Where we are in the leader track
Session 1 split every stream into what the room was given and what it did with it. Session 2 took the giveaway apart: more people, less sold per person. Session 3 was the say: a 12-minute script with a measured value per segment, and the urgency cue that pays for two minutes and costs from the third. Tonight is the show: the same minutes, seen instead of heard. The Lumen Glow data has a column for each thing on screen, minute by minute, across 60 streams, which is what lets you put a number on the camera without touching the script.
★ What you walk out with tonight. A shot list for the 12-minute cycle with a measured value next to each element: when the serum is in hand, when the camera is close, when the price card is up. And a one-page human-versus-AI decision that starts where it has to start, with the platform's LIVE quality rule, then prices the per-viewer gap on the hours where both kinds of host actually worked. The headline number to carry out: holding the product up on every CTA, with the words unchanged, is worth 9.7 percent more GMV a night on the replay bench.
LiveThe words move CTOR, the picture moves CTR3 min▶
The funnel from session 1 has two host-owned steps in a row: click, then buy. It turns out they answer to different people in the room. When the minute table is demeaned by stream, so each stream is only compared with itself, the visual columns move the click rate and barely touch the order rate. The script columns do the reverse. A CTA minute adds 6.50 points of CTOR and 3.28 points of CTR; the serum in hand adds 1.38 points of CTR and 0.04 of CTOR, a number that is indistinguishable from zero.
That split is the most useful thing in tonight's session, because it means you can coach two things separately. If a host's click rate is low, look at the camera. If people click and do not buy, look at the words. A review that says "the stream was flat" has not yet said which of the two it was, and the fixes do not overlap.
The product card gets shown every minute someone stays. A click is a viewer deciding the thing on screen is worth a closer look. An order is a viewer deciding the price and the proof are enough. The picture is what they are looking at when they decide the first; the host's voice is what they are listening to when they decide the second.
Self-studyThe vocabulary tonight adds4 min▶
- Product in hand - the minute has the serum visibly held, not sitting on the table. The builder track's detector reads a hand landmark and a bottle in the same frame.
- Close-up ratio - how much of the minute the product fills a large share of the frame. A number between 0 and 1; a dropper you can see the drop on is a close-up.
- Price overlay - a price card on screen. Detected by reading the text in the frame (OCR), so it is the price the viewer can read, not the price the host said.
- Scene cuts - how often the picture changes: camera moves, graphics in and out, the host walking off. Too few is a still; too many is noise.
- Host on camera - a human face is in the frame. The minute the host steps away to fetch stock is a minute without one.
- Avatar - a generated presenter, face and voice, driven by a script. Nova and Kai in the Lumen Glow data. Called a digital human in the platform's policy pages.
- Real-time interaction - the host hearing the room and replying now: reading a name, answering a question, adjusting the pitch. The platform's LIVE quality rule is written around this phrase.
What the camera adds
9 min liveThree visual columns carry almost all of the camera's value: the product in hand, the close-up ratio, and the price card on screen. Each adds about a point of click rate within a stream, a third to a half of what a CTA minute adds. None of them adds anything to the order rate once the words are held constant. That is not a disappointment. It is a division of labour.
LiveWhy these move clicks and not orders3 min▶
Run the same within-stream comparison on the order rate and the camera columns go quiet. Product in hand: +0.04 points of CTOR, p .81. Close-up: +0.46, p .65. Price overlay: -0.13, p .58. A p-value that high means the data cannot tell the effect from nothing; the sign is noise. The camera earned the tap and then handed the viewer to the words.
This is worth holding onto when a vendor shows you a slide saying "close-ups lift conversion 30 percent". Across a stream, yes, a stream with more close-ups may well convert better, because a host who shoots close-ups is also a host who does other things well. Within a stream, minute against minute, the close-up moves the click and the pitch moves the order. The cross-stream number is a description of good hosts. The within-stream number is an instruction you can give one.
| Element | CTR effect (points) | CTOR effect (points) | What to coach |
|---|---|---|---|
| Product in hand | +1.38 | +0.04, p .81 | Pick it up before every ask, every time |
| Close-up ratio | +1.01 | +0.46, p .65 | Dropper and texture in frame during demo minutes |
| Price card on screen | +0.97 | -0.13, p .58 | Card up when the price is said, down when it is not |
| CTA minute (script) | +3.28 | +6.50 | Session 3: where the ask sits in the cycle |
LiveWhat one camera rule is worth, in dollars3 min▶
A point of click rate is hard to feel. Dollars are easier. The replay bench in session 5 takes tonight's stream, Jules on a Friday at 20:00, and replays it 40 times with different seeds under a policy, then averages. The host's own plan averages 18,198 dollars a night. Add one rule, product in hand on every CTA, with the words unchanged, and the same 40 nights average 19,960 dollars: +9.7 percent.
