One desk per day

The company foldrun was built inside is a property inspection business in Victoria. Its marketing runs on seven workspaces — we call them desks — and one person who reads the headlines and decides the approvals. This is what that looks like on a normal week.

The desks

rank-desk measures. Every week it puts seventy-five target keywords through a localised search result each — Geelong, not Australia — and appends the position to a history file. Its headline is a count: how many rank, how many are absent, how many moved five or more places. A reporter turns the table into a paragraph. Cost per run: about fifteen cents in API calls and a dollar in model time.

link-desk acts on it. When rank-desk reports a page ranking with the wrong URL, or absent where it should be present, link-desk’s fix flow investigates, proposes internal links on specific addresses, and stops at a gate. A person approves some subset; the publisher applies exactly that subset. After 4 September it cannot apply anything else.

blog-desk listens. Search Console questions, forum threads that outrank the site’s own answers, and — since this week — queries that rivals rank for and the site does not. A translator re-angles each candidate as a question a buyer would ask; a validator judges it; the survivors become briefs. A person picks which briefs are written.

index-desk, aeo-desk, health-desk and gbp-desk each own one thing: which URLs the search engine has and has not indexed; how the site’s answers read to an answer engine; whether anything on the site is broken this morning; what the business profile should post next.

What the person does

Reads a column of headlines, most mornings, in about ten minutes. Opens two or three runs. Decides the approvals — usually one or two a day, occasionally none. Edits a paragraph in an agent when a headline was wrong in a way that will recur.

That is the whole job. The person is not an engineer and does not open a terminal. The agents are folders; the folders are in the dashboard; the paragraph is right there.

What it costs

Between them the desks run about thirty flows a week. The most expensive is the nightly leads pipeline at a median of $7.46 a night, most of it verification calls; the cheapest are the health checks at a few cents. The weekly total is under a hundred dollars in tokens and compute, which is less than an hour of the agency that used to do a fraction of this.

Every desk has a budget cap. Three have hit it. Each time, the run stopped between steps, the notification said which cap and by how much, and the fix was either a real bug (a translator that made one model call per row instead of two per run) or a real decision (that desk is worth more).

What we learned from running it

Every rule in the product came from one of these desks going wrong. The verdict headline, from a reporter that said “analysis complete.” Conclusions versus results, from a regex that could not find the first line. Every earlier group, from a reporter that invented numbers it had never seen. --only, from a hundred and forty links.

We could not have designed those rules in advance. We could only run the desks, read the headlines, and fix the paragraph. Which, it turns out, is the product.