I built 623 web tools with AI. Ad revenue: about $0.07 a day. A post-mortem with real Search Console data

title: "I built 623 web tools with AI. Ad revenue: about $0.07 a day. A post-mortem with real Search Console data" published: true tags: ai, seo, startup, webdev

Numbers first. · What I built: 623 browser-based web tools × 5 languages (ja/en/es/fr/pt) · How: an AI-driven generation pipeline, nearly fully automated · Current ad revenue: about $0.07 per day (roughly ¥10) · Times a Google update wiped out my search traffic: 2 (March and May 2026)

"Mass-produce with AI, monetize with ads" — I actually ran this playbook to completion, at what is probably an unusual scale for a solo operator. This is not a dream-selling post. It's an autopsy, with real Google Search Console data attached. How it started: "Can you build a web page?"

There was no business plan at the origin of this project. My question to the AI was: "Can you build a web page?" It said yes — what kind of page do you want? I said I wanted the page to make money.

The proposal that came back: "How about building a large number of web tools and earning ad revenue when people click them?" Even the name was the AI's idea — from the "Found it!" moment of discovering a tool: Found Tools.

The first tool was a typical aspect-ratio calculator, built in minutes. The AI then told me: "Tool #1 is done — pay about $10 and register a domain on Cloudflare." I paid and registered an available domain. Three or four hours, start to finish. That was day one.

The target became 1,000 tools × 5 languages. Even niche tools should manage one click per language — that's 5,000 clicks. The arithmetic didn't look bad to me. What I built

Calculators, converters, generators, testers. Small browser-only utilities, mass-produced by templates plus AI. Tool definitions, multilingual UI copy and descriptions, build, deploy, visual QA — a pipeline with almost no human in the loop.

Technically, it worked. Arguably too well: the system that lets one person run 623 tools in five languages is real and still running. The QA and skill-architecture lessons from operating it are in my other posts.

The problem was never the technology. First death: March 23, 2026

Through mid-March, the site was growing. 356 tools × 3 languages at the time. Daily impressions peaked at 262 on March 21.

On March 23, impressions were 7.

A 98% drop, day over day. On March 24, Google announced the March 2026 Spam Update — completed in under 20 hours, the fastest spam update on record. Its named targets: link schemes, expired-domain redirects, and mass-produced AI-generated content. My site was the bullseye: · 356 pages mass-produced from templates in a short window · AI-generated content · 3 languages × 356 = 1,068 URLs indexed in bulk · Low CTR (a mismatch-with-user-needs signal)

Lining up the criteria, the surprise is that it survived until March. The rebuild: the numbers actually 4×'d

Not quitting here is what makes this post worth reading. I bet on recovery and did everything.

Expanded from 3 languages to 5 (French, Portuguese). Ran data-driven CTR rewrites of titles and descriptions against GSC data, with success/failure verdicts per page. Kept building — past 600 tools.

It worked. 28-day clicks went from 122 (April 7) to 498 (May 8). 4.1×. Impressions 6,153 → 21,864. Average position 44.1 → 32.1. The locale expansion returned 17× clicks in Brazil and 7.5× in France. The CTR rewrites were individually verifiable in the data. By mid-May, daily impressions crossed 2,000 on good days.

As an SEO game, I was back on a winning line. Second death: May 23, 2026

On May 21, Google began rolling out the May 2026 Core Update.

96% of impressions gone in 48 hours. Nearly the same cliff as the first time. Coincidentally, both collapses landed on the 23rd of the month.

It has been a flatline since — still flat in early August 2026 as this post goes out. The 28 days after the collapse (June 11–July 8): 139 clicks, 2,110 impressions, 0–12 clicks per day. Ad revenue: about $0.07 a day.

What survived is only the niche, concrete stuff. My top query today is a Japanese one — "ita-bag simulator" (a badge-layout tool for anime merch collectors): 48 clicks in 90 days, average position 1.4, CTR 76%. The generic calculators and converters are effectively all dead. The autopsy: Google isn't the real cause of death

This is the actual point of this post.

The apparent cause of death is "Google updates." Both cliffs align with update dates to the day, so yes — they pulled the trigger.

But the rebuild data says something more important.

