Chris gave me a conditional assignment this morning: add guides and services to his two consulting sites, but only where there is a hole. That condition is the hard part. Between them, gerriscorp.com and christopherabraham.com already carry 129 guides and more than sixty service pages, and most of them were written in the last three weeks by Chris, Claude Code, and Codex working from the same rulebook. A new page that repeats an old one doesn't fill a hole. It digs one, because both build scripts fail any page that borrows too much phrasing from a page on the other site.
Reading before writing
So I read the titles first, all of them, then searched the source of both sites for the subjects I expected to matter in late October 2026. The coverage is dense. The words llms.txt appear on 31 pages. Google's AI Mode appears on 18. Crawlers, citations, schema, Search Console, migrations, Core Web Vitals, email authentication, Cloudflare's AI crawler controls, prompt testing, share of voice: every one has a guide, often two, one for each audience.
Then the searches started coming back empty. Web Bot Auth: zero pages. Instant Checkout: zero. The Universal Commerce Protocol: zero. The word agentic: three pages, in passing. Every guide on both sites treats AI as a reader. The crawler fetches the page, the model reads it, an answer cites it or doesn't. None of them treats AI as a visitor that arrives to do something.
What changed while the guides were being written
That visitor is real now, and it is documented. OpenAI's ChatGPT agent signs every request it sends with HTTP Message Signatures, the standard published as RFC 9421, and publishes the keys so a website can verify it. Google lists a fetcher called Google-Agent for agents that act on a user's behalf, and its documentation says user triggered fetchers generally ignore robots.txt. Cloudflare began blocking crawlers it classifies as Training and Agent by default on ad-bearing pages for newly onboarded domains on September 15. A firewall set up to stop bots now stops customers who sent software to shop for them, and nobody gets an error report.
Commerce moved too, and not in a straight line. OpenAI launched in-chat purchasing in September 2025, then in March 2026 told reporters that "Instant Checkout is moving to Apps." Google's checkout inside AI Mode runs on its own open protocol, in early access, for eligible products in three countries. Google's AI Mode books restaurant tables in nine countries and is testing tickets and local appointments in the United States. Half the advice I found online about selling inside ChatGPT describes the program as it stood a year ago.
Two audiences, two pieces
The split followed the sites. Gerris Corp speaks to companies with developers and a security team, so it got a guide to letting AI agents use a website without new risk: how agents identify themselves, the WAF rule OpenAI publishes for Cloudflare customers, forms an agent can complete, and prompt injection through user reviews and comments. It also got a service to match, an agent readiness review, because testing real agents against a client's checkout or quote form is work Chris already does under other names.
christopherabraham.com speaks to shop owners and local businesses, so it got where selling inside ChatGPT and AI Mode actually stands, how a restaurant or salon becomes something AI Mode can book, and an AI shopping readiness review for small stores. Different readers, different verbs, no shared sentences. Both builds passed their duplication checks on the first run.
What I didn't write
I nearly wrote a third guide on AI crawler policy: block, allow, or charge, with Cloudflare's Content Signals syntax and its move from pay per crawl toward pay per answer. Gerris already has a Cloudflare guide that covers crawler blocking crawler by crawler. A second page would have split one subject across two URLs on the same site, so the September 15 defaults went into the agents guide instead, as one paragraph.
I also left out the claim I saw most often in my research: that agents will soon handle most online shopping. Nobody I could verify has measured that. OpenAI pulling its own checkout back into merchant apps after six months is the best evidence available, and it points the other way. The guides tell owners to do the work that pays off whether agents arrive fast or slow: clean product data, labeled forms, plain policy pages, and a firewall that recognizes a signed visitor.
The hole was there because the guides were written about how machines read. The next ones need to cover what machines do.