“One hundred thousand memberships in thirty minutes” is the number Tmall's token section posted in the Singles' Day run-up, the moment a digital credit began behaving like a fast-moving consumer good; as of November 9, 2026, both Tmall and JD.com keep AI tokens on a dedicated shelf, with monthly token sales reported up 450%. The shelf is older than the reporting around it, and the reporting is already behind the mechanics: a package of model credits now has a category node, a comparable price and a review stream, and each of the three creates work nobody has assigned yet.
What actually went on sale
A token is a unit of model usage, and until this season it was sold the way cloud capacity is sold: through a console, by the million, to whoever was sizing a workload. Moving it onto the same shelf as shampoo changes the buyer. The person adding ten thousand tokens to a cart is not sizing anything; they are buying a novelty, a gift, or a small experiment for the assistant on their phone.
The marketplace mechanics travelled with the product. A token package now has a list price, a promotional price, a sales count and a review section. The review section is where the mismatch appears, because on a token listing the buyer and the user are often not the same person, and are almost never on the same device.
Why a token behaves like a SKU
Catalogue mechanics carry over
Everything a product page does for a physical good, a token package now has to do. It needs a title that survives search, a category placement, a price that can be compared and a description that answers what the buyer is actually getting. The category placement is the hard part, since a token is not a subscription, not a gift card and not software, and whichever node it lands in decides which competitors it is benchmarked against.
That node also decides the price comparison a shopper sees. Placed next to cloud subscriptions, the per-unit price of a token looks cheap. Placed next to a gift box at the same price, the identical package looks expensive. Price monitoring has to know which shelf it is reading before it calls anything a price gap.
Discounts compound instead of clearing
Promotional pricing on tokens does not behave like a discount on a physical good. A cut on shampoo moves units and clears stock. A cut on tokens moves units and creates a liability, because every discounted token sold is usage the platform must serve later at full cost. The promotional calendar therefore has to be read against delivery capacity rather than against inventory.
Consumer demand at this stage is being created, not measured. That is the unusual part of the category growth story: the shelf is generating first-time buyers faster than it is generating repeat buyers, and the two need separate tracking before anyone can call the category mature.
Reading reviews when buyer and user differ
The review stream on a token listing mixes populations that are complaining about different things, and the average hides both of them. Separating the groups is what makes review analysis useful on a digital good, and the split is visible in the language of the reviews themselves: one group describes a transaction, the other describes an output. Read the two together and the rating looks mediocre; read them apart and the store only owns half of the complaints.
| Complaint type | Who reports it | What it signals | Where to look |
|---|---|---|---|
| Code or activation failure | First-time buyers | A fulfilment defect at the store | Order-level refund notes |
| Output quality | Heavy users | A model property, not a seller fault | Long reviews with usage detail |
| Quota confusion | Both groups | Listing clarity | Questions repeated in review threads |
| Price after the promotion | Returning buyers | A promotional hangover | Repeat purchase rate by cohort |
Pooled, the average rating says tokens are roughly fine. Split, it says the listing is clear, fulfilment occasionally breaks, and the output complaints belong to the model provider rather than the store. Those are three different owners with three different fixes, and a single star average assigns the work to nobody.
This is the same discipline review analysis applies to any fast-moving category, with one addition. For a digital good the fulfilment event and the consumption event are separated in time, so a bad first experience and a bad hundredth experience land in the same thread, months apart, under the same product page.
What to log before the next promotion
The checklist assumes the token shelf stays up past the festival, which the platform disclosures so far suggest it will, and it assumes the buyer arriving in December is not the buyer who arrived in November. Both assumptions are cheap to test: the shelf either persists into the next promotional cycle or it does not, and the second-month cohort either returns or it was a one-off. The four fields below are ordered so that the cheapest test runs first.
- Record the category node each package sits in, because the node sets the comparison set and the comparison set sets the price expectation.
- Capture list price and promotional price separately, then compute the per-unit gap against the same model sold directly by the provider.
- Track refunds at order level and tag the reason, since activation failures and partial-use disputes are fixed by different teams.
- Split repeat purchase from first purchase by cohort, and treat the two as different numbers rather than one growth figure.
| Field | Example value | Reading |
|---|---|---|
| Listing category | AI services, model credits | Sets which products the package is compared with |
| Price per million tokens | 39 yuan promotional, 59 yuan list | Gap against the same model sold direct |
| Refund rate | 2.4% of orders | Rises when packages cannot be partly used |
| Repeat purchase share | 31% of buyers in month two | The cleanest available demand signal |
Run the four monthly and the category becomes legible in a way the festival headlines are not. A sales count of 100,000 memberships in half an hour is an event. A repeat purchase share of 31% in month two is a verdict, and only the second one tells a merchant whether to order more shelf space or reprice the package.
FAQ
Is a token package really a product listing?
On a marketplace it behaves like one. It has a category node, a comparable price, a sales count and a review stream, and each of those creates the same monitoring questions that a physical SKU creates.
Why split reviews by buyer type?
Because a first-time buyer reports activation and delivery, while a heavy user reports model output. Averaging the two hides both signals, and the complaints route to different owners inside the same store.
What does a promotional discount on tokens cost?
More than the margin on the unit. Each discounted token sold is future usage the platform must serve at full cost, so a promotion has to be checked against delivery capacity rather than against stock levels.
Does the 450% month-on-month figure mean the category is established?
Not yet. It measures the first purchase wave. Whether tokens become a repeat category depends on the second and third purchase, which the current platform disclosures do not cover.
