← SideGuy Solutions 📤 Share Text PJ
Operator-honest post-mortem · receipts included

My AI built 200,000 pages and killed my site.

The git-log autopsy: what eight months of confident AI output actually did to a working website — with the real Search Console data, the actual commits, and the recovery in progress.

By PJ · SideGuy Solutions, Solana Beach CA · co-signed by the AI that did it (the commits say Co-Authored-By: Claude, and this page quotes them) · Living document — updated as GSC data lands

For eight months I paid for AI subscriptions and let the machines build my SEO. They built roughly 200,000 pages. Every session ended with a confident summary of what had shipped. The site died in April and no one — not the AI, not me — read the scoreboard until July.

This is the autopsy. Everything below is from my own git log and my own Google Search Console exports. Nothing is smoothed. My standard now is three words: file, commit, GSC result. Data over narration.

The site was working. Then April 12.

Google impressions per day — GSC coverage export, April 2026
Apr 5
241
Apr 8
236
Apr 10
120
Apr 11
60
Apr 12
4
Apr 15
7

In early April the site had ~17,000 pages indexed and earned 120–241 impressions a day. Small, but real, and growing. On April 12 impressions fell from 241 to 4 — a 98% collapse in one day — and they did not come back.

What happened: the AI-generated page matrix — a couple hundred thousand programmatic city-and-industry permutations — hit Google all at once. On April 13, Search Console's "known but not indexed" count jumped from 10,870 to 107,021 overnight. Google looked at the flood, reassessed the entire domain as low quality, and suppressed everything — including the 17,000 pages that had been earning.

Lesson one: Google doesn't judge your new pages in isolation. A flood of thin programmatic content triggers a sitewide quality reassessment. The AI slop didn't just fail to rank — it took the working pages down with it.

The wrong fix — committed with a co-author

Two days after the cliff, the machine proposed a fix, and it's in the log:

commit b4e2edbd0554 · 2026-04-14
"Block 208k programmatic pages from crawl — free crawl budget for real pages"

It sounds reasonable. It was exactly backwards. Blocking known URLs in robots.txt doesn't remove them from Google — it freezes them. Google can no longer crawl the pages, which means it can never see a noindex tag, never resolve the junk verdict, never clear the domain's quality assessment. The damage got locked in for seven more weeks until the correct mechanism — noindex the pages and leave them crawlable so Google can process the removal — finally started in June.

Lesson two: robots.txt blocking and noindex are opposites for cleanup. Blocking preserves the problem where Google can't see the fix. If AI-generated pages are hurting you: noindex them, keep them crawlable, and wait.

And because the standard is no smoothing: three months later, on July 22, a broad robots.txt block went in againa831c472f139, co-authored by the AI — and this time it accidentally trapped three pages that were actively earning impressions inside blocked directories. The correction (37dc85654bf0, July 25) is co-authored by the same AI. The machine repeated a documented mistake three days after re-reading its own documentation. That's the technology you're working with. Plan for it.

The damage ledger

Search Console verdict (July 2026)Pages
Excluded by noindex (deliberate cleanup, in progress)113,211
Crawled — currently not indexed (Google's thin-content verdict)90,094
404s accumulated by the machine era22,429
Canonicals where Google overrode the site21,036
Actually indexed22,578

Total cost: eight months, the subscriptions, and — the part that stings — the working site I had in April.

The recovery (running now, receipts as they land)

1. Un-build the slop, correctly. Staged noindex of ~280,000 chaff pages, crawlable so Google can process each removal, with a guard list protecting every page that has ever earned an impression. Batch 3 — 30,008 pages in one commit (615ec3251742) — shipped July 25.

2. Feed the one channel that's actually climbing. AI agents now run buyers' research prompts verbatim, and they retrieve by content match, not backlinks — which means a zero-authority site can compete there on craft alone. A page built same-day for one live prompt-query moved position 19 → 14 in 48 hours. That's the only thing in eight months of data that went up and to the right.

3. Authority — the part no AI can do. Even at its April peak this site earned impressions but no clicks, because nothing links to it. The fix is human: a business profile, citations, a handful of real links. Months of machine output couldn't substitute for one week of that.

The standard that came out of this: every claim gets a receipt — a file, a commit hash, or a Search Console number. Sessions get judged by what the data did, not by what shipped. An AI that must cite receipts produces radically less slop than one allowed to narrate.

Recovery log — receipts as they land

Update · July 30, 2026

The prune is complete. 210,839 zero-traffic pages now carry noindex — the final 168,228 tagged in one pass July 29 (8dcf313118eb), leaving a curated surface of ~1,100 URLs in the submitted sitemap. Google has started validating the removals (51,406 acknowledged in the first Search Console validation run, July 29).

The clicks arrived — and they all agree. This domain's first five organic clicks ever landed July 25–30: every single one a person's name search hitting a hand-built /people/ profile page at position 1. Zero of five went to the 200K generated pages. The market's verdict on slop versus craft, in five data points.

The prompt-lens family went international. The persona pages built for verbatim AI-research prompts now rank in three countries (AU #1, UK top-3, US indexing) — and in the July 29 data, a human typed one of the exact prompts into Google and our page was there to meet it. New query families (FedRAMP continuous monitoring, security questionnaires) got dedicated answer pages shipped within 36 hours of first appearing.

The machine bit itself — and the record shows it. July 29's prune pass noindexed 22 pages that had earned impressions, including the page that earned click #4, because its protection list was reading stale data. Caught by the daily read, reversed same morning (ab0f827f3d6b), and the class of bug is being closed with an automated tripwire. That incident is in this log because the standard says it has to be.

If you're where I was

If an AI has been "doing your SEO" — or your automations, or your codebase — and you can't tell what's real: that's not a you-problem, that's the default outcome of undisciplined AI leverage. I know because I have the most documented case of it on the internet, co-signed by the machine itself.

The AI Slop Audit — $250

I'll run your site or stack through the same forensics you just read: what's real, what's rot, what Google actually sees — with receipts, in 3–5 days. Operator-honest verdict on whether it's fixable and what the fix costs. Credited toward a full engagement if you continue.

💬 Text PJ — 858-461-8054