AliExpress AI Verdict Explained 2026 — How the Dropship Score Works

AI Verdict Explained — How the Score Actually Works

Quick answer: The AI Verdict is a 0–100 composite of four weighted dimensions — product fundamentals (30%), profit estimate (30%), market saturation (20%) and risk (20%) — that resolves to one of four labels: Strong Buy at 80 and above, Buy Signal 60–79, Hold / Watch 40–59, Pass below 40. Beside it the panel shows a confidence percentage, which tells you how much real data the verdict rests on. Every dimension's own score is visible in the panel.

Open any AliExpress product page in AliShopping Tools and the Verdict tab gives you three things: a 0–100 score, a label, and a confidence percentage. This page explains all three, including the exact arithmetic, because a score whose formula you cannot check is not something you should be spending ad budget against.

The four dimensions and their weights

The composite is fixed and unweighted by user preference — it is the same for everyone:

overall = product × 0.30 + profit × 0.30 + saturation × 0.20 + risk × 0.20, rounded.

Two things are worth reading off that table. First, margin and competition together are half the verdict — which is exactly what a product-quality score on its own leaves out, and why a high Winning Score does not automatically produce a Strong Buy. Second, the profit dimension saturates: because the margin percentage is doubled and capped, anything at or above a 50% margin scores the maximum. Above that point extra margin no longer moves the verdict.

Where product fundamentals come from

The 30% product dimension is the Winning Score, computed server-side as its own 0–100 composite of seven weighted factors:

Its own bands are 75 and above for High Potential, 50–74 Medium Potential, below 50 Low Potential. It is identical for every user; there is no per-user weighting anywhere in the extension.

Where market saturation comes from

Saturation is derived from how many competing listings exist for the product, on a logarithmic scale — so the difference between 10 and 100 competing sellers moves the number far more than the difference between 1,000 and 1,100. It then enters the verdict inverted: a crowded market drags the score down.

Where risk comes from

The risk dimension is the supplier-and-listing risk read surfaced in the panel's own risk section, inverted the same way. We are not going to publish a list of exact red-flag thresholds here, because the honest position is that the risk read is a section you should open and look at rather than a number to memorise.

The band cut-offs, published

There are four states, not two. That distinction matters in practice: Hold / Watch is not a soft no. It is the band that says something real is being flagged and the answer depends on a judgement the model cannot make for you — so open the individual tabs rather than treating the label as a decision.

Publishing these cut-offs is a deliberate choice. An unpublished threshold does not make a model more honest about its precision; it just makes the output unauditable.

What the confidence percentage means

A confidence percentage is rendered next to the verdict. It is not a second opinion on the product — it is a statement about the evidence.

When the backend scores the listing, confidence is a mechanical function of data coverage: how many of six genuinely observed signals were available (winning score computed, risk computed, real 30-day order volume, real review count, real rating, and at least two real daily order snapshots to form a trend series), discounted when no real trend series exists. Six of six with a real series reads 100%; a listing missing half of them cannot report a confident verdict, by construction.

There is a floor. If fewer than two of the six signals are present, the backend refuses to emit a directional verdict at all and returns an explicit insufficient-data state instead of a degenerate "Pass". A thin listing gets told it is thin rather than quietly scored.

When the extension has to fall back to computing the verdict client-side, confidence is derived from the weakest of the underlying analyses rather than their average — so one badly-supported input caps the whole number.

Read it as a discount on the label, not as a separate score. A Buy Signal at high confidence and a Buy Signal at low confidence are different situations.

What happens when data is missing

Any dimension the extension cannot compute defaults to 50 — the neutral midpoint — rather than being dropped or guessed.

This is the single most important caveat on the page, so it gets its own paragraph: a verdict built on two real dimensions and two neutral placeholders produces the same headline number as one built on four real dimensions. The confidence percentage is what distinguishes them. If a listing is thin, read the four dimension scores in the panel, not the label.

Why the breakdown is visible

Every dimension's own score sits next to the verdict in the panel, so you can see which one carried the result. If profit is strong but saturation is dragging, that is visible; if the whole thing is resting on neutral defaults, the confidence number says so.

This exists because operators should not be asked to trust a black box. You can see what drove the label and decide whether you agree with the emphasis — if demand is strong but you know the category is seasonal, discount it yourself.

When the verdict is wrong

The verdict is directional, not predictive. Known weak spots:

It does not know your niche. The score is computed from public listing data. It has no view of your audience, your creative, or your ad costs. A Strong Buy can fail in a market you are not positioned for; a Pass can be right for an audience you happen to own.

Seasonality. A Christmas product scored in February shows a weak demand read. The same listing in October reads differently. The score reflects the moment, not the calendar.

Thin listings. New listings have too little order history for a meaningful velocity read. The verdict may be correct but it is resting on very little — which is what confidence is there to tell you.

Gamed signals. Seeded order counts and review farms exist. Obvious patterns get flagged in the Reviews tab, but a sophisticated one can move the demand and social-proof inputs.

Listing versus product. The model scores the listing. It has never held the item. A listing that presents better than the product will score better than the product deserves.

Fast-moving external shifts. Tariff and regulatory changes can move faster than the risk read updates.

In all of these the score is still computed. It just rests on less.

How to use it pragmatically

• Strong Buy / Buy Signal — a reason to look harder, not a reason to spend. Check the confidence number first, then the Profit tab for margin viability and the Reviews tab for fake-review flags.
• Hold / Watch — the label that most rewards opening the tabs. Something specific is being flagged; find out which dimension and decide whether it applies to you.
• Pass — move on unless you have knowledge the model provably lacks. A Pass usually means more than one dimension failed badly.

Compared to paid tools

Sell The Trend, Minea, Ecomhunt and Pexgle all ship scoring systems. The difference is not that ours is more accurate — we have no basis to claim that. The difference is that the weights, the cut-offs and the missing-data behaviour above are published, and theirs are not. Ours also runs on any AliExpress product page in your browser, free, without an account.

No scoring methodology is objectively correct. Pick on the basis of what you can audit.

Verify it yourself

Run the Verdict on 10 products you already have an opinion about — ones you are confident will sell, and ones you know are dead. Check whether the label agrees, and check the confidence number when it disagrees. After 10, you will know whether the scoring matches your judgement, which is more useful than anything on this page.

Install AliShopping Tools free — no account, no paywall.

---

Corrections, 2026-07-23: an earlier version of this page described a two-label verdict (buy versus skip) built from three sub-scores — trend, demand and margin and stated that no numeric cut-off was published. The shipping extension has always used four labels, four weighted dimensions and fixed cut-offs at 80 / 60 / 40; those are now published above. The same version also implied there was no confidence figure. There is, and it is described above.