Compliance keeps you listed. Competitive content wins the click.
Most content tools answer one question: does this listing break the retailer's rules? That's necessary and it isn't enough — a listing can be perfectly compliant and still lose every click to the three products sitting next to it. This prototype answers the second question too, and then does the part tools usually leave to the user: it drafts the fix, cites the rule behind it, and shows its evidence from the shelf.
Pick a SKU
Two client listings, each against its own competitive set. The dataset ships with a
competitor_group column, so selection is a lookup here — where it's absent, the set is built from
the category path plus similarity on title and attributes, excluding the client's own brand.
The shelf, measured
Where this listing sits against the products it competes with, before any opinion is offered.
| Brand | Title chars | Bullets | Avg bullet | Description | Images | Rule violations |
|---|
What's wrong, and how we know
Two kinds of finding, kept visibly apart. Compliance comes from the supplied ruleset and is judged listing by listing. Competitive comes from the shelf and has no rule behind it — a listing can pass every rule and still be the thinnest page in its category.
The three edits that matter most
Ranked compliance risk first, then discoverability, then conversion — and spread across fields, because three fixes to one field is one fix. Each proposed edit is re-checked against the ruleset after it's written, not before.
What we're deliberately not touching
The section most tools skip. A recommendation the user can't act on is noise, and a claim the data can't support is a liability.
Why it's built this way
The design decisions an evaluator would ask about, stated up front.
The model doesn't count
Every rule is tagged T1 (checkable in code), T2 (needs judgment) or T3 (a real gap that text can't fix). A
model asked to count characters is slower and less reliable than length. Judgment is where a
model earns its place.
Guidelines are a file, not code
The ruleset loads from JSON with its own source and version. A different retailer, or an updated policy, is a new file — not a rewrite. That's also why every finding can cite an ID.
Generated copy is re-checked
Whatever the model writes goes back through the same deterministic checks before a human sees it. A rewrite that fixes one rule and breaks another is caught here, not in Seller Central.
Missing facts get flagged, not invented
Where an improvement needs a fact the data doesn't hold — a flavour name, a size range, a certification — the edit stops and asks. Hallucinated product claims are the one failure mode that costs a brand more than bad copy.
The demo can't die
Deterministic analysis runs entirely in the browser. The drafting step calls a model through a serverless function, and falls back to pre-computed recommendations if the call fails — which are then validated the same way.
Where this goes next
Catalogue-wide, prioritised by revenue at risk rather than SKU by SKU; continuous monitoring that fires when a competitor updates their page; and optimising for conversational shopping queries, not just keyword search.