Where tonight sits on the track
Sessions 4 and 5 were the live channel: an event spec per minute, then the stage metrics as code. Tonight moves to the content half of the estate. Session 1 drew two chains that carry most of the value, and this is the second one: content archive, then tagged archive, then creator and content yield. The archive is owned; the yield is what brands ask about every quarter review; the tag in the middle is the link nobody budgets for. The data is six creators and one quarter, constructed, so your numbers will differ. The arithmetic and the traps will not.
★ What you walk out with tonight. A creator table in pandas that separates attributed from settled GMV, three rankings that answer three different questions, a commission check that shows what paying on attributed costs, and a working sketch of the asset-tag join that lets an enabler measure content the platforms never report back.
The chain: content, tag, order
6 min liveAn asset is only measurable if something on the order points back at it. The platforms credit a creator with a total, and the listing with a total, but not "this 30-second clip, placed in this stream, sold these units". So the enabler builds that link itself: a persistent asset id stamped at export, carried into every placement, and read back off the order.
LiveWhat the tag is, and where it has to ride4 min▶
The tag is one id per asset, minted in Serai's archive when the asset is exported, never re-minted when it is re-cut or re-posted. A re-cut is a new asset with a parent id; a re-post is the same asset in a new placement. Get that distinction wrong and one clip shows up as five, each with a fifth of the credit.
Then the id has to ride on something the order will carry back. Each placement offers a different carrier:
| Placement | Where the id rides | What the order gives back | Grain you can reach |
|---|---|---|---|
| Listing | Serai's catalogue maps the listing's content version to the asset id | Listing id and order time | Asset per listing version; mixed if two assets run at once |
| Live stream | The run sheet logs which asset was on screen, per minute | Stream id and minute, from session 4's minute table | Asset per stream minute |
| Creator's link | A sub-id on the affiliate link, set to the asset id | The sub-id on the attributed order | Asset per creator, cleanly |
The link is the cleanest carrier and the stream the most work, which is exactly why session 4 put one clock on the minute table. The platforms do not hand back content-level performance in a form you can pipe into a warehouse; the tag is how an enabler builds its own.
Self-studyWhy the archive is owned and the yield is only partial3 min▶
On the obtainability scale from session 1, the content archive is owned: Serai made every file and knows who, when and for which SKU. Creator results are partial: the affiliate report returns what the platform credited to each creator, in the platform's window, and nothing about which asset did the work. The tag does not make the platform's data any fuller. It makes Serai's side rich enough that a thin order row (an id, a sub-id, a minute) is enough to join on.
This is also why the tag has to exist before anyone asks for yield. An order placed last quarter against an untagged asset cannot be credited retroactively; there is nothing on the order to join. Every quarter without the tag is a quarter of yield you will never be able to report.
Constructed: the brand that asked "which video". A skincare brand on Serai's books (constructed) asked at its quarter review which of the 40 assets made that quarter it should re-shoot. The affiliate report said creator C4 sold well; it could not say with what. Serai's answer was a promise to tag from next quarter. The same review a quarter later had an answer per asset, because the link sub-ids had been set at export. Nothing about the platform changed; the join did.
Attributed is not settled
6 min liveAttributed GMV is the value of orders the platform credited to a creator. Settled GMV is what is left after cancellations and returns, the money that actually reached the brand. The gap is not constant across creators, so the ranking changes, and if commission is paid on the attributed number, so does the bill.
