Inside the AI Merchandiser Scout
Sourcing has a scheduling problem before it has an intelligence problem. Evaluating a product properly — landed cost, net margin at three price points, delivery days, stock, risk — takes a few minutes. Doing that across enough candidates to find one worth listing takes an evening. Doing it every week, across a portfolio of stores, takes an evening you do not have.
The AI Merchandiser Scout exists to move that work off your calendar. It runs on a schedule, does the screening, and emails you a ranked shortlist with the arithmetic already done.
What it actually does
On a regular schedule — Monday and Thursday by default — the Scout runs per store. The cadence is configurable, but a custom cron only takes effect once it’s registered on the deployment, so treat the default as what’s actually running unless you’ve confirmed otherwise:
- Sources candidates from CJ. Searches against your store’s categories plus any niche keywords you’ve set (“wooden toys, montessori, sensory”), then fills any remaining slots from what’s trending — so a store with categories configured keeps sourcing on-topic even if you’ve never typed a keyword.
- Evaluates each candidate with an AI model, which returns a recommended retail price anchored to the market, a low/high price range, a one-line rationale, a confidence value, and any risk flags it spots.
- Computes the economics — landed cost, gross margin, and net profit per unit after processing fees, ad cost, and a returns allowance, using your assumptions.
- Applies hard gates, discarding anything that fails.
- Scores and ranks the survivors, then emails you the top few.
The email is the interface. Each candidate arrives with its numbers and three buttons: Approve, Tweak, Reject. Approving imports the product into your catalogue — as a draft by default, or activated immediately if you’ve turned that on. The whole loop, from “I have no idea what to list this week” to “there is a new product in my catalogue”, can happen from a phone.
The gates matter more than the score
The Scout’s first job is rejection, and it is deliberately unsentimental. A candidate is dropped, not down-ranked, when it fails any of:
- Gross margin below your floor (default 45%).
- Net margin below your floor (default 15%) — computed after fees, ads, and returns, which is the number that decides whether a sale is worth making.
- Delivery time above your ceiling (default 15 days).
- Stock below your minimum (default 10 units) — when CJ actually reports a number.
- A blocking risk flag — anything reading as branded, licensed, counterfeit, or trademark-related. No margin makes an IP problem worth having.
- Low model confidence. If the evaluation isn’t confident about its own assessment, the candidate doesn’t reach you. A shortlist you can’t trust is worse than a short one.
Unknown delivery time is handled explicitly rather than optimistically: a product with no published delivery estimate isn’t quietly assumed to be fast — it’s scored against your ceiling and flagged as unknown, so you can see why it ranked where it did. Unknown stock is the one exception worth knowing about — when CJ doesn’t report per-variant inventory, the stock gate is skipped rather than assumed to fail it. The shortlist email doesn’t currently show a stock figure either way, so treat every candidate as needing a quick stock check on CJ itself before you commit to it.
How the ranking is weighted
Survivors get a weighted composite score, out of a maximum of 88 points before risk flags subtract up to another 12:
| Signal | Weight |
|---|---|
| Net margin percentage | 24% |
| Net contribution in dollars | 19% |
| Demand | 20% |
| Model confidence | 17% |
| Delivery speed | 8% |
| Risk flags | −12% (penalty) |
The reason margin appears twice is the trap described in our product evaluation guide: percentage margin flatters cheap products, and dollar contribution flatters expensive ones. Weighting both stops the shortlist filling up with $6 trinkets at 70% margin that each net you three dollars.
Risk subtracts rather than blocking here. Blocking flags are already gone; the remaining ones are softer concerns that should push a candidate down the list without removing your ability to look at it.
The assumptions are yours, and they should be
The Scout’s output is only as good as the cost model you give it, so all of it is editable per store: shipping cost model (a flat assumption or weight-scaled), processing fees, target ad cost as either a percentage of retail or a flat cost-per-acquisition, per-category ad ratios, returns allowance, margin floors, delivery ceilings, and stock minimums. The settings screen shows net profit on an example product as you change them, so you can watch a gate tighten before it silently rejects everything.
Two optional switches worth knowing about:
- Market anchoring (off by default) lets the evaluation consult live web comparables when pricing a candidate, rather than relying on CJ’s own signals alone.
- A display currency so shortlist prices arrive in the currency you think in, with tidy price endings applied. This is a plain rate conversion for readability, not the buffered conversion your storefront checkout uses — treat the shortlist number as a close estimate, and let the storefront price be the one you actually sell at.
You can also ignore the schedule entirely and hit Generate recommendations now when you want a shortlist today. The Scout needs two things to work: a connected CJ account and a configured AI provider. The settings screen shows both as readiness checks, along with the digest recipient, so it fails visibly rather than going quiet.
Where a human still decides
Nothing here picks your niche, and nothing here tells you the product will sell. The Scout screens on economics and delivery, which are the parts that can be computed. Whether a plush dinosaur belongs in your store, whether the ad creative writes itself, whether the product photographs well — those are still yours, and they’re the ones worth spending your evening on.
That’s the trade the shortlist is making. It removes the mechanical part of sourcing so the judgement part gets your attention instead.
The Scout is live in the product today and included in the public beta, which is free while we build. If you want to see the rest of what’s shipped, the roadmap lays it out honestly — including what hasn’t shipped yet.