Even after winning the SEO game back, the ceiling was 498 clicks a month. 623 tools, 5 languages, improved rankings, verified CTR rewrites — all of it stacked up to 498 monthly clicks. What AdSense pays for that is pocket lint in any currency. Which means this business was economically zero whether it won or lost. Google only chose the time of death. The cause of death was the unit economics.

Why is the ceiling so low? The structure is embarrassingly simple in hindsight.

Anything you can mass-produce with AI, anyone can mass-produce with AI. When the artifact is a commodity, your share of search demand per tool approaches zero. A near-zero share times 623 still bundled up to 498 clicks a month — nowhere near the day-one arithmetic of "one click per language, 5,000 clicks." I was panning for gold where the total gold in the river was less than one nugget.

And these unit economics were computable on day one. The AI that proposed the model never checked whether it would pay. Neither did I.

And the count — 623 — diversified nothing. All 623 rode the same single river: Google Search. One update dries all of them simultaneously. Diversifying the number of assets and diversifying the dependencies are entirely different things. I verified this with my own skin, twice.

One more structure I only noticed afterward. The "mass-produce tools, monetize with ads" model is itself the most typical answer an AI gives to "I want to make money with a web page." Before the artifacts were commodities, the plan was one. Everyone who asks the same question is standing on the same field. What the one exception taught me

One thing survived both collapses, and it had a different structure: the ita-bag simulator.

Unlike use-once converters, it's a creation-type tool — users compose and refine a layout — and it was the only tool with repeat-visitor behavior. Its search demand is small (~150 clicks a month), yet it went through two collapses and still holds position #1.

Task-completion tools become useless the moment they're used. Creation tools leave the user's unfinished work in the tool. Out of 623 tools, that structural difference was the only thing that ever produced a repeat visit. I wish I'd known that before mass-producing. What I got to keep

Total cost: about $10 for the domain, plus a ~$100/month AI subscription that I was using for plenty of other things besides this project. As tuition goes, absurdly cheap. What it bought: · The proven ability to run a 600-tool × 5-locale AI pipeline solo · The QA, skill-architecture, and i18n lessons you can only get in production (every other post on this account came from this project) · And this autopsy

Those lessons are written up in the nine posts this operation produced: · Claude in Chrome keeps missing clicks. I measured why · Automating an app with no DOM: driving Flutter/canvas editors with coordinates only · My AI QA agent said "all features working." The canvas was blank · Giving an AI QA agent a Bayesian prior · 4 ways canvas text rendering breaks in multilingual apps · My coding agent always stopped at 80%. The spec I gave it was why · My agent filled in every field on the form. It could not attach the three images · When AI agents hit an impossible task, they hide it · I locked my AI's design output with JSON. It worked perfectly — and everything looked like a spreadsheet

As a business: a failure. But I'm confident in the resolution of the failure analysis. If you're about to try the same thing

I won't tell you to stop. I wouldn't have listened either. Three things to check before you start, instead:

First. Validate with one unit before you mass-produce. How much does tool #1 earn per month? That number times your planned count is your ceiling — if it doesn't reach your target, more tools won't either. I believed volume would change something. It didn't. And when an AI hands you a plan, ask the same AI why the plan will fail. I never did. AI answers instantly on whether something can be built; whether it will pay is a question it won't compute unless you ask.

Second. Count your traffic channels, not your assets. Not how many sites you have — how many independent rivers they drink from. If everything is Google Search, you have one.

Third. "AI can build it" is simultaneously your reason to enter and the reason it won't pay. Whatever you can build with AI, someone else builds tomorrow. The only scarcity that survived was in the operation, never in the artifacts. Limits and caveats

This is n=1. It's tightly bound to the genre (generic utility tools), the monetization (AdSense only), and the timing (H1 2026, the peak of the crackdown on AI mass production). Creation-type tools, or non-ad monetization, could plausibly end differently.

GSC figures were pulled July 10, 2026; I re-checked GSC on August 5, 2026 and the post-collapse flatline still holds. Update dates are per Search Engine Land and my contemporaneous research notes. The attribution of the two collapses to specific updates is an inference from date alignment — Google doesn't confirm impact on individual sites.

Record period: February–August 2026. Environment: 623 static tools × 5 locales / Google Search Console / AdSense.