An Engine I Check Once a Week

I have run search advertising at the scale where it stops being a channel and becomes a P&L line. At a multibillion-dollar corporation I owned half a million dollars of search spend, returned four times that in web sales, and built the team of four that carried it after me. I know what a well-run ads operation looks like, and I know what it costs in people to run one.

The interesting problem is running it without the people. One operator, a dozen competing priorities, and a channel that punishes neglect. The answer is not working harder in the interface, and it is not hiring an agency to do the clicking for you. The answer is process: an engine designed so that a week of attention compresses into one deliberate look, without giving up the control that tells you the money is well spent.

I got there the long way, because this is also the story of how I started trusting AI again after it had cost me.

What I inherited

The Google Ads account was four or five years old and several agencies had been through it. The copy tried to be all things to all people, which was precisely the opposite of where our new CEO wanted the company positioned. But the account was producing leads, I was new to the market, and there was a lot to learn before touching a machine that worked. So I left it alone and watched.

That watching was less rigorous than I would have told you at the time. It was glancing. And in mid-December we made a change on the website that cut inbound demo requests hard, for a reason we had not foreseen. It took until January for me to see how far the numbers had actually fallen, and by then the leads from paid had thinned to almost nothing. I had owned that account for months.

This was also the winter after I had broken our CRM by trusting a model’s confident wrong answers. I was off AI for anything that mattered. So the project that followed started from an uncomfortable double position: a channel I had neglected, and a set of tools I no longer trusted.

What I tried

I went in to rebuild the account around the new positioning: new copy, new structure, a strategy with an actual point of view instead of a message for everyone. And I did what a competent operator does. I pulled performance apart, ran analysis through AI to find what was working, killed the weak spots, refreshed the creative.

It moved the edges. Single-digit improvements on a channel that had structurally slowed. Meanwhile the real cost was invisible on any dashboard: my attention. Every hour inside the ads interface was an hour not spent on the community, the content engine, the operations stack, or the strategy work only I could do. Optimizing harder was the wrong shape of answer. I needed the channel to take big steps, and then I needed it to mostly run without me.

The tools were half there, which was the frustrating part. Google had shipped an MCP connector that could pull anything out of the account, and it genuinely sharpened my view of what was happening. Semrush connected to Claude covered the keyword research. Claude wrote strong ads once it had the context. So I could watch the account in real time, research in minutes what used to take a morning, and draft a full refresh in an afternoon.

And then every one of those threads dead-ended at the same wall. Analysis, research, and copy all landed in my lap as things a human now had to type into the Google Ads interface by hand. The intelligence was automated. The action was not. The seam between them was me.

So I built the missing piece: a write connector into Google Ads, so the same system that reads the account can change it. And because I was carrying a fresh scar, the constraints went in first, not as an afterthought. Everything the connector creates arrives paused. Every change runs as a dry run before it can commit. It acts on an explicitly named account or it fails loudly. Every write is logged. I had learned, expensively, what happens when you give automation a live system on faith, and this time the guardrails were the foundation the tool was built on.

The failure in this story has nothing to do with AI. I missed a collapse in my own channel for weeks because I was glancing at the account instead of tracking it, while a website change quietly cut off the leads upstream. There is an irony in that I have to sit with: the entire project was about building an engine I would only need to check once a week, and what pushed me to build it was discovering that my casual checking had already failed. The wound and the cure have the same shape. The difference is that a deliberate weekly look at an instrumented system is a discipline, and an occasional glance at a black box is a hope.

The other thing I got wrong was slower to admit. After the CRM failure I had pulled back from AI across the board, and that was overcorrection. The lesson was never that the technology could not be trusted. It was that I had pointed it at the wrong kind of work, the kind where being wrong hides. I spent months being cautious in general when I should have been precise about where.

What I would do now

This is the project where I got my footing back, and the way it happened is the lesson I would hand anyone rebuilding their own confidence in these tools. When Claude Code and the first MCPs arrived that winter, I did not return to asking a model what to think. I gave it tasks. Pull this data. Compare these campaigns. Stage this change, paused, and show me the dry run. Task work is verifiable: it either ran correctly or it did not, and the logs settle the question. Trust came back the same way it does with a person, through a body of small checkable work done right, not through a persuasive answer.

The engine now runs the way I wanted it to. Once a week I sit down with it deliberately: real numbers, anomalies surfaced, opportunities ranked. Roughly monthly, a campaign gets a full refresh, researched through Semrush, written with Claude, staged through the connector, reviewed by me, and only then set live. When something breaks or an opportunity opens midweek, I can act in real time instead of adding it to a backlog behind a login. The channel gets more consistent attention than it ever got from my hours in the interface, and it gets almost none of my week.

What I took from this is a definition of process I now apply everywhere. Assemble existing pieces first, build only what is genuinely missing, and spend the build where the loop fails to close. Most of my stack was other people’s work: Google’s connector, Semrush, Claude. The part I built was small, and it was the part that turned analysis into action without a human courier in between. Knowing which piece that is, and having the scar tissue to build it with the brakes on, is the actual job.