How We're Growing About This Car With an AI-Run Marketing System

Most companies that say “we use AI for marketing” mean it wrote some social captions. We do something narrower and, we think, more useful to show you: an AI agent runs the actual growth process for one of our live products, About This Car, under a written set of rules it isn't allowed to break. This is what that looks like from the inside, including the parts that haven't worked yet.

The problem with “just give the AI access”

Handing an AI agent write access to a production codebase and a “go grow this” instruction is how you end up with either nothing happening (too cautious to act) or something breaking (too eager to act, no guardrails). We wanted a third option: an agent that can move fast on the boring, safe stuff and stops to ask before doing anything a person should sign off on.

So before any code got touched, we wrote a permission ledger. Every category of action the agent might take, from “edit a metadata tag” to “deploy to production”, starts in an ask-first state. Each time it wants to do something new, it has to name the category, explain the exact scope, and propose a rollback plan before asking. If we say yes and mean it every time, that category stays ask-first. If we say “yes, and remember this”, it becomes autonomous, but only within the scope we agreed to.

What's actually been graduated to autonomous

Not much, on purpose. Two categories currently run without asking each time:

• Invisible code changes on a branch. Things a user would never see: meta tags, sitemaps, structured data, analytics wiring. The agent can make these changes and push them to a working branch by itself. Merging that branch into the live site still needs a yes from us, every time.
• Publishing guide content, once we'd seen three examples and agreed they matched the bar we wanted. After that we told it not to ask permission for each new guide, just to tell us after the fact so we can review, edit or pull it down.

Everything else, including anything that changes how the product looks or behaves, stays ask-first. No exceptions we've made so far have been permanent; we can revoke any of them at any time, and one grant was accidentally given during a UI test and reverted the same day once we noticed.

What it has actually produced

Four published guides on About This Car answering real questions used-car buyers search for (MOT history, DVLA vehicle data, car tax, a pre-purchase checklist), each one drafted, checked against the style rules, and shipped without us writing a word of the copy ourselves, though we read every one before it went live. Technical SEO foundations (sitemap, structured data, Open Graph tags) that used to be the kind of task that sits on a list for months got done in an afternoon.

What it hasn't produced yet: real growth numbers we can show you. This is early. The measurement plumbing (analytics, Search Console) went live at the same time as the first content, which means we genuinely don't have enough weeks of data yet to claim a result. We'd rather tell you that plainly than round up.

Why this is the same thing we're offering you

If you're an indie developer with something live and you're wondering whether “getting help” means losing control of it, this is the honest answer to what that could look like. The system doesn't get to decide things for us. It proposes, we decide, and every decision it's allowed to act on without asking again is one we chose to grant, on purpose, in writing, and can take back. That's the model we think works for a developer handing off the parts of running a business that aren't the part they actually want to do, compliance, hosting, marketing, while keeping the product and the decisions that matter theirs.

We're not done building this. When we have real numbers, we'll publish those too.

FAQ