Methodology

A scan is an outside-in hypothesis built from public company information only. It is not an audit, not a certification, and not a view inside your business. Every claim is either quoted verbatim from a page we read, or labelled an inference — and the whole result is built to be challenged: every opportunity carries an “Is this right?” control, and your corrections recalculate the analysis.

The scanner reads 3–10 of your most informative public pages: the home page first, then sections it recognises as about, services, careers, news, clients, pricing and contact. It respects robots rules and applies size and time limits per page.

Pages about individual people — team, staff, leadership profiles — are deliberately skipped: the scan works at company level and does not collect personal data. Nothing behind a login is ever read, and the text of your website is always treated as data, never as instructions.

Every result lists the exact pages read, with links. If some pages could not be read, the result says so rather than pretending completeness.

Each opportunity carries three separate measures, kept separate on purpose:

A scan returns at most five opportunities, strongest first. Value is always framed as capacity, speed, quality, revenue or risk — never as replacing staff.

Confidence is bounded by hard, deterministic rules that run after the analysis and outside the model:

• An opportunity resting on public pages alone can never exceed 85% — outside-in evidence has a ceiling by design.
• An opportunity supported by a single verified quote can never exceed 70%.
• Only evidence you supply yourself — answers in “Refine my scan”, marked You told us — can lift the public-only ceiling, because you know your business and we only know your website.

Confidence is intended to be calibrated: a 60% should turn out right roughly six times in ten. We have not yet accumulated enough feedback to publish calibration figures; until we do, treat confidence as an evidence-weighted estimate under the caps above. The “Is this right?” answers are what will let us test it — and we will publish the results, including where we were wrong.

Every fact and every opportunity must cite at least one quote from your pages. After the analysis, a deterministic check — plain string matching, no model involved — verifies that each quote appears word for word in the source it cites, ignoring only letter case, spacing and typographic quotes and dashes. Quotes shorter than eight characters are rejected outright.

A claim whose quotes all fail this check is not softened or reworded — it is dropped from the result. The counter at the top of every report (“N of M quoted pieces of evidence matched your public pages word for word”) reports this honestly, including the misses.

We would rather you know where this breaks than discover it yourself:

• We overweight what websites talk about. Websites talk about clients and services, not admin — so back-office work is systematically under-represented in outside-in results.
• Brochure sites defeat the method. A site that says little about day-to-day work produces an honest “not enough evidence” rather than a padded result.
• Public pages lag reality. Your website may describe the business you were a year ago.
• Inference is fallible. The reasoning between verified quotes is model judgment. That is why facts and inferences are labelled separately, and why the report asks you to challenge it.

Scan results live at an unlisted link, private to whoever holds it. “Delete my scan data” on every report permanently removes the scan, its results and your answers. Details on what we store and for how long: how we use data.

The methodology is versioned like software: every report records the methodology version that produced it, and internally we log the prompt versions, model and cost of every run. When the methodology changes, historical results are preserved as issued — never silently recomputed.