That is a rule with no cost. No new script, no extra minutes, no tool to buy. It is a habit: the serum goes into the hand before the ask leaves the mouth. The bench prices it because the click rate on CTA minutes rises and the CTA minute's own order rate then acts on more clicks. Camera plus script, in that order.
A Lumen Glow host, four streams into her run, had the script down: CTA on the right minute, urgency held to two. Her click rate on CTA minutes still trailed the other humans. The recording showed why. During every ask she set the serum down to gesture with both hands. The words were right and the product was out of frame. The fix was one sentence, pick it up before you ask, and it needed no rehearsal.
Self-studyTikTok Shop's own show rules4 min▶
The platform publishes a shape for a good LIVE in its Academy pages, and the data agrees with it more often than not. Three rules worth knowing before you write a shot list.
- Do not switch the pinned product faster than every 10 to 15 minutes in a two-hour LIVE. The pin is the card viewers tap; a card that changes under them resets the click. The 12-minute cycle from session 3 fits inside this by design.
- Ten or more products in the bag. The Academy's line is that LIVEs with 10+ items see 2x higher shopping bag clicks. Lumen Glow sells one hero serum, so this is where the two-pack bundle and the rest of the range earn their place in the bag even when the pitch is about one bottle.
- No stills or slideshows covering more than 50 percent of the screen. This sits in the Requirements for High-Quality Videos and LIVEs, effective 2026-09-17 in the US. A full-screen price card breaks it; a card in a corner does not. It also matters for Part 2.
The tools the platform offers for the show, LIVE Flash Sale with a countdown, LIVE Coupon, Creator Exclusive Price, put a graphic on screen with a real end time attached. A countdown that the platform draws is one that the platform can also check. Session 6 comes back to why that matters.
Human host or avatar
8 min liveTwo of Lumen Glow's six hosts are avatars, Nova and Kai, running the same script. The per-stream gap is large: human streams average 19,850 dollars, avatar streams 8,132. Most of that gap is not the avatar. Some of it is. And on TikTok Shop, a fully AI-voiced avatar host is not permitted on a promotional LIVE, so the numbers describe a trade you would be making somewhere else.
LiveStart with the rule3 min▶
Before any number, the platform. TikTok Shop's Requirements for High-Quality Videos and LIVEs, in force in the US from 2026-09-17, say three things that decide the avatar question on their own:
- "Interact with viewers through real-time verbal or sign language communication."
- "Do not use non-real-time verbal interaction such as AI-generated voices, audio recordings, or radio."
- "Do not use animated figures or content that covers more than 50% of the screen."
Put together: a fully AI-voiced avatar host is not permitted on a TikTok Shop promotional LIVE. Not discouraged, not penalised in reach; not permitted. The consequences for a shop run from warnings to traffic limits to account removal, and a Risk label on the LIVE Diagnosis page sits above every GMV number you have. So the decision page you build tonight starts with one line, "on TikTok Shop: no", and everything after it is about the platforms and formats where the rule does not apply: an avatar on a recorded video with disclosure, a human-voiced digital presenter, or a LIVE on a marketplace with different rules.
The Lumen Glow avatar streams exist in the data so you can see what the trade would be where it is allowed. Read them that way.
LiveThe per-viewer gap is smaller than the GMV gap3 min▶
Human streams out-earn avatar streams by more than two to one, 19,850 against 8,132 dollars. But 95 percent of human streams ran in the evening and only 26 percent of avatar streams did. The avatars got the daytime hours because they were cheap enough to run there. So the raw gap mixes two things: who was hosting, and when.
Take only the hours where both kinds of host actually worked, 17:00 and 22:00, and compare again. GPM: humans 1,910, avatars 1,258. Watch time: 9.91 minutes against 7.29. The avatar still loses on every per-viewer number, but the gap is now a per-viewer gap and not a per-viewer gap wearing a scheduling gap on top. That is the number that belongs on the decision page, because it is the one that survives someone in finance asking "but weren't the avatars on at lunchtime?"
| Comparison | GPM human | GPM avatar | Watch human | Watch avatar | What it mixes in |
|---|---|---|---|---|---|
| All streams | 1,981 | 1,252 | 9.80 | 7.28 | The hour: 95% vs 26% evening |
| Same hours only, 17:00 and 22:00 | 1,910 | 1,258 | 9.91 | 7.29 | Host type alone, as near as the data allows |
Session 2 made the same move with the giveaway: the per-stream number flattered the thing that changed the traffic, and the per-viewer number told the truth. The hour is a confound of its own, and the fix is the same: compare like with like before you compare at all.
LiveWhat the avatar cannot do, and the room notices2 min▶
Human hosts in the Lumen Glow data reply to the room 95.6 times a stream. The avatars reply zero times, because a scripted presenter has nothing to hear with. Watch the room chart: both rooms fill on the first passes of the script and peak on the third, and both fade as the script repeats. The avatar's room fades from a lower peak, and there is nothing in the script to stop it, because the thing that holds a fading room is the host noticing it.
