Recognition judge
A closed question · you or not you
- The question it asks
- Is this answer about you, a namesake, or a guess?
- What it reads
- The engine’s answer alone.
- What it refuses
- Wrong-store answers never score as yours.
Mechanism
How Burnish works.
An always-on instrument with four jobs. This page is the whole mechanism.
Two judges read every answer.
The twelve-question battery fires in a continuous round-robin; these grade what comes back.
Accuracy judge
An open grading · claim against page
- The question it asks
- Is what the engine claimed actually true?
- What it reads
- Your live policy pages, fetched fresh, beside the claim.
- What it refuses
- A claim is graded against your own pages, not taken at its word.
The brand brain.
Before Burnish says a word about your brand, it builds the manifesto.
[01]
A fact spine, with sources attached
Extracted from your store, invention forbidden. Nothing enters that cannot be pointed at.
[02]
A narrative pass you approve
You approve and can regenerate it. Every engine-facing word traces back — discovery queries, battery tokens, articles, rewrites, schema.
Four jobs, on a loop.
Each one hands its output to the next, and the last one starts it again.
-
Measure
What five assistants say about your store, sampled continuously rather than once.
-
Diagnose
Every gap comes with a why — which field, which product, which question.
-
Fix, with your approval
Twenty-nine of the forty-one fix types wait for you. The rest sit behind floors.
-
Prove
What the work earned, measured on the same instrument and labelled with its limits.
Representative demo · Threshold Footwear
Burnish design sandbox
Living Loop · Concept A
01
Catalog mapped47 prompts tracked02
Sampling now4 engines · 1 min agoView log →03
ResolvedLast autopilot run · 9:43 PM · 4 fixes appliedQueue →04
Tracking lift+$1,240 · 9 fixesSee lift →Earned autonomy.
Never assumed. Each level is unlocked by a record, not by a plan.
L1 · Review everything
Burnish proposes. Nothing is written until you say so.
L2 · 20 approvals
Half the queue starts auto-publishing. The other half still waits.
L3 · 50 approvals
The widest lane, and it still needs explicit consent.
How the proof works.
Four signals, and one sentence about what the result is not.
| Signal | What it is |
|---|---|
| First-party pixel | Consent-gated and zero-PII. Classifies AI-referred sessions and ties them to orders inside a 60-day window. |
| Shopify order journeys | Corroborate the pixel. First and last visit source only — no customer data. |
| GA4 | Optional cross-validation. Never a requirement. |
| Per-action lift | Ships with bootstrap 95% confidence intervals. |
| What it is not | Directional, not a controlled trial. The product labels it that way wherever it appears. |
Four guardrails, always on.
They do not switch off at the higher autonomy levels.
Start with the reading.
The free audit runs the instrument against your store and shows you what came back, engine by engine.