About 16,000 search visits in the peak month, from a program run by one person with AI agents

What made this possible

Most of the jobs a search program needs are now jobs AI agents can do, and how to run a good one is written down in public, so a company can run one without first hiring the people it used to need. The program covered the jobs a search team does, and agents did each one. Curriculum data and verified numbers fed the research and the pages, the technical side sent them to Google, Bing and AI answer engines (chat assistants that answer a question and cite web pages), and a daily job and an agent watched the numbers. Nothing went live without passing the checks, and what they found shaped the next topics.

The challenge

I'd spent about seven years selling data and analytics products to the SEO industry, and earlier I'd managed a content marketing and SEO program, so I knew the field. I was never a practitioner, though. The bet was simple. Could I take what I knew and, with AI agents, close the gap to what a practitioner knows?

The constraint was real. A proper search program takes two or three people. In Europe that is around €6,000 to €9,500 a month. An agency charges a similar monthly fee. What was at stake was our reputation if we published low-quality AI text, and a slower company if we did nothing.

What the program was

First, AI agents sorted real search queries by what the searcher wanted and joined each kind to what visitors then did, so we knew what searchers did on arrival. Agents built the pages per topic and grade, in each country's own language. Data pages tied every number to one verified record. Tests on the main jobs decided which AI model did the work.

Automatic checks, some scripts and some AI models, caught wrong citations. Failed work wasn't published. Before anything shipped, an audit checked the research citations the AI had written against the real papers. Every incident became a new check. Agents built the technical side into every page, and a daily job sent new pages to Bing.

A daily job saved Google's numbers and the analytics tool counted what visitors did. A statistical alarm flagged drops outside its normal range, and an agent found the cause. Once agents had tested the cause, I decided what to change. Later, only non-routine releases, spending and messages to people outside the company came to me, the calls that changed direction.

Choosing a model per job

The method improved as we used it. A near-miss model gets a second chance before a winner is picked. In the test for the curriculum pages, a cheaper model lost on formatting, not writing. After the second chance it took the job. The model that had won the first round would have cost roughly seven times as much.

pareto-evals on GitHub

The peak and the fall

The rise was fast. Our pages appeared in Google results 32,015 times in February 2026 and about a million times a month by April. Then, after a Google update in late May, visits from Google fell 96% in one week. By late June almost all the pages we had built per topic and grade were gone from Google's results.

Testing the explanation

Before believing any explanation, we tested it. Google's own records showed it had never downloaded just under half the pages we asked it to read, so for half of what we built, the writing can't have decided anything. Then a quality test, written down before it ran, compared pages Google kept with pages it refused, scored by a judge that never knew which group a page came from. The difference was about zero.

The best-supported explanation is a very large, very new set of templated pages on a site Google didn't know yet.

What the traffic made possible

Enough visitors changed how we worked. Recordings of real visits let us find problems early, about twenty of them. In one case an AI answer engine sent a visitor from Brazil to a Spanish curriculum page, a dead end; the recording showed it, and the fix shipped the next day. We ran split tests. People in 17 countries answered our surveys.

Bing took longer to warm up, and it's now the main search channel, still listing about 101,000 of the same pages.

What I would tell the team that owns this work

Knowing the field still matters. I couldn't have run this without knowing the tools, the tactics and how to check an agent's work, and I'd say that holds in any field, not only search. Learn what great looks like, so you can direct your agents.

Believe in the channel. Search is one of the best returns for many brands, especially closer to consumers. Start early, because it takes a while to warm up.

Be careful with large-scale actions. Don't generate a lot of AI content at once, and I'd launch fewer markets first as a precaution. A lot works without those, technical SEO, honest digital PR, a good experience on the site, and a presence on review platforms, which matter even more now that AI answer engines choose the sources they cite.