Q&A minutes, when a host reads names and answers questions, add 0.96 points of CTOR within a stream. Small next to a CTA, but they are the minutes where a viewer decides the room is live and worth staying in, and staying is what watch time measures. The avatar can say every word of the script and none of the words the room asked for.
Self-studyDisclosure, and where the digital-human question goes next4 min▶
Where a synthetic presenter is allowed at all, it has to be labelled, and the labelling rules have hardened this year.
- TikTok Shop AI-Generated Content Restrictions (US, effective 2026-09-02): synthetic faces, voices and digital humans must be disclosed, and the same policy says "Do not use AI to make a product look more powerful or effective than it really is". A generated hand applying a generated glow is a claims problem before it is a disclosure problem.
- UK AIGC label: content marked "Disclosed by creator as AI-generated" carries that label permanently; it cannot be removed after the fact.
- EU AI Act Article 50 applies from 2 August 2026: transparency duties for AI systems interacting with people and for synthetic content, which covers a digital presenter talking to a European room.
Consent and likeness, whose face and voice the avatar is built from and what they agreed to, is a subject of its own and belongs to the AI + Digital Human course. Tonight's decision page only needs the first line and the per-viewer numbers.
The shot list and the host decision
★ 20 min · everyone buildsTwo pages. First, take the 12-minute cycle you wrote in session 3 and write the camera into it, minute by minute. Second, a one-page human-versus-AI decision that starts with the platform rule and prices the gap on like-for-like hours. Neither needs a spreadsheet.
Lay the 12-minute cycle from session 3 down the left of a page. Hook, build, demo, price, CTA, urgency, Q&A, in the order and minute counts you gave them. Four columns to the right: in hand, close-up, price card, host reads names. Tick each column for each minute where it applies.
Write the measured value next to each tick. In hand +1.38 CTR. Close-up +1.01 CTR. Price card +0.97 CTR. None of them CTOR. That last part goes on the page in words, "these move clicks, the script moves orders", because the first person who reads the shot list will ask why the order rate did not go up.
Check the shot list against the platform's show rules. Pin held at least 10 to 15 minutes: one pin per cycle, fine. Ten-plus products in the bag: list the two-pack and the rest of the range even when the cycle sells one bottle. No still or graphic over half the screen: the price card is a corner card. Tick all three at the bottom of the page.
Now the host decision page. Line one is the platform rule. "TikTok Shop promotional LIVE: real-time verbal interaction required, no AI-generated voices, no animated figure over 50 percent of the screen. An AI-voiced avatar host is not permitted here." Everything under that line is about other platforms and other formats.
Finish with the hours question on its own line. The avatar numbers looked cheap partly because they ran when nobody was watching. Whether Lumen Glow should stream at all at 03:00 is a question about the feed, not about the host, and putting it on its own line stops it from being smuggled into the avatar case.
Write the value next to the instruction, every time. "Hold the serum up during the CTA" is a note a host forgets. "Hold the serum up during the CTA: +1.38 points of click rate, +9.7 percent GMV a night on the bench" is a note a host keeps. The number is not for you. It is for the person being asked to change a habit.
Before session 5
25 min- Watch ten minutes of your best host's last stream with the sound off. Note each minute: product in hand or not, close-up or not, price readable on screen or not. Then watch the same ten minutes with the picture off. Write one sentence on what each version was selling.
- Take your shot list from tonight and give it to a host without explanation. Ask them to circle anything they do not already do. Those circles are the coaching notes for session 5.
- Find the Requirements for High-Quality Videos and LIVEs in Seller Center for your own market and copy the interaction rule word for word into your decision page. Rules differ by region and change by date; the page should carry its own.
- Take the last three streams by your two most different hosts and write down the hour each one ran. Before you compare any per-viewer number between them, ask whether the hours match. If they do not, say so on the page.
Three questions
A host starts holding the serum up during every CTA, with the words unchanged. Which number should you expect to move within her stream?
Within a stream, product in hand adds 1.38 points of click rate and an order-rate effect the data cannot tell from zero. The GMV lift on the bench, +9.7 percent, comes from more clicks reaching a CTA minute whose order rate is set by the script.
A vendor offers a fully AI-voiced avatar to host Lumen Glow's TikTok Shop LIVE overnight. What is the first line of your answer?
The numbers and the disclosure come after the rule, and on TikTok Shop the rule closes the question. The avatar data describes the trade on a platform that permits it.
Human streams average 19,850 dollars and avatar streams 8,132. Why does the decision page quote the same-hours comparison (GPM 1,910 vs 1,258) instead?
The hour is a confound of its own. Comparing 17:00 and 22:00 streams only, the avatar still trails on every per-viewer number, and now the gap is one finance cannot explain away with the schedule.