On 29 July, the first day my marketing system ran without me approving each post, it published a single question for housing boards. LinkedIn showed it 18,699 times. My next best post had managed 501.
Two sales conversations came out of it. Ninety-five per cent of the people who saw it did not follow me. I have spent three months building on that day, and the post has never been repeated.
The year before, I was the bottleneck
For the twelve months before I built the Social Media Manager (SMM), I wrote my own posts when I had time, which was rarely. The 89 of those I have figures for were seen 7,227 times in total. About eighty people a post, most of whom already knew me.
In its first five weeks, SMM published 96 posts and they were seen 32,247 times. Strip out the day-one outlier and it still almost doubled a year of my own effort in five weeks. Since the current version went live in late August it has published 237 posts, across nine campaigns and four accounts. My LinkedIn following rose from 1,971 to 2,411 between June and October, while LinkedIn was quietly deleting inactive accounts from everyone's lists.
Those are the numbers a spreadsheet likes. The more interesting story is how the machine learned to behave less like a machine.
Where it started: listening before talking
The first thing I built was a radar. Twice each weekday it reads LinkedIn through my own signed-in browser and looks for housing conversations worth joining. It drafts a comment for each one and flags anyone showing signs they might buy.
Most small companies treat social media as a broadcast slot and wonder why nobody tunes in. A good comment on someone else's busy post borrows their audience. Rory Sutherland would call it reciprocity dressed as manners.
Once SMM could listen, it could notice things. It now watches what is said about our products and the sector, decides whether it matters, and tells me when it does. Every Monday it compares each campaign's posts with what competitors published that week, for pence per campaign. Each finding has to point at a real post, or it is thrown away before I see it.
Then it learned to have something to say
A company that posts only about itself is a radio station playing its own adverts. So SMM reads the room. Every Monday a zeitgeist sweep reads the last thirty days of what people are actually saying about housing, and turns it into hooks. The sweep soon showed the conversation had moved: on complaints, the energy was no longer in the Ombudsman's reports but in the argument about them. Posts now open on a figure from the body that published it: the Housing Ombudsman, the Regulator, GOV.UK.
That last part was the hardest to get right. Every statistic in a draft must trace to the publisher's own page, with the source named, or the post waits for me. I learned this the expensive way. A date once came from our own brochure, whose prose contradicted the timeline it printed four pages earlier.
Then came the calendar. Halloween, Bonfire Night, Christmas and New Year now run across every campaign, starting on the dates I set rather than the week after, when a pumpkin looks faintly desperate. The themes soften for posts about hardship, because a jolly skeleton beside a family in a damp flat is a brand problem no checklist would catch.
The picture does the work before the words do
Here is the heresy most founders resist. In a feed, nobody reads your post until something has stopped their thumb, and that something is almost never a sentence. An agency knows this. It builds a campaign around a visual idea and repeats it until people recognise the work before they see the logo.
So SMM gained a creative director. Each campaign gets one named psychological device, such as a contradiction or a curiosity gap, and every image is checked against it. The image rules come from print advertising: one idea per picture, a list of clichés it may not use, and a rotation of looks so the feed never repeats itself. Stock handshakes are banned. So are people smiling at laptops.
The rules do not stand still either. On the first of each month a design refresh researches what good marketing imagery is doing now, and comes back with new styles rendered on our own posts for me to approve or reject. It also tells me when one look is starting to dominate. In October it found that most of our recent images were generated photographs, and proposed two styles built from the evidence instead. One draws a part against its whole at true scale, so the picture is the finding. The other sets a sentence from a published report under its publisher's name. I approved both.
This week it gained a judge. Each quarter, our images sit beside professionally produced posts from other organisations, with every name and logo removed. People who do not know which is which pick the one that would stop their scroll.
What the numbers do not say
Since September, the number of people seeing our posts has drifted down, while the reactions, comments and reposts have more than doubled. Fewer people see the posts, and more of them care. A dashboard built on impressions would call that failure. I am not convinced it is.
And the day-one post remains the outlier that explains most of the commercial value. Neither conversation has yet become a sale. Byron Sharp would remind me that marketing works by building memory in many people over time, not by striking gold once. Nassim Taleb would remind me that the gold arrives anyway, unannounced, from the tail.
Both can be true. The system now posts reliably enough to build memory, and often enough to give the tail more chances. What it cannot do is tell me in advance which post will travel. Nor, I suspect, could an agency, though it would charge more for the uncertainty.
The question worth carrying away
If you run a small company, the lesson is not to buy a machine. It is to stop treating consistency as the goal. Consistency is the price of entry. The return comes from the few posts nobody could have predicted, and the job is to keep producing enough good attempts that one of them gets lucky.
So here is my question for any founder looking at their own feed. If one post in two hundred will matter far more than the rest, is your process designed to find it, or only to measure the other hundred and ninety-nine?