On-Brand Content at Volume

AI collapses the cost of producing content, and that much is real: a draft that used to take a writer an afternoon now costs cents and seconds. The cheapness is also the trap. When the marginal cost of a draft goes to near zero, the thing that used to be scarce stops being the bottleneck, and a new one takes its place: not whether you can produce a hundred pieces, but whether anyone can tell your hundred apart from everyone else's now-equally-cheap hundred. Most "AI for content" advice optimizes the part that already got cheap and ignores the part that got expensive. On-brand volume is the skill. Volume on its own is just a faster way to arrive at the average.

The reason it lands on the average is structural, not a quality bug you can prompt away. A model predicts the most probable next token, and most-probable means most-typical, so a request with nothing specific in it returns the statistical center of everything written on the topic — which is, by construction, what every other person's model returns from the same empty request. The output reads as plausible and competent and says nothing only you could have said. That is the failure underneath both of the symptoms operators actually get punished for.

The distribution symptom is where the abstract risk becomes a number you can verify. Google's May 2026 core update produced a clean dose-response between how much of a site was scaled programmatic AI content and what happened to its traffic. A SaaS site running roughly 70% programmatic pages dropped about 78% of its traffic. One at 45% programmatic dropped about 52%. An editorial-only B2B blog on the same update gained about 4%. The relationship runs one way: the more of your output was the average, the harder you fell, and the penalty was sitewide — Google read the programmatic ratio as a domain-level quality signal, so the editorial pages on a heavily-programmatic domain took collateral damage from the slop sitting next to them. The lesson is not "AI content gets penalized." Editorial content authored with AI did fine. Ungated volume that regressed to the mean is what got demoted, and it dragged the rest of the domain down with it.

So the operator question is not how to generate more, it is how to keep the model off the path to average while it generates. Two moves do almost all the work, and the order they run in is the whole point.

The first is voice. The instinct is to draft fast and then edit the brand back in at the end, the way a human editor would. That does not work, and the reason is mechanical rather than a matter of effort. The defect the model produces is not surface noise sitting on top of good structure that you can sand off — it is baked into the sentence shapes, the cadence, the rhythm of how each paragraph opens and resolves. A late edit pass that swaps words keeps that skeleton intact, so you can change every adjective and still have something that reads exactly as machine-shaped as before. Voice has to be present in the call that writes the section, not applied to the section after it exists. The form that actually holds is a voice profile treated as code: scrape ten to twenty of an author's real published pieces, extract the patterns that recur — average sentence length, how they open (a war story, a contrarian claim, a number), the moves they always make, the words and constructions they never use — and inject that profile into every section call as a hard constraint. Adjectives like "bold, witty, human" do nothing here; they are unfalsifiable and the model ignores them. A do-not list and concrete before/after exemplars are falsifiable, and falsifiable is what the model can actually obey.

The second move is to stop asking for the whole piece in one shot. A single long generation drifts in a way that is independent of how good the model is — past the first thousand-odd words it loses track of what it said at the top, starts repeating itself to fill the word budget, and treats its own weaker earlier text as ground truth. The fix is to break the work down: write an outline first, turn each section of that outline into its own brief carrying a target claim and the evidence it should rest on, generate the sections against those briefs, then run one low-temperature pass to stitch them together. Per-section briefs beat one big prompt because the model only has to hold one section in attention at a time, and because any section that comes out wrong can be regenerated on its own without re-rolling the whole piece. The voice profile rides into every one of those section calls. That is what keeps eight separately-generated sections from reading like eight different writers — the stitch pass at the end can prevent drift it never had to repair, but it cannot install a voice that was absent from the writing.

A voice profile is only as good as the voice it encodes. Scrape it from ten generic pieces and it just encodes generic more precisely — garbage origination in, polished average out. If the underlying writing has no stance, no proprietary observation, nothing the model could not have guessed, no profile rescues it, and the volume only multiplies the blandness across more pages.

The per-section pipeline also has a cost it is honest to name. It is real engineering, not a single prompt, and below a certain volume it is not worth building. For one carefully-considered essay a week, a human writing with a voice file open beats the whole apparatus. The pipeline earns its complexity when the volume is high enough that one-section-at-a-time regeneration and the stitch pass save more than they cost to maintain.

And the human edit pass is load-bearing in a way that is easy to let slip, because the failure is invisible. The defect that survives every other gate is plausible emptiness — prose that is fluent, on-voice, factually unobjectionable, and says nothing. It passes a skim, which is exactly why a skim is not enough. Editing AI output is a harder skill than writing fresh, not an easier one, and a team that assumes AI saves edit time tends to ship the wordiness and end up slower per finished piece, not faster. The edit pass is where a human injects the concrete detail the model structurally could not know. Drop it to chase throughput and you are back to producing the average faster.