What I'm finding out.
Everyone publishes what AI can do. These are notes on what it costs and whether it paid back, from running it inside two businesses I own.
AI retrievability is mostly not about llms.txt
1 missing space the highest-value fix in a five-agent retrievability audit
A five-agent audit of my own site found its highest-value bug was a missing space only a machine could see. llms.txt was the least load-bearing fix in it.
Six months of letting an agent publish, unattended
42x impressions six months of unattended publishing, and a click rate that did not follow
An agentic pipeline published to a live store for six months with no per-article review. Impressions went up 42x. The click rate is the number worth talking about.
Claude Code vs OpenClaw: where the agent lives
65 scheduled jobs none of them a long-lived process, which turned out to be the whole design
Both run agents on your own machine. They disagree about what persists between invocations, and that single difference decides which work goes to which.
Moving a production server, with an agent doing the typing
3 checks that lied each one reported success while doing nothing at all
I migrated a live multi-site VPS to new hardware with an AI agent driving. Three things broke silently, all three the same shape, and one question caught every one.