keep the existing winners running while you learn. Changing the bid, budget, audience, page, and creative at once makes the next review pretty useless.
Use the experiment template, bidding guide, and audience guide. For product-page and offer work, there's a separate commerce testing guide.
Have AI prepare the decisions, then review them
This is where Codex helps me a lot. Pulling data, comparing periods, finding the changes, and organizing the evidence are all things I can hand over. I still want to decide whether the explanation makes sense and whether the proposed move is worth the money.
The Monday review ends with a small number of changes, investigations, and tests. Every meaningful recommendation needs a reason. Sometimes that reason supports leaving it alone.
The daily workflow catches issues between reviews. The change-management guide covers approvals and rollback. In my own workflow, account changes go through review before execution.
You can copy the Monday memo, daily report, and change log.
Try it on one account
Start with the agent instructions and context-loading guide. Have your agent read the files for the task you're doing. You don't need to stuff every playbook into every conversation.
The repo contains instructions and templates. It doesn't connect to your account by itself. Use your existing connection or provide exports; the API setup repo is available separately.
Here's a starting prompt:
Read the Google Ads agent playbook's instructions, Monday review, measurement guide, and change-management guide. Review my last seven completed days against the previous seven using the data I provide or my authorized account connection. Ask for missing business targets. Separate established campaigns from tests and broken setups. Show me up to three decisions, the evidence behind each, and what would change your recommendation. Flag missing data. Do not make any account changes.
Read what comes back. Challenge the explanation. Ask it to show its work when something feels off. That's where your experience still matters, and it's how this gets more useful every week.