Method

Every article here is researched and drafted with AI assistance, and no article publishes until Scott Zakrajsek has read it, edited it, and approved it. This page says what happens in between.

The stages

An article starts as a brief: the thesis, the reader, what the reader should be able to do afterward, and what is out of scope. A research stage opens and reads sources against a short list of questions and records every source with its date, its publisher, and the passage that supports each finding. An outline is built from the brief and the research, and the author approves or revises it before a word of prose is drafted.

The draft is then reviewed by separate passes, each starting fresh: three on a short piece, four on a long one. One reads as a sharp CMO with no engineering background and marks every place it had to reread. One checks the writing against the house rules. One argues the opposite case and hunts for the weakest claim and the objection the piece dodges. On a long piece a fourth reads as a senior data leader and looks for what a practitioner would dispute. An editor resolves their findings, and the panel reads again, up to three times.

Sources

A source counts only if the page was opened and read. Search snippets are leads, never citations. Every statistic records who measured it, when, the sample, and the exact wording. A vendor's claim about its own results is labeled as a vendor claim. A source with no findable date is marked undated and cannot support a fact that changes over time.

Research prefers sources from the last twelve months. A fast-moving fact, such as a product capability, a price, or a model version, needs a source from the last six months or a primary source. When two sources disagree, the newer primary source wins and the conflict is recorded. Sources that were read for context and not cited are named at the end of each article under Notes on confidence.

The fact-check

After the review loop, every checkable statement in the draft is listed as a claim with the source it rests on. A separate pass that did not write the draft, and cannot see the research notes or the reviews, reopens every cited page and checks each claim against it: supported, partly supported, or unsupported. A claim that falls short is rewritten to match its source or cut. Nothing reaches the author's final read with a claim in any state other than supported.

The two approvals

The author decides twice. At the outline, where the argument is set and nothing has been drafted. And at the end, on the finished article, where he fills in his own experience, edits freely, and approves. That second approval is recorded with a date in the article's record, and the site refuses to publish an article that lacks it. No article ships on automation alone.

Last verified

Each article shows the date its claims were last checked, in green, under the byline and again under Notes on confidence. Claims that can go stale are rechecked on a schedule, about every ninety days, and an article that changes carries a dated note in the body saying what changed.

The disclosure line

Every article carries the same sentence: "Researched and drafted with AI assistance. Reviewed, edited, and approved by Scott Zakrajsek. Every claim is checked against its source." It is true because of the second approval above.

What the process does not promise

It cannot promise that a source is right. The fact-check confirms that the article matches the source; it cannot confirm that the source matches the world. The source tiers, a second source for any surprising number, the reviewer who argues the opposite case, and the author's own read are what reduce that risk. If you find an error, the author wants to hear it.