The numbers nobody publishes

This studio is forty-nine days old and every number it produces is measured with the same tooling. This page is that set, pulled the morning it went up, including the ones that make the operation look worse.

Nobody publishes this set. Builders publish the numbers when the numbers are the story, and go quiet when they are not. The quiet part is more useful to read, so here it is.

What exists

The operations repo's first commit is 2026-06-26. The first product repo's first commit is 2026-07-05. Today is 2026-08-14, so the whole thing is forty-nine days old and the oldest product is forty days old.

The roster holds 91 entries: 35 products, 30 infrastructure repos, and 25 archived. Of the 35 products, 23 carry a price and 20 carry a live subdomain in the roster. They cover accessibility reporting, invoice chasing, document extraction, pharmacy tracing, agent-skill indexing, App Store readiness, payroll reconciliation and about a dozen other things that have nothing to do with each other.

Every number, and where it came from

Everything below was pulled on 2026-08-14, the day this went up.

The PostHog rows are one HogQL query with this machine's own addresses excluded, and the raw row is the same query without that exclusion. The AEO rows are a nightly probe that asks five engines the questions a buyer would type for each of 23 products. The Supabase rows are social_account_metrics and social_post_metrics, latest row per account.

That repository row said 8 when this page first went up. A ninth went public the same afternoon: deferless, MIT, the fail-closed gates this estate runs on its own agent output. It has zero stars, so the number underneath it did not move. The row is corrected here rather than quietly, because a page about publishing the losing numbers does not get to edit one silently.

Half of my own traffic was me

The traffic row deserves its own paragraph, because for six weeks it was wrong and nothing said so.

Every product on this estate is screenshotted, gated and probed by automation running on one machine. Those visits are pageviews. PostHog's own internal-user filter could never catch them: the shared analytics component runs with person_profiles: 'identified_only', so a self-visit never creates a person profile to hang the internal flag on, and the cohort behind that filter has zero members by construction.

The query underneath the morning report filtered on $host NOT ILIKE '%localhost%' and nothing else. An audit on 2026-08-12 found that 764 of 1,674 production pageviews in the preceding fortnight came from this Mac. The damage was not spread evenly, which is what hid it. The apex was 38% self and still had real readers. CiteRank was 97% self. ShipWall 96%. BreachProbe 94%. StackTab was 100%. Six products reported one or two visitors for the week and every one of those visitors was a daemon.

One automation browser profile accounted for 531 pageviews across 29 different products in twelve days. PostHog stitches a browser profile into a single person, so it added exactly one phantom visitor to twenty-nine per-product rows at once.

The fix is an address filter applied at query time, which cleans the history retroactively. Today's honest figure is 1,065 real pageviews in fourteen days against 2,124 raw. Half the traffic to this estate is still its own machinery.

Zero citations out of ninety-two

A nightly probe asks five answer engines the questions a buyer would actually type for each of 23 products. On 2026-08-14 it asked 115 queries, got 92 answers back, and found zero citations of anything on this estate.

Not a low rate. Zero.

The engines answered every one of those questions with somebody. Asked about shadcn registries they cite ui.shadcn.com and shadcnblocks.com. Asked about invoice chasing they cite chaserhq.com and paidnice.com. Asked about ADA Title II reporting they cite deque.com and levelaccess.com. The pages exist, they are indexed, they have schema, they have llms.txt, they load fast. None of that is the mechanism. Roughly 85% of brand mentions in AI answers originate on third-party pages rather than on the brand's own site, and every hour spent tuning my own landing pages was spent on the wrong side of that number.

The retention number that changed what I did today

751 followers across 25 accounts sounds like a small audience. The split inside it is the useful part.

Same machine, same cadence, same content pipeline, forty-eight times the retention.

The sharpest single case is a breaking-news wire I run on Threads. It has published 1,676 posts, they have been seen 516,323 times, its best post reached 2,885 people, and it holds eleven followers. That is 0.021 followers per thousand views, roughly three hundred times worse than the account with a name on it. Four short-video accounts have about 109,000 views between them and two followers.

The content was being consumed. It was consumed by strangers who had no reason to remember a faceless brand they would never see again, and there was nothing on any of those surfaces telling them who was behind it.

So as of today the wires publish under one masthead with an editor on it: kynth.studio/wires. One editor, one method, one place to send a correction. None of the accounts were retired and none of the cadences changed. The only thing that changed is that the work now has a name attached to it.

What is still wrong

The citation number will not move by editing my own pages, which is the thing I have been doing.

1,065 pageviews in a fortnight is not a market, and no amount of work on the products changes that number. The distribution is the problem, and the distribution is measured here for the first time.

The follower split says anonymity has cost effectively all the retention on views already earned, and the masthead is a fix for the wires only. The short-video lanes still have no answer.

Every figure above can be re-pulled. The traffic exclusion, the citation probe and the social tables all run nightly, and the next set will be worse or better on the same instruments. The gates that hold the rest of this estate to the same standard are public and MIT at github.com/kyisaiah47/deferless, including the one that refuses to let a finding be deferred.