Experiment

This started as a review of something that looked like it was already working.

A daily routine has been publishing one glossary term to AI.AS400Software.com since September 7, 2026. Five runs, no failures, terms landing clean. It felt like early traction, and the plan was to copy it across the portfolio.

Then I actually pulled the numbers, and they said something less flattering.

Crawler requests averaged 27.3 a day in the thirteen days before the routine started and 32.0 a day in the five days after. Search impressions went the other way, 168 a day before and 140 a day after. On five days of data, neither of those is a result. Worse, the real movement on that site happened on August 25, when impressions went from 9 to 120 to 260 in three days. That was Google finding the site for the first time. It predates the routine entirely.

So the honest read is that the machinery works and the payoff is unproven. That is a fine place to start an experiment from. It is a terrible place to start a rollout from.

That reframed the question. Instead of copying a win, this became a test of whether the win exists at all, run across enough sites and enough time to actually answer it.

Selection came out of the Digital Karma warehouse. All 126 deployed sites were scored on three things: how long since anything on the site changed, how many warehouse events it had in the last 60 days, and how much verified crawler attention it was getting. Sites already under active work were excluded, because a routine publishing alongside a human is a confounded test. Sites with almost no crawler or search signal were excluded too, because you cannot read a result off a flat line.

Forty-eight sites cleared the neglect filter. Ten were selected on vocabulary depth, meaning the subject has enough real terminology to support months of honest records without padding.

The ten, with 28-day baselines to September 9, 2026 for search and September 11, 2026 for crawler data:

• NaturopathicScience.com ... 1,756 impressions, 0 clicks, position 73.8, 1,494 crawler requests, 74 days idle, 27 files
• MedMarketingDr.com ... 2,728 impressions, 0 clicks, position 73.9, 458 crawler requests, 49 days idle, 29 files
• HealthyWebAI.com ... 2,308 impressions, 0 clicks, position 66.2, 356 crawler requests, 80 days idle, 43 files
• SupplementsNoMin.com ... 2,081 impressions, 2 clicks, position 63.8, 385 crawler requests, 66 days idle, 64 files
• eInkGallery.com ... 1,779 impressions, 12 clicks, position 42.9, 901 crawler requests, 43 days idle, 36 files
• Serv-Pak.com ... 1,802 impressions, 24 clicks, position 26.0, 906 crawler requests, 21 days idle, 46 files
• SpaForBusiness.com ... 1,571 impressions, 2 clicks, position 53.4, 818 crawler requests, 63 days idle, 62 files
• FlexPackMFG.com ... 1,278 impressions, 2 clicks, position 59.7, 704 crawler requests, 56 days idle, 57 files
• AI.AS400System.com ... 846 impressions, 1 click, position 58.4, 486 crawler requests, 33 days idle, 33 files
• RescueOurPlanet.com ... 218 impressions, 0 clicks, position 76.7, 385 crawler requests, 72 days idle, 65 files

One routine serves all ten. It takes two sites per run, so each site receives a new record roughly every five days. Portfolio-wide that is two new records a day, which is deliberately quiet. The point is to test cadence, not to flood ten sites with filler and then claim the filler worked.

Everything else about these sites stays frozen for the duration. No new templates, no internal linking push, no schema changes, no backlinks. If something moves, the publishing cadence is the only thing that could have moved it.

Two constraints shaped the build more than anything else. Anthropic cloud sandboxes can only open outbound connections on ports 80 and 443, so the routine cannot use ssh or scp to reach the server. Everything goes through an authenticated HTTPS endpoint that accepts JSON and nothing else. And the ten sites turned out to carry six different glossary schemas between them, which is its own small lesson about what happens when a content type grows up across a portfolio without a contract. The roster file carries the exact field shape per site so an unattended run cannot guess.

Status: active, and deliberately early.

What would support the hypothesis: crawler request frequency on a roster site rising within two to three weeks of its first record, followed by impression growth on the glossary URLs specifically. The glossary is the cleanest possible test surface because each term is its own URL with its own query surface, so attribution does not require untangling a redesign.

What would challenge it: crawl frequency staying flat through 90 days of steady additions. That would mean crawlers are budgeting these sites on something other than freshness, and the useful question becomes what that something is.

The result I am least prepared for, and the one worth watching: crawler attention rising while impressions stay dead. That would separate two things that usually get discussed as one, being read by a machine and being shown to a person. AI.AS400Software.com is already hinting at that shape, with crawler traffic holding steady and impressions drifting down. Five days is not evidence. Ninety days across ten sites might be.

One caveat worth stating plainly, because it applies to every number above. The 28-day crawler baselines come from verified classified crawler requests, not raw log lines, and a server request proves only that a request reached a URL. It does not prove indexing, a query, a citation, or a reader. Nothing in this experiment establishes causation on its own. It establishes whether a correlation exists that is worth chasing.