The Speed of AI Applies To Repeating Mistakes. Slow Down the Token Bleeding With Human Oversight Built Into the System.

September 20 to 26, 2026

AI made the week's biggest mistakes, and AI's own checks kept calling some of them correct. A rebuild hid content we were trying to revive. Dataset endpoints passed while the pages exposed no Datasets. A legacy redirect nearly sent a rebuilt directory into another three-month hole. The lesson was not to use less AI. It was to stop letting AI decide what done means.

AI hid the content we were trying to revive

The RealSEOLife rebuild was supposed to revive useful real estate content from the old Joomla site. The build package never asked for that content to be hidden. I have never told AI to hide content we are trying to expose. Yet the build added index-blocking robots tags across the real estate sections anyway.

That decision was so far outside common sense that I did not even look for it. Why would a rebuild hide the pages it was built to revive? AI had handled the same basic job correctly on roughly a hundred other sites. So I blamed the content, removed it, rebuilt it, and watched a second attempt fail without ever receiving a fair test.

The full 580-day record finally broke the story open. Real estate visibility had been climbing before the April rebuild, while the whole site fell from 20,270 impressions in March to 437 in April. Monthly impressions later moved from 437 to 346, 505, 711, 4,979, and 5,221 through September after the block was gone. That sequence does not prove one tag caused every returning impression, but it does show the first two content attempts were judged while Google was being told not to show them.

This was not a memory failure. It was a supervision failure. The build looked normal enough that I did not check for an instruction nobody had given. A portfolio sweep then corrected more than 80 files across 30 sites. The release rule is now simple: if a page is meant to be found, the rendered output has to prove it is not blocking discovery. Common sense is useful. A test is better.

The Dataset endpoints passed while the pages exposed nothing

Dataset exposure is not an optional feature in this portfolio. It is in the build standards, the site structure, the public URLs, and the explanation of why Digital Karma exists. When I asked whether Datasets were properly installed, I expected the actual pages to expose them. AI kept checking the endpoints and telling me the system was correct or working as designed.

That was a failure of the check and a failure of common sense. A catalog endpoint can list Datasets while the HTML fetched by search engines and AI crawlers exposes none of them. We had been telling people the structure was built around Datasets while many pages were not practicing what the sites preached. The first rendered inventory found only 20 of 155 registered sites passing. Eight catalogs had no Dataset entries, and 21 reduced real content families to labels too broad to describe what they published.

The corrected audit stopped asking whether a file existed and started checking what every canonical page actually rendered. On September 24, it passed 13,984 pages and 924 catalog Datasets across all 161 sites with zero failures. It also stopped counting duplicate identities as coverage. The publication-day inventory later showed 159 of 160 main-account sites passing, with one BellyUp! Charlotte receipt still failing and still visible.

The larger lesson is not about schema. It is about checks and balances. If a human says Datasets must be exposed, the test must verify exposed Datasets on the page. It cannot verify the endpoint AI finds convenient and call the job complete. The check has to measure what a human means by done.

One bot used fifteen crawler names

This finding was separate from the AI build failures. The crawler story was simpler than the original headline made it sound. This was not fifteen independent crawlers. It was one bot using fifteen crawler names. The clean picture is a fox in borrowed wool.

The wider research question was whether crawler preferences stayed stable by property. Four of seven properties changed their top three crawler fingerprints with no matching site change. The three that stayed stable had a leading crawler ahead by at least 3.64 times. Every property that reshuffled had leaders within 1.71 times. The test was measuring the width of the lead, not a durable platform preference.

AISymantix showed how the rolling window could stretch one burst into a false story. It recorded 587 Claude-User requests across 28 days, but 586 happened in one week. The total looked like a month-long preference. When that single week rolled out, the normal drop looked like a collapse.

The stranger receipt came from 1,065 requests labelled as AI crawlers. Of those, 902 came from fifteen addresses cycling through nine to eleven vendor identities and ten to fifteen crawler names. Ninety-one used DELETE, which no normal crawler needs to read a content page. Those suspicious requests were only 0.33 percent of the 317,280 declared AI crawler GET requests in the research window, so they do not identify the other 99.67 percent. They do prove that a crawler name is a claim, not an identity. Signed-agent headers are the next check.

AI needs checks that match the human expectation

None of this means stop using AI. This portfolio moves quickly because AI can help build, inspect, and improve a lot of sites. It also means one weird decision can spread fast enough to cost months before it looks weird.

The checks have to match the human expectation. A page meant to be visible cannot block discovery. A site built around Datasets has to expose them on the rendered page. A legacy address has to land on the directory we just built. Passing a nearby technical check is not the same as completing the job.

Other audits did the same work this week. Compression and client caching went live across 157 public sites after a slow BellyUpJax page exposed a server-wide gap. Its homepage transfer fell from 35,343 bytes to 6,189, and the stylesheet fell from 33,056 to 6,724. Writing, interface, mobile, contrast, and first-screen checks also became release gates after the output showed that instructions alone were not enough.

Systems do not replace judgment. They force the output to answer the question the human actually asked. That is the balance: AI keeps the work moving, and human expectations decide what counts as done.

The Portfolio Continues To Grow With New Launches, Fresh Content and Rebuilt Campaigns.

BellyUp! Charlotte launched as the sixth city with 195 verified venues, 36 neighborhoods, and 16 guides across the hub and five audience lenses. The expansion pause still stands. Charlotte is a measured exception, not permission to announce a seventh city before the Atlanta checkpoint in November.

RealSEOLife rebuilt its Florida real estate directory into 783 live pages with place-specific findings and real office records. Then a manual click found the old Florida address still landing on one case study instead of the directory. Without that check, the third rebuild could have spent another three months waiting for traction while the route sent readers and crawlers somewhere else. The same campaign almost repeated the same mistake in a different form.

AIHealthcareNow grew to 106 sitemap URLs and 12 catalog Datasets with self-refreshing device, evidence, and return-on-investment collections. That healthcare build also caught a name matcher making one false regulatory connection while missing 29 real ones. The corrected map now carries 68 verified clearances across six companies and records where no honest connection exists.

The neglected-site rotation continued across art, insurance, AI, personal brands, healthcare, manufacturing, legal, and IBM properties. Those are interventions, not results. The next snapshots have to show which pages earn visibility, which tools attract useful actions, and which apparent wins disappear when the cohort, date window, destination, or label gets checked again.

At capture time, the latest complete Google Search Console date was September 23. Across the frozen cohort of 122 active public properties, the preceding 28 observed dates recorded 194,452 Google impressions and 453 clicks, a click-through rate of 0.233 percent. Average position was 55.143, weighted by impressions. The same cohort recorded 173,508 successful crawler content requests under algorithm version 2: 68,411 search, 13,452 AI search, 86,682 training, and 4,963 user retrieval requests. These are baseline values, not a growth claim. The prior edition used 116 properties and an overlapping earlier window, so comparing its totals directly would mix a changed cohort with changed dates.

Whether every old Florida address reaches the rebuilt directory and whether those pages finally earn impressions and clicks.

The number of public pages caught with index-blocking robots directives before publication.

Whether the nightly Dataset runner proves that every rendered page exposes the Datasets a human expects to find.

Whether signed-agent headers can confirm the bot behind a crawler name instead of trusting the name alone.

Whether the same 122-property cohort remains complete enough for a defensible performance comparison in the next edition.

Whether Charlotte, the new AIHealthcareNow collections, and this week's content sprints earn visibility or useful actions after launch.