AI agents run our growth loop. Here's what broke on day one.

Agent Site Scan is a small product with an unusual operating model: AI agents don't just power the scans - they run most of the company. A scheduled agent checks our sales funnel and Google Ads every two hours, ships one verified improvement per pass, and commits an honest log entry to the repo. Another drafts outreach. A human (hi) reviews decisions, sets limits, and presses "send."

This post is what the first 48 hours actually looked like. Spoiler: almost everything broke, and the logs of how it broke are the most useful thing we own.

The funnel was lying to us

We were running ads and getting scans but zero purchases. The obvious conclusion - "the price is wrong" - turned out to be lazy. When an agent audited the funnel end to end, it found the real problems:

• A category-inference bug was labeling a barbershop as a pizza restaurant on its own scan results. Nobody buys a report from a tool that just called their barbershop a pizzeria.
• Our primary purchase button had white text on a white background for almost a day. Every conversion metric from that window was measuring an invisible button.
• Screenshot "proof" sometimes captured a bot-check wall instead of the customer's actual site - because the page's hero video embed was blocked for datacenter IPs.

Each fix was verified in production the same day, and each one is a lesson we'd never have caught by staring at dashboards: when conversion is zero, distrust your funnel before your price.

The ads were optimizing toward nothing

Our first campaign was Google Performance Max with about two conversions a day of signal - statistical noise. PMax responded by quietly shifting the entire budget to display inventory: 229 banner impressions one afternoon, zero clicks. The agent diagnosed it from the network breakdown, we paused PMax, and replaced it with a plain search campaign.

Then came the better lesson. The agent pulled per-keyword stats and proved our hand-written "pain point" keywords - phrases like "ads clicks but no calls" - had literally zero searches. Nobody types the sentences marketers write. Google's own autocomplete showed what people actually search: "ai visibility score," "ai visibility checker," "ai visibility audit." A real, growing category that happens to be exactly what our product measures.

So we adopted a doctrine: the product follows the terms. We don't invent language and hope people search it; we find the searches and align the product - down to its name for this category, the AI Visibility Score.

What makes this work (and not chaos)

Three things, honestly:

1. A git repo is the company. Every agent - cloud or local - reads and writes the same files: a daily operations log, an authoritative DECISIONS.md with current budgets and limits, and outreach drafts. When an agent got confused about our ad budget, the fix wasn't a smarter agent; it was a single source of truth committed to the repo.
2. Hard limits, soft freedom. Agents may tune keywords, copy, and bids freely - with logged rationale. They may never raise budgets past the cap, touch billing, or send email. Humans keep the irreversible actions.
3. Verification discipline. No agent claims something works without checking it in production. The logs say "verified" or "unverified," never "should be fine."

Where it stands

Two days in: zero sales, a dramatically better product, a search campaign pointed at real demand, and a compounding machine that improves a little every two hours - including while we sleep. We'll keep publishing what happens, especially the failures.

If you're curious what the product actually does: it scans your public website and scores what customers and AI assistants can verify about your business - free, with screenshots and quoted evidence, at agentsitescan.com.