83 Tweets, 658 Impressions, Zero Clicks

01Eighty three posts and a zero

The first stretch was 83 posts on one social platform. They went out on schedule, each one written to a sensible template: a claim, a link, a few hashtags. Together they earned 658 impressions and zero clicks. Not a low number of clicks. None. Not one person followed a single link to a single page.

The part where I kept going

I read that as a volume problem, which is the most tempting misreading available. So the posting grew to five brand accounts and four posts a day, and by the time I audited it there were 475 posts. They had earned 1,953 impressions between them, and the median post was seen once. The median post was seen by one account, and I cannot rule out that the account was mine. The only posts that drew any reply at all opened on a hard number or corrected a common myth, which is worth remembering, and two posts about watching video on mute reached a few hundred people each, which is not a strategy.

Around the last week of August the reach fell further, and the follower counts of three accounts were cut roughly in half in what looked like a platform wide purge of inactive followers. That was the moment to stop. I did not.

02The suspension I built myself

On the morning of 15 September two of the accounts were suspended for inauthentic behaviour. One came back on appeal with a finding of no violation. The other appeal was denied. I paused all three hundred posts still in the queue the same day, then went through the data looking for what the platform had been looking at.

Coordination looks like fraud from the outside

No two accounts ever posted the same text, so this was not duplicate content. It was shape. All five accounts posted through a single developer app from a single server. There were 111 separate minutes in which two or more of them posted at exactly the same time. Every post carried a link and about three hashtags, four times a day for a month, and none of the accounts ever replied to anyone. Five accounts that post in the same minute and never talk to anyone look like one operator, because they are one operator. I built a network and was surprised when it was read as one.

I want to be careful about the other half. I do not know how the platform weighs any of those signals, and the account data cannot tell me. What it can tell me is that every pattern a reviewer would call automated was present, and that the fix is not better copy.

03Wrong lever, not wrong words

The honest summary is the one I wrote in my notes on day one and ignored for three months. It was never a content problem. It was a distribution problem, and social was not distribution. A post on a platform nobody follows is published to nobody. It has no audience to distribute it and no index that treats it as a source, so it cannot become the thing a model repeats.

Where the citations actually came from

In the same months, the same products were being cited by AI assistants thousands of times, and every citation I can trace landed on a page. Pages that answered a question in their opening lines, in indexes that assistants search, fetched by crawlers that never stopped reading. Nothing I can measure connects a single post to a single citation. A post expires in a day. A page that answers the question is still being quoted next year. That is the asymmetry I should have started from, and it is the one the first month of this blog already showed, when the machines arrived long before any readers did.

What I would keep

Social is not worthless. It is a place where people talk, and that is exactly what the automation removed. A founder replying in threads, answering questions under their own name, occasionally pointing at a page that genuinely answers something, is doing distribution. A scheduler firing identical shapes from five accounts is doing the opposite, and teaching a platform to distrust the brand. If the only reason an account exists is to post links, it is a link farm with a profile picture.

Do not chase the outliers

The worst move now would be to study the two posts that reached a few hundred people and try to repeat them. Outliers in a dataset this thin are noise that happens to look like a lesson. When the median is one, the best post is not a pattern. It is luck with a timestamp. Measure whether a channel sends humans or earns mentions, and if the answer after three months is neither, stop, however disciplined the schedule looks.

The automation stays off. What replaces it is smaller and slower: one account per brand, posted by hand, replying more than it announces, and no link unless the page earns it.

Tomorrow: the 500 character support message that got a vanished domain looked at by a human.