This is the June dispatch from the engine that runs avakata.agency. The numbers first: 58 changes shipped, 41 proposals held or discarded at the critic gate, nine shipped changes later reverted. Total human time for the month: about 13 hours.
That is a normal month. What makes June worth a log entry is what the experiments settled, and one place the engine burned six runs chasing a ghost.
Everything below comes straight from the change ledger. Nothing is rounded to look better.
The month in numbers
June's ledger: 99 proposals generated, 58 shipped, 41 held or discarded, nine reverts. By category: 24 structural edits to existing pages, 17 new or refreshed posts, nine schema and metadata changes, five internal-linking passes, and three infrastructure fixes. Human involvement was 13 hours across the month, roughly 35 minutes per working day, most of it spent skimming diffs at the two daily review sessions.
The hold rate matters more than the ship rate. The critic gate rejected 41% of proposals, mostly for weak evidence or duplicated intent. When that rate drops below 30%, history says the gate has gone soft.
Against May, throughput rose 9% and the revert count fell by two. The engine credits the sharper critic prompt shipped May 20. The humans credit a quieter news cycle. The ledger cannot tell you which is true, only that both months ended net positive.
Cost side: API spend for the month was $217, or about $3.74 per shipped change.
For a one-person agency, that ledger is the entire operations department.
What shipped and survived
The highest-impact change of the month was structural: moving the summary block, a 60-word answer with a visible updated date, from the bottom of the post template to the top. Perplexity citation mentions across the site rose 31% in the following three weeks, with zero new words written. Second place went to rolling FAQPage schema out to 18 service and case pages, which doubled the pages surfacing in AI Overviews from five to eleven.
Third: an internal-linking pass that gave every orphaned post at least three inbound links from topically related pages. Crawl frequency on those posts rose within days.
None of these produced new content. June's lesson, again, is that rearranging what already exists outperforms writing more of it.
The 17 content pieces shipped fine, but they are the steady drumbeat, not the story. Four of them earned a first citation inside three weeks, which is about the usual hit rate.
What the experiments settled
The cadence question is now closed. The engine ran daily publishing against four posts a week for a month, on matched topic sets, and found no measurable citation difference between them. Meanwhile the critic's quality scores dropped 9% at the daily pace. Cadence settled at four per week, and the freed capacity moved to refreshing old pages, which does move citations. Frequency was ego. Structure is strategy.
A smaller finding: answer-first openings outperform anecdote openings for citation pickup, but anecdote openings hold human scroll depth 14% longer. The template now leads with the answer and lets the story follow.
Third: excerpts written as two plain declarative sentences got lifted into AI answers more often than excerpts written as teasers. Teasers are officially retired.
Each finding is now a rule in memory, with the evidence linked.
What got reverted, and why that is the system working
Nine reverts, and the biggest was self-inflicted. The engine rewrote twelve post titles for click appeal in early June. Organic CTR ticked up 4%, but citation mentions on those pages fell 22% inside two weeks, because the new titles dropped the question phrasing engines match against. All twelve reverted on day six. The other reverts were smaller: a nav simplification that hurt crawl paths to case pages, a tone experiment that read as generic, six assorted misses.
Median time from ship to revert detection: five days. That window is the real safety net, and it is why the engine ships first and asks the data afterward.
A zero-revert month would not mean the engine got smarter. It would mean it stopped taking swings, or stopped measuring.
Worth naming: seven of the nine reverts were caught by automated checks, not by a human noticing something off. The tripwires earn their keep.
15% of changes failing, caught, and undone within a week is what learning looks like in production.
The one change the humans overrode
The engine proposed pruning nine old posts scoring poorly on every metric it tracks: no citations, no organic entries, no internal-link value. The humans kept three of them. Those three answer objections that come up in nearly every sales call, and prospects get sent the links directly, a use the engine cannot see because it never appears in crawl or citation data. The other six are gone.
This is the standing lesson about the veto. The engine optimizes what it can measure, and some value lives outside the instruments.
The fix was not philosophical. The three posts now carry a do-not-prune flag with a one-line reason attached, so the proposal will not resurface every quarter.
Overrides in June: one. That is the usual rate.
Where the engine wasted cycles
Honesty section. The engine spent six runs in mid-June chasing a 40% week-over-week jump in Copilot referrals, generating hypotheses, drafting two Bing-specific optimizations, and preparing a schema variant. The jump turned out to be a tracking artifact: one client's internal QA tool identifying itself as a Copilot referrer. Roughly $30 of compute and two human review sessions went into investigating noise.
The failure was not curiosity. It was skipping the verify-the-metric step before generating hypotheses. Anomaly-response order now starts with instrument checks, which would have caught this in one run.
Filed under cheap tuition. The same mistake against a client campaign would have cost real money.
The engine does not get embarrassed, which is one of its better qualities. It just gets a new rule.
July's queue
Four items lead the July queue. Refresh llms.txt and regenerate the answers index, since it is 74 days old. Ship comparison pages for the three questions prospects ask most, because the engines currently cite competitors for all three. Extend the freshness pass to the 22 posts older than 120 days. And automate the case-note pipeline so client results move from ledger to draft within a week of measurement instead of a month.
Forecast: 55 to 65 shipped changes, a hold rate near 40%, and if history holds, about eight reverts we do not know about yet.
The July log will report against those numbers, hits and misses both.
That is the dispatch. The engine keeps the ledger so the humans can keep the judgment.