LiveThe six creators, and the two rankings4 min▶
One quarter, commission 8 percent of attributed GMV. All constructed.
| Creator | Followers | Attributed GMV | Settled GMV | Not settled |
|---|---|---|---|---|
| C1 | 2,100,000 | $42,000 | $29,400 | 30% |
| C2 | 180,000 | $18,500 | $16,650 | 10% |
| C3 | 650,000 | $25,000 | $21,250 | 15% |
| C4 | 45,000 | $9,800 | $9,212 | 6% |
| C5 | 1,400,000 | $31,000 | $19,840 | 36% |
| C6 | 90,000 | $12,400 | $11,160 | 10% |
Ranked by attributed: C1, C5, C3, C2, C6, C4. Ranked by settled: C1, C3, C5, C2, C6, C4. Only one pair swaps, and it is the pair that matters: C5, the second-biggest name on the books, is the one whose buyers send the most back. On attributed, C5 is paid 8 percent of $31,000, which is $2,480. On settled it would be $1,587.20.
For the equation GMV itself is built from, link out to the GMV equation in the ecommerce metrics course. The enabler twist is that there are two parties' GMV on one order, the platform's credit and the brand's settlement, and they are not the same number.
Self-studyWhat a high not-settled share usually means, and what it does not3 min▶
A creator whose attributed orders settle at 64 percent, like C5, is worth a conversation, not a verdict. The usual readings: the audience buys on impulse and returns; the creator's claims about the product did not match the product; an attribution window credits the creator with orders they barely touched, which buyers cancel. The data on this table cannot tell those apart. The order-level return reasons, where the platform gives them, sometimes can.
Two things the map does not record, so check them for your own programmes before you call anything overpaid: whether a platform claws back commission on a returned order, and how long after the order a return can still land. Both decide when "settled" is final, and a settled figure read too early is just a smaller attributed figure.
Serai's own network keeps a share of the commissions it brokers, so a commission paid on attributed flatters Serai's line too. Say that out loud when you propose moving to settled; the room will already be thinking it.
A listed marketplace counts orders before they settle, and says so. Sea Limited's annual report on Form 20-F for fiscal 2019 defines Shopee's GMV as including shipping and other charges, and counts an order "regardless of whether the transaction is settled or if the item is returned". That is a legitimate headline definition for a marketplace, written down where a reader can find it. The lesson for an enabler is not that either number is wrong. It is that the platform's number and the brand's number are defined differently on purpose, and a creator ranking must say which one it used. (Sea Limited, Annual Report on Form 20-F for the fiscal year 2019, "Conventions" definitions.)
Yield per what
4 min live"Yield" is a ratio, and the denominator is the question. Divide settled GMV by followers and the list inverts: the smallest creator comes first and the biggest comes last. Neither list is wrong. They answer different questions, and the analyst's job is to say which question each one answers before anyone rebooks anybody.
LiveFour denominators, four questions3 min▶
| Yield | Question it answers | Who wins here (constructed) | The trap |
|---|---|---|---|
| Settled GMV | Who moves the most money, and who reaches most? | C1 | Rewards audience size; says nothing about fit |
| Settled per 1,000 followers | Whose audience buys? Who to rebook, or book more of? | C4, at $204.7 | Small bases are noisy; one quarter is one draw |
| Settled per asset | Which content to remake, and which to retire? | Needs the tag | An asset placed ten times gets ten times the chances |
| Settled per placement | Where to put the next asset: listing, stream or link? | Needs the tag | Placements differ in traffic; compare within a SKU |
The first two need only the affiliate report plus a settlement feed. The last two need the tag from Part 1, which is the whole argument for building it. The build-along computes both on a tiny constructed archive.
Self-studyWhat per-follower yield cannot tell you3 min▶
C4 settled $204.7 per 1,000 followers in one constructed quarter. Whether that holds if Serai books C4 three times as often is not in the data. A small, loyal audience can tire of being sold to, and nothing on this table measures that. Treat the per-follower ranking as the shortlist for a test, not as the answer: book C4 more for one quarter and watch whether settled per 1,000 followers holds.
Which content traits actually sell (the hook, the length, the host, the first minute) is a modelling question, and it has its own home: the live commerce analytics course builds those models. The tag you build tonight is what makes its methods possible on an enabler's own archive.
