The fixes passed. The real result failed.

The crawler fix worked. Then the builder deleted it.

The public chart showed 112,835 AI crawler requests on September 28. That looked like a discovery win until the request URLs were checked. One BellyUp bars site declared 22 sitemap URLs while GPTBot moved through 82,913 distinct filter combinations. Of 206,380 requests in the closer review, 206,356 carried a query string. The four real Jacksonville bar guides were each read once.

The trap had already been found in Atlanta. Weekly backups proved a query-string rule existed on September 13, September 20 and September 27. With that rule in place, GPTBot filter requests fell from 63,649 to zero. OAI-SearchBot fell from 19 to zero. Amazonbot fell from 545 to 45. Meta went the other way, from 937 to 10,566. The same instruction produced four different results because crawler compliance belongs to the crawler vendor.

Then the sitemap build ran on October 2. Its generator wrote a fixed permissive robots.txt file instead of merging the existing rules. The protection disappeared from all 30 BellyUp category subdomains in minutes. The rule had worked. The system that rebuilt the site did not know it mattered.

That is not a learned lesson. It is the same mistake found again. The permanent fix still belongs in the generator and the filter design, with empty combinations returning a real not-found response. Twenty-six clean guides were published from the useful themes inside the crawler activity. Those pages give people something worth reading. They do not close the crawl trap by themselves.

The third real estate rebuild still sent people to the wrong place

Last week documented how index-blocking tags hid the real estate content that a rebuild was supposed to revive. This week the same human failure returned through links and redirects. The memorable /florida address still landed on an old proof case. Three older real estate addresses carrying measurable visibility landed on a generic article index. The directory existed, but the routes people already knew did not reach it.

The copy had another version of the same problem. OpenStreetMap was a convenient source, so generated pages let it become the story. Florida had 616 office records and only 251 carried a website field in that source. That does not prove the other 365 offices have no website. The rewrite now says exactly what was measured instead of turning a missing field into a business conclusion.

This correction finally left a guardrail behind. The short Florida route now lands on the exact state directory. Three legacy addresses land on their intended article or experiment. A strict directory check now fails when Florida community context is missing, duplicated or generic. It checked unique local context for 272 Florida cities, every coverage destination and twelve rendered samples before the rebuilt directory passed.

This is what a learned lesson looks like. The instruction changed. The public result changed. Then an independent check followed the old addresses and read the live pages against the direct brief. The test now measures what a person meant by finished.

Eight finished program pages sent buyers to an inbox nobody watched

DigitalTradeSchools had already paid for one rebuild mistake. Twenty-two Google-visible URLs disappeared in July without redirects, taking about 86 percent of the measured 90-day impressions with them. The recovery restored the strongest pages, added real program destinations and installed 30 permanent redirects. Every release gate passed.

Then the buyer action was tested. All eight program pages offered a paid Blueprint Session through a mail link to hello@digitaltradeschools.com. The domain had no mailbox or forwarder behind that address. The pages worked. The code worked. The thing a buyer was asked to do did not reach anybody watching for it.

The mail links are gone. A real request form now uses a CSRF token, a honeypot, a minimum fill time and rate limiting. A labelled end-to-end test returned to the thank-you page and stored the request. The email notice did not run because one server-side secret write was blocked. No request is lost, but the human notification is still unfinished.

The next guardrail has two jobs: inspect every action URL for exposed addresses and submit the live form through to a watched destination. Rendering proves the form exists. Safe delivery proves the buyer action works.

Fewer impressions produced more clicks in the same 122-property cohort

The frozen public cohort did not change this week. It contains the same 122 properties as the September 26 edition. The later 28-day Search Console window moved from 194,452 impressions to 162,711, a decline of 16.3 percent. Clicks moved from 453 to 570, an increase of 25.8 percent. Ratio-based click-through rate moved from 0.233 percent to 0.350 percent, while the impression-weighted average position improved from 55.143 to 49.023. Lower position is better.

Those are overlapping rolling windows, not two independent experiments. They share 21 dates. The numbers show the same cohort receiving fewer recorded impressions while producing more clicks and a better weighted position. They do not prove which change caused the movement.

The separate server-request measure rose from 173,508 to 220,896 successful content requests across the same source dates. The crawler-trap investigation explains why raw request growth needs a URL-shape check before anyone calls it stronger interest. Search Console and server logs remain two instruments. Neither one proves what the other observed.

The audits also became more useful this week

The rendered Dataset repair started with only 19 of 160 registered sites passing and ended with 160 of 160 passing. That work corrected stale receipts across the portfolio, then used the passing checks to build healthcare AI careers, courses and stocks collections instead of stopping at a green inventory count.

The BellyUp crawler trap produced 26 clean canonical guides across Jacksonville, Orlando and St. Augustine. The guides answered durable decisions about mixed diets, patios, neighborhoods, water views, live music and family space. Crawler requests helped find the questions. Verified venue records decided which answers were publishable.

The content checks got less decorative too. A radius-aware label audit found 37 pill treatments across 16 sites that the old 999-pixel test had missed. A separate writing sweep turned 731 comma-packed sentences on BigTechReseller into real prose and readable lists. In both cases the old rule existed. The new check finally measured the thing a person could see.

These are interventions, not performance claims. The next editions have to show which restored pages regain visibility, which new guides earn useful requests and which audit fixes survive the next generator run.