Methodology
How we make our content
AI-accelerated, human-reviewed. Every number and claim verified before publishing.
AI drafts here. A person directs the work, checks every figure against its source, and holds responsibility for what goes out under this name. That is the whole claim, and the rest of this page is the machinery that makes it checkable rather than something you have to take our word for.
The review comes first
The order matters. A person decides what is worth writing, what the piece has to prove, and whether the draft in front of them is true. AI makes the drafting fast. It does not decide what ships, and it never settles a number.
Under the EU AI Act's Article 50(4), content that has been through real editorial review by a person with relevant judgement, where a named person holds editorial responsibility, is treated differently from unreviewed output. The Commission's guidance is explicit that a spelling pass does not count. Our byline is Logan Adams because he does that review and carries that responsibility. It is not a claim that no AI was involved, which is why this page exists.
What a piece has to clear
Every piece runs through the same gates, in code, on the way to the site. These are the ones that actually stop a publish today:
- A statistic with no citation and no hedge holds the post. The render-time honesty scan reads the finished prose, not the plan. A bare external figure that cannot be traced is held for review and never written to the live site. If the checker itself crashes, the post is held too: a gate that could not run is not a pass.
- An objective performance claim needs evidence and a date. Every claim of that kind on the site or in a deliverable is registered with its evidence, its source and an ISO date, and the build fails if one ships without them. The same check runs per piece over the rendered prose.
- No outcome promises. That phrasing is banned across every customer-facing surface and checked on each build. Nobody can honestly say what an AI engine will answer next month, so we do not say it.
- Client work carries a recorded human approval of the exact version. Not of an earlier draft. Change the piece and the approval goes stale, so it has to be signed off again before it can publish.
- Provenance travels with the piece. A published item carries the visible line at the top of this page plus a machine-readable marker, and a build that skips it fails a named gate rather than shipping quietly.
Two more run on every piece and report rather than block, which is worth saying plainly because the difference is the difference between a standard and a slogan. A written voice standard grades each piece against eleven rules, including a ceiling of two em dashes per thousand words, and a quality bar scores structure and specificity. Both are graded and reported. Neither one holds a post on its own.
What we refuse to publish
This is the useful half of the list, because it is the half that costs something:
- A number we did not measure and cannot cite. Not an estimate dressed as a finding, not a competitor's figure repeated without its source.
- A chart without a dated source. Every figure we publish carries its method, its interval, the engine set and the date it was measured. A chart that cannot carry those does not get drawn.
- A claim with no source. Including a flattering one about our own work.
- A case study that is not a real client's real result. Illustrative examples are labelled illustrative, everywhere, in the same size type as the number.
- A promise about where an engine will place you. We commit to the method, at any price, and never to an outcome that sits inside somebody else's model.
None of that is a promise about intentions. Each line above is a check that runs on every build, and the build fails rather than warns.
Where the disclosure appears
It goes in a different place on each surface, and the placement is a decision rather than an oversight.
On a piece we publish
A line in the footer of the piece, plus a machine-readable marker and a link here. In the footer rather than beside the byline: a chip next to a name reads as a warning label, a footer line reads as a standard.
In a client report
Inside the methodology section, next to the run counts and the confidence intervals, where it reads as part of the rigor. Not at the top, where it would read as a caveat on work somebody paid for.
In a CSV or JSON export
A provenance field. Machine formats carry provenance natively, and the field says honestly whether the contents were drafted or measured. A data export of measured values does not claim to be AI-written.
In content you publish under your brand
The disclosure travels in your delivery record and in the agreement, not baked into the asset. Your channel, your name, your disclosure decision. We tell you, in writing, so you can make it.
The loop, with the receipt
The test of a measurement business is what it does when its own measurement comes back against it.
On 2026-08-21 we retired the three paid directory-listing tiers we had been selling. The first cycle of our own directory study returned a null result, and the pre-committed outcome, sealed before the answer was known, said that a measurable nothing meant we say so publicly and retire or reprice the tier. We took the retirement. The study, its method and its limitations are published in full, and the decision is recorded in our decision register with the evidence attached.
Read the study, including what it cannot yet tell us →
Research that only ever confirms the price list is marketing. This page is worth reading because that one is on it.
Keep reading