Constructed: the rebooking meeting that ran two lists. Serai's creator team (constructed) came to a planning meeting with one list, ranked by attributed GMV, and proposed renewing C5 at a higher rate. The analyst brought two more: settled, where C5 slipped to third, and settled per 1,000 followers, where C5 was fifth and C4 first. The meeting kept C1 for reach, moved C5 to a settled-GMV trial, and booked C4 for a test quarter. Three lists, three decisions, and nobody argued about which list was "right", because each one said its denominator in the header.
Rank six creators three ways
★ 22 min · everyone buildsSix rows of constructed creator data, then a tiny constructed archive of four assets and eight orders. Build the creator table, compute settled and the not-settled share, rank three ways, price the commission both ways, then sketch the asset-tag join. Every output below is what the code printed on pandas 2.1, run top to bottom as one script; the data is constructed, so on your own estate your numbers will differ.
Build the six-row creator table, and compute settled from what did not settle. Keep the cancelled and returned value as its own column: settled is a subtraction you can audit, not a number someone typed. The figures are the constructed creator table, one quarter.
Rank three ways from one function. One helper sorts by any column, so the three lists cannot drift apart in how they break ties or which way they sort. Print the per-follower table too: the ranking is the headline, the values are the evidence.
Price the commission both ways. The constructed contract pays 8 percent of attributed GMV. Put the settled alternative next to it, per creator, and print the difference. It is the number that starts the renegotiation, so it has to be one anyone can re-run.
Sketch the tagged archive and join the orders back. A constructed archive of four assets and eight orders. Each order carries the asset id that rode in on its placement, and one does not. The validate="many_to_one" argument makes pandas fail loudly if an asset id appears twice in the archive, instead of quietly doubling every order that joins to it.
Read yield per asset and per placement, and report the untagged share with it. A-103 was credited 180 and settled 60: the per-asset view finds the same problem the creator view found for C5, one level down, where it can be fixed by changing the content. The untagged order stays out of the asset view and in the report as a coverage line.
Report the tag coverage on every yield table. "Settled per asset, 7 of 8 orders tagged" is a number a brand can trust; the same table without the coverage line quietly assumes the untagged orders look like the tagged ones. Session 7 puts both the yield definition and its coverage rule into the metric registry.
Before session 7
30 min- Pull one quarter of your own creator or affiliate report and add a settled column from your settlement feed. Rank both ways and write down every creator who moves by more than one place.
- Find out, for each programme you run, whether commission is paid on attributed or settled GMV and when a return stops counting. Write both on one line per programme, with the date you checked.
- Pick 20 assets from last month and trace each one: does any order carry something that points back to it? Count how many you could join today, and name the carrier for each placement.
- Write the header for each of the four yield tables (settled, per 1,000 followers, per asset, per placement) as one sentence: the question it answers and the decision it feeds.
Three questions
Clear all three and Analyst session 6 gets stamped in your passport.
1 · Why does an enabler need its own asset id on every piece of content?
The archive is owned and the creator results are partial. The tag makes Serai's side rich enough that a thin order row joins to the asset that sold it, and an order placed against an untagged asset can never be credited later.
2 · C5 is second on attributed GMV and third on settled. What explains the move?
$31,000 attributed became $19,840 settled; C3's $25,000 became $21,250. Same rate for everyone, different return behaviour. On attributed, C5 is paid $2,480; on settled it would be $1,587.20.
3 · On settled GMV per 1,000 followers, C4 is first and C1 last. What should the planning meeting do with that list?
Each denominator answers its own question. Settled GMV says who reaches most; per 1,000 followers says whose audience converts. One constructed quarter is one draw, so the per-follower list earns a test, not a verdict.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Analyst session 6 cheat sheet · pin this
Tonight you wrote "settled GMV per asset" three different ways before settling on one. Session 7 stops that from happening across a whole company: a YAML metric registry with one definition per metric, its lineage and a version, and tests that fail the day someone changes what GMV, take rate or yield means without saying so.