# How KnitKnot measures movement without overstating it

> Every cycle begins with a preserved baseline, keys the record to the published revision, and keeps thin data and outside changes visible.

- Canonical: https://knitknot.ai/methodology/
- Measurement lab: https://knitknot.ai/experiments/

## Journey, cycle, version

- Journey: the buyer questions and category scope the customer chose to track.
- Cycle: the evidence packet, proposed changes, approvals, publishes, declines, and measurement window in one record.
- Version: the immutable published artifact, base content hash, overlay identity, and evidence snapshot active for that window.

## Freeze the before-state first

A completed benchmark must be frozen before the first managed revision is published for a workspace or newly managed page class. Otherwise the original representation is gone and the comparison is not valid.

Available traffic history is labeled with its source and marked as before KnitKnot. If the deployment mode has no pre-install traffic baseline, the product says so.

## Changed, then measured

The default statement is changed -> measured movement. The cycle report shows visibility, head-to-head, and misrepresentation movement beside the revision and its receipts.

A later movement can have more than one explanation: other published pages, competitor changes, model updates, retrieval changes, or normal answer variation. KnitKnot keeps those caveats beside the result.

## Insufficient data is a result

A journey or cycle below its reporting floor is labeled insufficient data. KnitKnot does not print a delta from a sample that is too thin to interpret. The rule applies before a favorable or unfavorable story can be written.

## Causal language requires a qualifying experiment

A holdout comparison must include both cohorts above the floor and support the specific per-page contrast. Headline journey and workspace rates remain association-grade. The permission is machine-gated for the result it covers.

## Keep evidence lanes separate

Verified crawler fetches, delivery outcomes, agent tool calls, AI-referred visits, and benchmark results answer different questions. KnitKnot displays them separately, does not add them into one total, and does not treat them as substitutes for journey-level measurement.

## Publishing enforcement

Every managed claim must trace to the customer's existing page or an approved fact with a receipt. Human content is not withheld, keyword stuffing and fake freshness are blocked, and every revision is reversible.

Revoked or refuted claims take a priority strip path. Automatic publishing only becomes available after the relevant evaluation gates pass, and rewrites remain manual.

## Open notebook

The measurement lab records scorer repeatability, placebo behavior, selection bias, and variance-reduction work. These historical write-ups explain why the current reporting rules exist: https://knitknot.ai/experiments/