I run five websites, and every one of them is a real site doing a real job. This weblog is my lab notebook. The Chris Abraham Show site is the home of my podcast. Hill Mole is the spy novel I serialized on Movable Type starting in 2005. christopherabraham.com sells my consulting, and gerriscorp.com is my firm. Each of them is also a laboratory for search, for the old kind with ten blue links and for the new kind where ChatGPT, Claude, Perplexity, and Google's AI answers decide whom to name. The tools I use change every week, and these sites are where I find out what works before I sell it to anyone.
The rule is that nothing sneaky goes on any of them. No cloaking, no hidden text, no fake reviews, no doorway pages, no keyword stuffing, nothing that would get a site sandboxed if a Google engineer looked at it. Everything I test is something I would put on a client's site with my name on the invoice.
What they test is everything else. All five are static pages on GitHub Pages, built by Claude Code on my server, so a technical change that would take a WordPress plugin and a prayer takes a commit. On October 14 we ran a speed pass across the four older sites: Google Analytics, on the three that use it, now waits for the first scroll or tap, Chrome prefetches the next page when a reader hovers over a link, the robots files welcome the AI crawlers by name, and the podcast covers went to WebP. Every site has an llms.txt for AI assistants, structured data on every page, and IndexNow telling Bing about new pages the minute they go up. As of today, all four sites that run analytics use Google's Consent Mode, with no analytics cookies for visitors in Europe or for anyone whose browser asks not to be tracked, and Hill Mole, which runs no analytics at all, got a privacy page saying so.
Each site also tests something of its own. The podcast site has a page for every one of its 458 episodes, transcripts where Spotify made them, a glossary, and a hand-built back-of-the-book index with 1,774 headings, because an index is the kind of thing an AI can cite and a person can trust. Hill Mole wraps its untouched chapters in a reader's glossary, an index, sourced endnotes that catch the narrator in his mistakes, and an atlas, and keeps its 371 old Movable Type archive pages out of search so nothing appears twice. The two business sites publish guides and case studies with their own glossaries and site indexes, and their build refuses to publish a page that copies phrasing from the other site. This weblog is the newest, and it scores 100 on Google's Lighthouse for accessibility on every kind of page.
The lab results go somewhere. A small tracker I built asks the AI assistants the same sixteen questions and records whether my sites get named and cited, at about $1.60 a run, so I run it only when I mean to. What I learn on these five sites becomes what I do for clients, and what I can't make work here I don't sell.
Update, October 16, 2026, 6:36 PM: Later the same night Hill Mole got Google Analytics too, with the same Consent Mode settings as the other sites, and its privacy page now says so. The tracker's latest scorecard is live on the wiki: AI visibility report.