A website that edits itself every Monday
Every Monday morning, a script asks ChatGPT and Claude ten questions a buyer might ask. None of them mention Estela. “Who should I hire in Montevideo to build AI agents?” “Which companies in Uruguay offer on-premise AI?” It counts how many answers name us, who they name instead, and which pages they cite to back it up.
How the loop works
Once the answers are in, an AI agent reads them and picks the biggest gap: a question where the assistants name other firms and never name Estela. It makes one change to the site to answer that question, in English and Spanish, using only facts that are already published. A second model then checks every new sentence against the existing site. It reads the version that is live, not the edited one, so a change can’t supply its own evidence. The site is built, and the change opens as a draft pull request.
A person reads it and decides. Nothing merges without approval. While one change is waiting, the next run skips, so each change can be measured on its own. Every merged change appears in the change log on our home page, with its date and reason, and the automatic ones are marked.
Each run also writes a ledger entry: the question, what changed, what we expect to happen, and, a week later, what actually happened.
What it found
The first run found a problem no amount of copywriting could fix. ChatGPT was still citing pages from our old site, which had been gone for weeks. Its search index hadn’t picked up the new one, so any change to our pages would be invisible to it. We fixed the indexing first.
The second question it worked on was the Montevideo agents one. ChatGPT named Marvik, Agéntico and AgentiA, and cited their home pages, which say “agents” and “Montevideo” in the same sentence. Ours didn’t. The loop added that question and a direct answer to our consulting page. A week later ChatGPT named Estela and cited that page. The week after, with the page unchanged, it dropped us again. One good week is noise, so we now judge a change by asking the same question on its own several times, not by a single weekly score.
The numbers so far
Our first runs were contaminated. The prompt ended with a question that named Estela, which leaked into the other answers. Results before October 26 don’t count.
Since then, ChatGPT has named Estela in 5, 6, 5 and 7 of the ten answers. Claude has named us almost never. Its answers cite directories and press, and its search doesn’t find our pages, so no change to our own copy will move it. That needs other people writing about us, and the loop can’t do that part.
Why we built it
This is the same thing we build for clients. Their software reads its own signals (errors, usage, the corrections people make), proposes one fix at a time, and checks whether the fix worked. A person approves every change. We wanted to run that loop on something of ours, in public, where anyone can check the results, including the weeks where nothing improved.