A channel nobody knew how to sell on

Pasar Live was new, so the marketplace lent Serai its own creator roster to host the first streams: people with followings, used to entertainment rooms where the job is to keep a crowd laughing for two hours. They did exactly that. The rooms were big, the chat was fast, and every week the GMV table ranked them first, because a room of 9,600 people that converts badly still out-sells a room of 1,600 that converts well.

That is the whole trap in one sentence: a big room with weak selling beats a small room that converts, every time, on GMV. The number is not wrong. It is the right number for the marketplace, whose fees are a share of it. It is the wrong number for anyone who has to decide which host to book next week, what to say to the host who lost the room at minute 14, or whether the brand's own product manager, quietly converting three times better in a room a fifth the size, should be given the bigger slot. GMV cannot answer any of those, because it is a total, and a total has no denominator.

In the constructed eight-stream table on the figure, the four platform-host streams produced 245 orders and the four brand-host streams 192. On GMV the platform hosts win. The question Serai's live ops lead finally asked, "how many people were in the room?", turned 245 into 245 out of 38,830 entries, and 192 into 192 out of 6,890. The second host type had been converting more than four times better for a quarter and nobody had a column that showed it.

Pasar Live's own naming, constructed for this story, follows the viewer through the room. Every stage is a ratio, and every ratio names its denominator:

Four stages because four different people can fix four different things. A stream with high Stop and low Hold is a good card on a host who cannot keep a room; a stream with high Hold and a low click rate is a host people love who never asks for the sale. GMV shows one number for both. The stages show which conversation to have.

The live commerce analytics course teaches the marketplace platform's own version of this idea, named Enter, Stay, Click and Buy, and works it down to the minute: read the room. Pasar Live's names are its own, and the point tonight is not the names. It is that a channel judged on a total was re-judged on ratios, and the ratios could be coached.

A North Star for a channel has one job: to let you compare two streams that were not the same size, on the same night, with the same confidence. A total cannot do that, so a North Star is always a ratio. Then the whole argument moves to the denominator, because the denominator decides what "better" means.

• Orders per room entry (Pasar Live's choice) says: a host is good if the people who came in bought. It rewards Hold and Buy, and it does not punish a host for a weak card, because the card is somebody else's job.
• Orders per live card impression would fold Stop in, and would reward reach. Platform hosts would look better on it, because their faces pull entries. It is a fair number for the traffic team and an unfair one for a brand's product expert.
• Orders per completer would grade only the ask, and would flatter a host who loses everyone but the buyers.

There is a second, quieter denominator decision: how you pool. The eight-stream table's overall conversion is 437 orders over 45,720 entries, which is 0.96 percent. The mean of the eight streams' own conversion rates is 1.71 percent. Both are honest. The first weights every room by its size, so the four big platform rooms drag it down; the second weights every stream equally, so the four small brand rooms lift it. A leader who quotes one without saying which will be corrected in the next meeting by someone quoting the other. Say "ratio of sums" or "mean of ratios", every time, and the room stops arguing about arithmetic and starts arguing about hosts.

The instinct when eight teams want numbers is to build eight views. Serai tried that for a month on Pasar Live and every Monday meeting became a reconciliation: traffic ops said the push went out at 20:14, live ops had the host's clock saying 20:11, tech's event log was in a different timezone, and the brand team's stream had "started" when the host went live while the analysts' had started when the card first appeared in the feed. Three minutes of clock disagreement is fatal when the thing you are trying to see is the minute a room emptied.

The table that replaced them is dull by design: one row per stream, per minute, per SKU, with the five events counted in it (room entries, product card impressions, product clicks, add to cart, orders), on one clock (UTC, and the stream's first minute is minute 0). Every stage in Part 1 is a sum of two columns of that table divided. Every team's question is a filter on it. Nobody owns a dashboard; each team owns the events it fires and the labels it writes, and reads back the answer it needs. The event spec itself, who owns each of the five events and what a missing one costs, is written as a contract in analyst session 4.

The estate argument matters more than the technical one. Eight dashboards are eight sources of truth, and two of them will disagree next quarter with nobody able to say which is right. One table with eight readers disagrees with nobody. It is the same rule session 6 applies to GMV and take rate for the whole company, met here first, on one channel, at the grain of a minute.

