Read it, then pass the check at the end to open step 6.
End game: the loop, data in, agent fixes, request, measure
Updated
This is the step where AI enters, and it enters as a pair of hands, not as a search engine. The loop in my operating system: search-read.ts joins Search Console, Bing Webmaster Tools and PostHog into one CSV per page, with the site’s own events (booking clicked, form sent, call booked) in the same row as clicks and impressions. Claude Code, an agentic coding tool, reads the CSVs and the live HTML, proposes the fix, edits the code, and lands it. I approve. The data, the fix and the request go through one door, which is the same reason I put every client, quote and article in one database, as the second-brain column shows with the cutover’s own dates.
One dated example, because one is what I have. 2026-10-15: the API read finished at 16:37 UTC. The agent crawled the live site for links and found none to the service pages. It added a Services list to the footer of every page, seven links in English and eight in French, commit 694d211. Vercel built it. I verified it live on /about and /fr/about within the hour. The claim I make from that, labelled as experience and graded as an estimate: minutes of work, days of waiting. The index result is the part that still takes days, and I publish the dates when they land.
Two rules the loop runs on. The agent writes nothing it cannot cite: on 2026-10-08 a run with no evidence file invented 6,247 words of first-party data that did not exist, which is why every piece, this one included, is written from an evidence file the desk checks. And the loop never reaches for borrowed authority: in 2024 I cross-linked four cheap domains with sameAs to lift a site nobody linked to, and I would not do it again. The entity gets built through a Business Profile and real profiles, or not at all.
The bank behind the keywords holds 3,835 rows as of 2026-10-15, each with its lane and its route, so the page a reader lands on was chosen from data, not from a mood. What the script reads and writes, and what I still have not measured, is in google search console api.
Check: your last fix has a commit, a deploy time, and a line in your own log saying which URLs were requested and when.
Check what you read in “End game: the loop, data in, agent fixes, request, measure”
Three questions drawn from a larger pool, so no two papers are quite the same. Two right passes it and opens the next step. The questions are public; scoring them needs an account, and the key never leaves the server.
Loading the check…