Content generation.

What we run when we run content#

A research-to-publish pipeline operated by humans, with models in the middle. We do the research, we structure the argument, we draft against the firm’s actual voice and corpus, we edit, and we publish. The model is a leverage tool inside the pipeline — it does not write the work alone, and the work does not read like it did.

Why most “AI content” doesn’t work#

Three failure modes show up almost universally in agency LLM-content output:

1. The voice converges to the model’s mean. Without retrieval over the firm’s previous work and tuned prompts, every piece reads like every other piece on the internet — confident, well-structured, and forgettable. The non-commodity test from our SEO service fails by default; the playbook (Part 4) calls this out as the #1 publish-blocker.
2. The facts drift. Models hallucinate at modest but meaningful rates. Without a citation discipline and a fact-check pass, the work cites things that aren’t true. Once published, that’s a permanent credibility liability.
3. The structure flattens. LLMs over-favor symmetric, three-bullet, opening-and-closing-summary structures. Page after page in the same template stops earning the reader’s attention. Real essays have rhythm.

How we operate it#

Voice modeling against the firm’s existing corpus#

Every engagement starts with the firm’s published work — articles, internal documents, transcripts, slide decks, whatever exists. We build a retrieval index over it and tune prompts so first drafts pull from the firm’s actual phrasing, examples, and frame, not from the model’s training data. The savings aren’t in drafting speed — they’re in editing time, which roughly halves once voice modeling is in place.

Research before drafting, not during#

We do source research as a discrete first stage: real citations, real numbers, dated, attributable. Then we structure the argument. Then we draft. Skipping the research stage and asking the model to “include statistics” is how false numbers enter the corpus.

Editorial pass on every piece#

Every published piece is read by a human editor — for fact-check, for argument structure, for the non-commodity test, and for the voice. Pieces that fail the non-commodity test get rebuilt or killed. We don’t ship work just because the queue says we should.

Frequently asked questions#

Are these articles written by a human or by AI?

By a human, with AI as a research and drafting tool. The structure, the argument, the editorial judgment, and the final pass are human work. The model is leverage inside that process. We disclose the model usage in the editorial process where the question matters — for example, when the work is published under a named byline.

Will the content actually rank?

Substantive, non-commodity content that links into the rest of the entity graph ranks reliably. We don’t promise specific positions on specific keywords — see the SEO service page for why that’s a category error — but we operate the work alongside the rest of the SEO function so what we publish compounds.

How many pieces per month?

Whatever’s right for the editorial calendar, never a fixed deliverable count. A typical engagement publishes one or two substantive long-form pieces a month plus 4–6 short essays or briefs, give or take. We don’t optimize the rate; we optimize the work.

Can you write in a voice that isn’t ours?

Yes — for ghost-written executive content, founder-voice essays, and client case studies. We build the voice model from the named author’s existing work; if it doesn’t exist yet, we interview the author until it does. We don’t fabricate a voice from prompts alone.