The machine half of the table is the five events. The human half is what Serai's live ops team was already doing: sitting in the room with a notebook, writing "minute 14, host went off on a story about her holiday, chat slowed", "minute 22, price reveal, cart spike", "minute 31, sound dropped". Before the table those notes lived in a chat thread and were gone by Thursday.

Attached to the minute, they become labels. Now the analyst can see that Hold collapsed at minute 14 in three of a host's streams and that the note at minute 14 says the same thing each time. The coaching conversation stops being "your Hold is low" and becomes "you lose the room when you leave the product; here are the three minutes it happened". That sentence is the whole reason the table exists. It is also the reason live ops, the least technical team of the eight, is the one the analysts cannot do without.

A scorecard that is read once is a report. A scorecard that is read before every stream is a loop, and the loop is the product. Pasar Live's ran weekly, constructed as follows:

1. Score every stream the morning after, on the four stages and the North Star, from the minute table. No exceptions for small rooms; small rooms are where brand hosts live.
2. Coach against the minute the room was lost. Not against the average. One host, one stream, the two or three minutes where Hold fell, with the live ops note beside each.
3. Go live with one change agreed in that conversation, and only one, so that the next stream can tell you whether it worked.
4. Measure the next stream the same way, on the same table, with the same denominators.
5. Repeat, and keep the scorecards, because two quarters of them are what let you say anything about a lift at all.

The cadence is a leadership decision, not an analytical one. Weekly worked for Pasar Live because hosts streamed two or three times a week. A channel with one stream a month needs a different loop, and a channel with hourly streams needs the scoring automated before the loop can be human at all.

Platform hosts came from the marketplace's roster. Their gift is reach: a face people already follow pulls entries, so Stop is high (8.39 percent pooled, constructed). Their weakness is everything after the door. Hold is low because the room was built for entertainment and drifts from the product; the click rate is low because the ask, "tap the card", is either missing or apologetic. Their coaching line is about structure: stay on the product, ask for the click every four minutes, and let live ops show you the three minutes you lost the room.

Brand hosts are the brand's own staff or merchant creators. Nobody knows their face, so Stop is low (3.98 percent). But the people who do come in stay, because the host knows the product and answers real questions; Hold is 33.24 percent, Love 8.71, and the click rate 12.79. Their coaching line is about the card and the slot: a better cover, a better title, a traffic push timed to the price reveal, because their problem is the door, not the room.

On GMV only, platform hosts win, 245 orders to 192, and Serai would have kept booking them into the best slots. On the North Star, brand hosts convert 4.4 times better (2.79 against 0.63 percent). The right answer was neither "book the brand hosts" nor "coach the platform hosts". It was both, on different stages, with the scorecard deciding which.

Over two quarters of the weekly loop, Pasar Live's conversion, orders per room entry, went from 0.8 percent to 2.4 percent. Serai called it "the 3x". This result is constructed, like everything else in the story; it is the shape of what a scorecard-and-loop can do on a channel that was being run on opinion, not a claim about any real channel.

Even as a constructed number it carries a warning worth hearing as a leader. A rise over two quarters is not, on its own, proof that coaching caused it. The channel matured, the marketplace changed its feed, the brand mix shifted, the weak hosts left. Proving that a lift like this was the coaching, by comparing coached hosts against hosts who were not, is a difference-in-differences question and belongs to the live commerce analytics course: prove it worked. What you need to hold as the leader is simpler: keep the scorecards from before the loop started, because without a before there is nothing to prove.

Three things changed once the scorecard existed, none of them analytical. Slot allocation stopped being a negotiation between whoever shouted loudest and became a question of which host type the brand needed that week: a launch wanted a platform host's door, a repeat-purchase product wanted a brand host's room. Brand reviews gained a page the brand could not get from the marketplace, the per-host funnel for its own streams, which is the first of the "knowing" products session 5 prices. And the marketplace itself started asking Serai for the host-type comparison, because its own reporting was GMV by room and it had the same blind spot Serai had started with.

That last one is the estate lesson. A channel nobody knew how to sell on is a channel where the first team to write down a scorecard becomes the team everyone else asks. The scorecard cost eight teams a month of agreeing on a clock. It paid back as the one thing on the channel that was not opinion.