Score the content

At the stream grain the funnel ratios are fine: an average stream has 3,997 clicks and 484 orders, and 12.1 percent is a stable number. Divide the same stream into 120 minutes and each minute has a few dozen clicks and a few orders. A minute with 3 clicks and 2 orders has a click-to-order rate of 67 percent. A minute with 1 click and 0 orders has a rate of zero. Neither number means anything, and averaging 7,200 of them means nothing with more decimals.

The cure is not a bigger sample; it is a bigger denominator. Every minute has a room, and the room is never small: even a quiet stream holds hundreds of viewers a minute. So the rate that is honest at the minute grain is orders divided by the room, scaled to orders per 1,000 viewer-minutes. Across the whole table it averages 5.80. Its numerator is still small, but its denominator is stable enough that averaging over a segment's minutes lands in the same place whichever way you sum it, which is the property the click-based ratios lack.

• Viewer-minute - one viewer watching for one minute. The sum of the room column over a stream is its viewer-minutes; dividing by views gives average watch time, as session 1 showed.
• Weighted mean - a ratio's average across rows computed as the sum of numerators over the sum of denominators. Rows with a bigger denominator count for more, which is the point.
• Small-denominator bias - what an unweighted mean of per-row ratios does: the rows with the fewest clicks contribute the most extreme values and the same weight as everyone else.
• Lag - a column shifted in time within each stream. orders_next1 at minute 40 is the orders at minute 41. Session 6's regressions need it because an order takes a moment to place.
• Run length - how many consecutive minutes a segment has been running, counting the current one. The fomo_run column session 4 built.

Three things the table says, in the order a leader will hear them. CTA minutes sell. 11.57 orders per thousand viewer-minutes, against a table mean of 5.80: twice the average minute, and about six times a hook minute at 1.98. Price and demo are the next tier, 7.69 and 7.72, close enough that the difference is not worth a script line. The bottom three are not failures. The hook orders at 1.98 because nobody buys in the first minutes of a show; they arrive. It is 12.8 percent of all minutes and the room curve is built there. Filler at 1.40 and dead air at 0.71 are the two rows a host can actually remove.

Now read the right-hand columns with the denominator in mind. CTOR for CTA is 16.73 and for dead air 5.07; the ranking mostly agrees with the bar, which is reassuring, but the click-based rates are computed on a few clicks a minute and they wobble. The bar does not. When a per-segment CTOR and a per-segment orders-per-viewer-minute disagree on a small segment, believe the bar.

To average a ratio across minutes, sum the numerators and sum the denominators, then divide. Never average the per-minute ratio. That is the whole rule, and it is the difference between a number the dashboard would agree with and one it would not.

Watch the two answers separate on a toy pair. Minute A: 3 clicks, 2 orders, rate 67 percent. Minute B: 300 clicks, 30 orders, rate 10 percent. The mean of the two rates is 38.5 percent. The pooled rate is 32 orders over 303 clicks, 10.6 percent. The dashboard, which counts orders and clicks for the whole stream, reports the second. The first is a number about minute A's three clicks, dressed up as a number about the stream.

The rule holds for every ratio in the table. It is why orders per thousand viewer-minutes is computed as sum(orders) over sum(room) within a segment, and why CTR and CTOR in the segment table should be read as pooled clicks over pooled impressions and pooled orders over pooled clicks. When the denominator is the room, the pooled and the averaged versions land within a tenth of each other, because the room is never small; that closeness is the test that the denominator is safe. When the denominator is clicks, they can separate by a factor of three on a small segment, and the pooled one is right.

Add a fourth column to the table: comments per 100 viewers, computed the same weighted way, sum of comments over sum of room. Q&A minutes run at 3.6; every other segment sits between 1.6 and 2.0, with filler and dead air below one. The Q&A minute is the only kind in which the room speaks at nearly twice its normal rate, and it is also the segment with plus 0.96 points of click-to-order within a stream: not a selling minute, but the minute the room decides to stay.

