How to Automate Google Ads With AI Agents
By Dan Ahmadi, Co-Founder at Upside ·
What if your Google Ads account could optimize itself every morning before you open your laptop? Dan Ahmadi built exactly that. In this GTM Clawllective session, he walked through APEX, an open-source multi-agent system that manages paid search on autopilot using Claude, MCPs, and first-party GTM data. Here is everything he shared.
What Is APEX?
APEX stands for Autonomous Paid Execution. It is a multi-agent AI system that runs Google Ads end-to-end: keyword optimization, ad creation, landing page generation, and experiment tracking, all without requiring a human to log in to the platform. It runs one to two times daily.
APEX is not a Google Ads script or a native platform feature. It sits on top of Google Ads and connects it to the rest of your GTM stack, giving it business context that Google's own AI does not have. Dan has open-sourced the full skill kit at upside.tech/skills/apex so anyone can run their own version.
The Architecture: Four Data Sources, Seven Agents
What data does APEX use?
The system is powered by four MCPs (Model Context Protocols) that feed it real-time context:
- Notion MCP — brand voice, personas, target audience definitions, and product messaging
- EXA MCP — live web search for competitive and market research
- Upside MCP — GTM data pulled from Salesforce, HubSpot, Marketo, and call transcripts
- Google Ads MCP — read and write access to the Google Ads account
One orchestrator agent manages seven specialized sub-agents. A "commander's intent" document defines the program's overall goals, and a strategy log acts as the system's memory, recording every decision and the rationale behind it so the system builds institutional knowledge over time.
The system uses two Claude models with distinct roles: Claude Opus 4.7 handles the heavy lifting (orchestration, writing responsive search ads, landing page markdown, and research), while Claude Sonnet 4.6 runs the daily search-term classification where speed and cost efficiency at scale matter more than depth.
What does each of the seven agents do?
| Agent | Function |
|---|---|
| Keyword Optimizer | Pulls the last 24 hours of impression keywords, analyzes each in parallel, and automatically adds negative keywords or creates new targets |
| Ad Optimizer | Identifies underperforming ads, auto-pauses low CTR ads, and queues new RSA challengers |
| Ad Creator | Reads landing pages, generates 15 headlines and 4 descriptions, and requests human review when uncertain |
| Landing Page Prioritizer | Ranks which landing pages to build next by ROI and outputs exact H1 recommendations |
| Landing Page Researcher | Takes a page concept, researches competitors and messaging, and creates a comprehensive brief |
| Landing Page Creator | Converts the research brief into a structured markdown landing page and opens a GitHub pull request |
| Experiment Manager | Tracks multi-week optimization tests, decides rollouts, and updates the strategy log when experiments conclude |
Technical Implementation
What infrastructure does APEX run on?
Dan migrated from Webflow to GitHub and Cloudflare for the landing page layer. This enables faster iteration, custom experiences, and AI-generated pages. All pages are automatically tracked through existing Google Tag Manager with no additional setup.
For the AI harness, Dan uses Craft Agents: a free, open-source tool that runs locally and can connect to any API or service. For Google Ads API access, an MCC account is required and edit permissions need a one to two day approval from Google.
Landing pages follow a structured template with dynamic variables. A custom endpoint exposes all available pages to Claude so the AI understands what already exists before creating anything new.
Does It Actually Work? The Results
These are real numbers from APEX running on Upside's own Google Ads account:
| Metric | Result |
|---|---|
| Junk-click rate | Dropped from 81% to 0% over 7 weeks |
| Wasted-spend share | Reduced from 48% to 27% (14-day rolling average) |
| Cost per qualified click | $14, held steady through an 8x keyword expansion |
| Junk keyword variants permanently blocked | 587 (zero at launch) |
| Runs unattended | ~21 hours between optimizer runs |
| Claude usage cost | ~$100/month |
| Time per optimization run | 5 to 10 minutes |
| Scalability | One operator running a similar setup at $300K/month ad spend |
Why Is This Better Than Google's Native AI?
- Google only sees Google Ads data. APEX also reads call transcripts, CRM data, and messaging docs, giving it full revenue context for every decision.
- Google's AI is designed to increase spend. APEX uses a more conservative approach aligned with actual business goals.
- Custom business rules are built in. Naming conventions, campaign structures, and strategic guardrails are embedded in the system, not approximated by a black-box model.
Key Takeaways for GTM Teams
You do not need to be an engineer to build this. Dan built APEX as a marketer with deep operational instincts. Craft Agents, MCPs, and Claude lower the barrier significantly.
Business context is your moat. The more first-party context you feed the system (call transcripts, CRM signals, messaging docs), the better the output compared to platform-native AI that only sees its own data.
The strategy log is the secret weapon. Logging every decision and rationale creates institutional memory so the system improves over time rather than resetting with every run.
Start at any scale. The architecture works at $100/day or $300K/month. There is no minimum viable budget to get started.
Want to Attend the Next Session?
GTM Clawllective is a community for GTM practitioners sharing what is actually working. Sessions are not recorded, so this recap is the closest thing to being there. If you want in on the next one, join the community.
About the Speaker
Dan Ahmadi is the Co-Founder of Upside, a B2B forensic revenue intelligence platform that helps teams selling big-ticket items understand what influenced each deal and what to do next to win it efficiently. Before founding Upside, Dan spent 12 to 14 years in demand generation, managing ad budgets as large as $20 to $25 million per year at companies like MuleSoft and Branch. He is a graduate of UC Davis with a degree in Biochemistry and Technology Management. At Chili Palooza, he won an award for his two-minute demo of APEX.
You can follow Dan on LinkedIn or explore the APEX skill kit at upside.tech/skills/apex.