For six months I let a pipeline write and publish articles to a commercial store I own, without reading them first. Not “draft for review”. Published, on a schedule, while I was asleep or doing something else.
Here is what Search Console says about the window from 1 February to 31 July 2026, measured against the six months before it.
Clicks grew roughly six-fold. Impressions grew 42-fold. Average position settled at 7.7.
The 42x is the number everybody quotes back to me. It is also the least interesting number in this post.
The number that actually matters
All of that traffic lands at a click-through rate of 0.58%, from an average position of 7.7.
Position 7.7 should do considerably better than that. The pipeline is very good at getting a page in front of people and weak at earning the click. Impressions at that volume with a click rate that low is not a content problem, it is a packaging problem: titles and meta descriptions written to satisfy a quality gate rather than to win a click in a crowded result.
The pipeline optimises for the thing it can measure at write time. Nobody clicks at write time.
This is the honest shape of the result. Volume was the easy part. I would trade a large chunk of those impressions for a click rate that reflected the position.
What the gate actually does
Every article runs a fixed chain before it can go anywhere:
factcheck → analyze (≥80) → seo-check → geo → brand-voiceanalyze scores the draft out of 100 across content quality, SEO, E-E-A-T, technical
elements and citation readiness. Under 80 and the article does not publish. A second
failure abandons it to draft rather than retrying forever, which matters more than it
sounds: without that rule the pipeline will happily burn an afternoon polishing
something that was never going to work.
There is a separate policy layer on top for where a passing article lands. 85 and above goes live. 80 to 84 goes live and posts a flag to Slack so a human sees it. Anything below 80 stays a draft. Pillar and hub pages never auto-publish regardless of score.
The rule that does the most work is the dullest one: no fabricated statistics and no unsourced claims. A generated article will produce a confident, plausible, entirely invented number every time you let it, and on a commercial site in a regulated-adjacent category that is not a stylistic problem, it is a liability.
The measurement I did not build
The gate writes its scores to a per-article JSON file that the pipeline consumes and then discards, so the rejection rate was never persisted. The mechanism is precise and unreportable, which is exactly the statistic that would make this post better.
It is a two-line write to a log directory, and it is the first thing I would build into the next one. A pipeline that works is the easiest possible reason to skip instrumenting it.
What I would not change
Publishing without per-article review was correct, and it is the part people expect me to walk back.
The reason it works is that the review happens at the gate rather than at the article. Once you trust the chain, reading every output is theatre. Once you do not trust the chain, reading every output does not save you either, because the failure mode of generated content is not obvious badness, it is confident mediocrity that reads fine at a glance.
The useful question is never “did a human look at it”. It is “what would have stopped this if it were wrong”, and the answer has to be a mechanism, not attention.
What actually generalises
Three things carried over to every pipeline I have built since.
Measure the thing you are optimising, not the thing that is easy to count. The gate scores drafts. Nobody reads a draft score. Position and impressions were never the goal either; they were the proxy I could see at 8am on a Monday.
A gate needs to be observable or it is decorative. A control you cannot report on is indistinguishable from no control at all.
Volume is the cheap half. Getting an agent to produce 121 competent articles is a solved problem and was the least interesting part of six months. Getting them to earn a click at position 7.7 is the part still open, and it is where the next version of this lives.
Sources
- Search Console performance report metricsGoogle
- Creating helpful, reliable, people-first contentGoogle Search Central