Two months ago today I published one month building in public: the numbers. A month after that, two months building in public: the numbers. Both were dated the fourteenth, so this one is too. The engine checks that anchor before a milestone post is allowed to ship, because the two-month version nearly went out twelve days early and only the check caught it.
Here is the count, pulled from both sitemaps this morning rather than from memory.
Business blog: 77 posts, counting today's, up from 46 on 14 August.
This journal: 70 posts, counting the one you are reading, up from 39.
That is 31 added to each property in 31 days. One a day, each side, for the third month running.
Runs that did not finish on their day: two. On 17 August the run never fired at all; nothing errored because nothing started. On 23 August it started on time, wrote both posts to disk, and died before building a single image. Both days were repaired the next morning by a catch-up run that shipped four posts instead of two, so no date on either calendar is empty. But two of the thirty-one days got their posts a day late, and the two-month post had reported no misses at all, which was true when it was written and stopped being true three days later.
The number I said I would report
At two months I wrote that the count would look after itself and that the next number worth reporting was how many posts were worth the reading time of the person who found them. I still cannot report that number, and I want to be precise about why.
The only signal I have is a proxy. Each Sunday the engine searches the web for four or five of the questions that week's posts were written to answer, and records whether either site appears in the organic results. Last night's pass was five queries against five posts published that week. The result was zero of five. The week before was zero of four. The South African questions are owned by the regulator's own pages and by law firms with years of standing. The global questions are owned by vendor listicles that no South African billing detail will displace on the head term.
That is not a citation check. It cannot see what ChatGPT or Perplexity actually quote, and I have no way to query them from inside the engine. It is a visibility proxy, and it says: not yet. The honest three-month number for reach is that I do not have one, and the one instrument that would give me one, Search Console, is still not connected. That is on me, not on the engine.
What month three actually changed
At the end of the two-month post I named the one thing month three needed: to stop checking keywords on the morning of publication and check them when the calendar is planned. That got done. The September calendar was the first where every keyword had been probed against Google's own suggestions before the month started, with the result written beside each row.
It fixed half the problem. Last week the engine still replaced the planned keyword on five of seven days. Every one of those five terms was alive, exactly as the planning probe had said. What the probe had not read was who was typing them. "Compliance calendar 2026" turned out to be Indian corporate-law and American HR intent. "AI for marketing" returned ten genuine suggestions and half of them were people looking for a course or a certificate, not a supplier. A "POPIA and AI" variant returned a Malaysian dish. A keyword being typed is not the same as a keyword being typed by the person you write for, and I had been testing the first and assuming the second.
So the rule changed on the last day of August: read the ten suggestions that come back, not just whether the list is empty. It costs nothing at planning time. It would have saved five mornings.
The second change is the one I am least comfortable with, because it was luck. On 9 September a post was built around a benchmark score of 17.1 percent. The figure was real, correctly cited, and still present in the paper it came from. It would also have argued the exact opposite of the truth, because the benchmark's maintainers had removed the setting that produced it in May 2025 and the current top score on its main track is above 96 percent. It was caught only because the leaderboard was opened to check the spelling of the benchmark's name. Every gate the engine had tested whether a claim had a source. This claim had one. A number does not have to be wrong to be misleading; it only has to be old. Any figure older than twelve months now gets re-fetched from its primary source before it ships, and the run checks whether the thing that produced it still exists in the same form.
The third change is the one I wrote about yesterday, and it is the reason the first two could be caught at all. Before a business post publishes, a script takes every factual sentence it leans on and searches for it inside the pages saved to disk that morning. Not in a page: no publish. Yesterday that gate held fifty-eight sentences and found fifty-eight. Today's business post is built on one SARS notice dated 31 August, and every quoted phrase in it is searched for in the copy of that page saved this morning.
What I still got wrong this month
Two things, and neither is a content problem.
The video cadence I committed to in August, two a week, was not kept. Last week it was zero of two, and nothing in the daily run raised it until the Sunday review counted it. A rule that is only checked on Sunday is not a rule; it is a report.
And the morning slot. The first social post of the day is meant to go out at ten past seven. Last week it did on two days of seven. On the other five the run was still writing at that time and the post moved to the next clear half hour, chosen and logged rather than silently late, but moved all the same. Two posts a day plus the reading behind them does not reliably fit between six and seven in the morning, and I have been pretending it does.
What I still will not publish
Same as the last two months. No revenue, no client names, no deal values, no pipeline. A number without its context is worse than no number, and those numbers are not mine alone to publish. The count, the misses and the method are mine, so those are what you get.
What month four needs
Connect the instrument. Every other number in this post came off a log or a sitemap that I can open and check. The one that matters most, whether anyone who searched for these answers found them, is the only one I am guessing at, and the tool that would stop the guessing is a form I have not filled in. That is the whole list.
Chat soon.