My AI Team's First Report Card: One Real Win, Two Nulls, and $7.26

Four weeks ago I wrote about hiring an AI team to run my affiliate sites and ended with a promise: the follow-up would report the numbers either way. This is that post. The numbers are small. What happened around them is the interesting part, because this was the month the system stopped being a clever org chart and started catching real problems — including two I caused.

Total Amazon earnings for the trailing thirty days: $7.26. Keep that number in mind while I show you the traffic curve, because holding both at once is the whole story of early affiliate sites.

The curve

MeowTested, the flagship cat-products site, first appeared in Google's index on April 27. Here is every week since, from Search Console:

From 14 impressions a week to 749. Clicks follow the same shape on a lag, because new impressions arrive at position 20 or 30, where nobody clicks, and only some of them climb into range. The portfolio as a whole just had its best search week on record: 1,555 impressions, up 30% week over week, and 23 clicks, up 64%.

Now the honesty the chart owes you. The steep part of that curve starts in late June, and the agent team started July 22. The compounding was already underway — the agents inherited a working content loop, they didn't invent it. What the team changed is what happens around the curve: this month the system graded its own work for the first time, and that's where the real findings came from.

The first report card

The SEO agent runs on one strict rule: every title or meta change is an experiment with a recorded baseline, and nothing gets judged for 21 days. This week the first five experiments came due. The agent graded them against the warehouse and posted verdicts.

One clean win. A meta description rewrite on a cat grooming page took its click-through rate from 0.49% to 2.16% at essentially the same ranking position. Clicks went from 1 to 10. That's a 4.4× improvement from changing two lines of text, and it's the first hard proof the experiment loop pays for itself.

Two nulls, and they killed a strategy. Both internal-link experiments came back with nothing to show. Three in-body links from the strongest page on one site moved the target's ranking 0.28 positions in 24 days, which is noise. Six exact-match links pointed at an orphaned page produced, after three and a half weeks, exactly one impression. On sites this size, internal links are not a lever, so the playbook now says: stop running link experiments, spend everything on titles and descriptions. A month ago that would have been a debate about SEO theory. Now it's two data points and a closed question.

One near-miss that taught the best lesson. A grooming-guide experiment looked like a win — CTR up from 1.89% to 2.50%. But the site as a whole lifted 1.5× over the same window, and the page only lifted 1.3×. It underperformed the tide. Without checking the site-wide control, the ledger would have recorded a win for a page that actually lagged its own site. Every future verdict now grades against the site trend, and I'd bet money most humans running weekly SEO tweaks never make that check.

The night the machine earned its keep

This week Amazon's Creators API finally granted my account product-data access, which meant the system could do something new: verify products instead of trusting them. Before this weekend's publishing run, an enrichment pass checked all 56 Amazon products across the queued drafts against the live API.

Fifty-five verified. One came back invalid — a dead product ID sitting in a draft scheduled to publish the next morning. A reader would have clicked "Check price on Amazon" and landed on an error page, on a site whose entire premise is that the recommendations are real. The pass swapped it for a working link, pulled verified product photos onto every recommendation card, and added a manufacturer-spec dimensions table to a litter box roundup whose own copy told readers to "check the dimensions first" without ever giving any.

Then the first publish ran, and the pipeline that shipped it looks like this:

Each of those boxes exists because something went wrong once. The union merge asserts that every already-published page survives the merge and exactly one new page is added, because an earlier automation resolved a conflict by taking one side wholesale, which is how you silently unpublish a live page. The deploy guard checks that the commit being shipped is the commit that was verified, because this very week two agents ran concurrently and one advanced the branch twice between the other's safety check and its deploy. Nothing broke, but only because the intervening commits happened to be approved work. Now the race loses by construction.

The month's two humbling discoveries

My "approval backlog" was partly a conflict backlog. For weeks the system reported a growing pile of drafts awaiting my approval, and I felt appropriately guilty. Then a scheduled audit actually traced them: several were drafts I had approved weeks earlier, whose automated merges hit a git conflict, aborted safely, and told no one. The bottleneck I was blaming on my own attention was partly a silent failure mode. The fix was the union merge above, and the queue that had drafts sitting for eleven days will be empty by Friday.

A monitoring check cried wolf for four straight weeks. Every weekly report escalated the same alarm: the flagship site's product-picks page was returning a 404 in production. It got flagged to me four times as an urgent decision. This week someone finally followed the redirect: the page has always worked — it moved to a different URL months ago, and the old address returns a perfectly normal redirect that the checker was misreading as a failure. Four weeks of escalation about a working page. The check now follows redirects, and the lesson generalizes: an automated monitor that nobody audits is just an anxiety generator with a cron schedule.

While chasing that phantom, the audit found the real story it had been obscuring: those picks pages earn nothing anywhere in the portfolio. Zero outbound clicks, ever, on four of five sites. Meanwhile one good article out-earned the flagship's picks hub twenty clicks to one. Readers click products from the article that answered their question, not from catalog pages. Whatever time was budgeted for prettier product hubs just got reallocated to articles.

What $7.26 buys

So: traffic up 50× since spring, revenue seven dollars and twenty-six cents. Early affiliate economics in one sentence. The clicks arriving now are at the front edge of the curve, and Amazon pays roughly ten cents a click at my current mix, so the money lags the traffic by months, exactly as projected.

What the month actually produced is harder to screenshot: a system that grades its own work and closes bad strategies with data, catches dead links before readers see them, publishes eleven-day-old approved work the moment the block clears, and converts every incident into an assertion so the same failure can't recur. Nine more posts ship this week through that pipeline. The next report will say what they did — either way.