Two things follow. The comment column is a denominator problem too: comments per minute is a count that grows with the room, and only per-viewer is comparable across segments. And the AI-host rows, coming up in Part 2, have no Q&A minutes at all, which this column predicts will show up in the room curve before it shows up in orders.

Urgency minutes as one row in the segment table score 5.83 per thousand viewer-minutes, a hair above the mean. That single number hides two very different minutes. In the first and second minute of a run the click-to-order rate is 15.12 and 15.44 percent, better than every segment except CTA. From the third it is 8.30, then 8.33, 8.43, 9.09: the room has heard "almost gone" three times in three minutes and stopped believing it. Session 6 will price the two halves separately, plus 4.93 points within a stream for short urgency and minus 1.84 for stretched, and the sign flips only because the run length was a column.

So the fomo_run column session 4 built is not a diagnostic; it is a feature. A segment's value depends on how long it has already been running, and the only way a model can learn that is if the table says how long. The same logic applies more gently to CTA (a fourth consecutive CTA minute is a host who has run out of things to say) and to Q&A, but urgency is where the cliff is sharp enough to draw.

A CTA at minute 40 sells at minute 40 and at minute 41. Placing an order takes a tap, a size, a checkout, and for a fair share of viewers that crosses the minute boundary. If the regression in session 6 only looks at same-minute orders it under-credits every segment and mis-credits whatever segment came next. So tonight builds orders_next1, each minute's orders shifted back by one within the stream, and session 6 uses the sum of the current and the next minute as its outcome.

The shift must be within the stream. A plain shift(-1) on the whole table hands the last minute of stream 16 the first minute of stream 17, and the first minute of a stream is a hook with a room of near nothing. The last minute of every stream gets a missing value, which is correct: nothing was observed after it.

Two of the six hosts are AI avatars, Nova and Kai. At the minute grain their rows are unusual in three ways that the stream-level comparison (GMV 8,132 versus 19,850 for humans) cannot see. There are no Q&A minutes. The tagger never fires the name-plus-question rule because the avatar never reads a name. Replies are zero, every minute of every AI stream, against 95.6 per human stream. And the room fades by script cycle: the average room across the eight cycles of the script runs 216, 442, 484, 468, 441, 413, 383, 366 for the AI hosts and 419, 856, 961, 961, 935, 880, 847, 790 for the humans. The AI room peaks lower, peaks earlier, and gives back more of its peak.

What the table cannot say is why. The AI streams also sit mostly outside the evening (26 percent of them in the evening against 95 percent of human streams), and traffic follows the hour. Session 6 holds the hour still and finds the gap survives it. Before that, one fact sits outside the data entirely: TikTok Shop's requirements for LIVEs, in force since 2026-09-17, say "Do not use non-real-time verbal interaction such as AI-generated voices, audio recordings, or radio", so on that platform a fully AI-voiced avatar host is a compliance stop before it is an analytics question. The dataset's AI rows show what the trade would be on a platform that allows it.

The segment table is pooled across 60 streams. That is its strength for a leader and its weakness for a scientist. A segment that appears more often in high-traffic streams borrows those streams' room, and orders per viewer-minute is a rate, so it borrows their conversion too: a segment that good hosts use more will look better than it is, and one that weak hosts lean on will look worse. Urgency is the obvious case: the hosts who stretch it are not a random sample of hosts.

The fix is to compare minutes within a stream, subtracting each stream's own mean before asking what a segment does. Session 6 does exactly that, with stream fixed effects and clustered standard errors, and its numbers (CTA plus 6.50 points of CTOR within a stream, short urgency plus 4.93, stretched minus 1.84) are the ones that survive the confound. Tonight's table is the map; session 6 checks the map against the territory. Read tonight's numbers as "what the minutes did", never yet as "what the segment causes".