Where the leader track ends up
Session 1 split every stream into what the room was given and what it did with it. Session 2 showed the giveaway buying the first and paying for it with the second. Session 3 wrote the say, session 4 the show, and session 5 put a five-minute nowcast in the producer's ear with three cues worth giving. Tonight all of it goes onto one page, and the page gets two things it did not have: a way to check whether a host followed it, and a way to check whether following it paid. The Lumen Glow data supplies the numbers; the shape is yours to reuse on any brand.
★ What you walk out with tonight. A host scorecard built on per-viewer numbers, GPM, CTOR, watch time, replies, with a compliance line above them and the given line kept out of the score. A test design you can explain to a finance director in plain words: ten-minute blocks, cue on and cue off, inside the same night. And the claims list that keeps the account alive, in one table, with what to say instead. Stapled together, that is the playbook: the say card, the shot list, the three cues, the scorecard, the test plan, the claims list. One page, defensible.
LiveWhy "we told the host and GMV rose" is not evidence3 min▶
In session 2 the giveaway looked like the best idea in live commerce. Nine of the ten top streams by GMV opened with one, and a regression across streams put its value at 5,576 dollars a night. Then the per-viewer numbers said the opposite: GPM 1,547 with a giveaway against 1,983 without, CTOR 11.16 against 13.10. The giveaway bought traffic and the traffic wore the giveaway's name.
A before-and-after claim about a cue has the same problem in a different coat. You gave the instruction on Monday. Tuesday's stream did better. But Tuesday's room was also different: a different feed, a different boost, a different day of the week, perhaps a different host. Nothing in "before and after" separates the cue from the night it landed on. The room-size correlation from session 1 is the reason: GMV tracks peak room at 0.96, and peak room is the platform's decision, not the host's. Any night-to-night comparison is mostly a comparison of two feeds.
The test you want is inside one night, not across two. If the cue is on for some minutes and off for others in the same stream, the feed, the day and the host are all held constant by construction. That is the whole idea of Part 2, and it needs no statistics to explain.
Self-studyThe vocabulary tonight adds4 min▶
- Scorecard - one page per host per period with the given line (not scored) and the did line (scored), plus a compliance line above both.
- Switchback - a test where the same unit, here one stream, alternates between treatment and control over time. Cue on for a block, cue off for the next, in a random order.
- Carryover - the effect of one block leaking into the next. A CTA at minute 38 still sells at minute 42, so the first minutes of an "off" block are not really off.
- Baseline - what the host would have done without the cue. In a switchback, the off blocks of the same night. In a before-and-after, a different night, which is the problem.
- Policy-risk flag - a minute where the transcript contains a phrase the platform's misleading-content guides name, or an urgency run past its second minute. Session 5's bench counts them; the scorecard reports them.
- Disclosure - the label that says a face, a voice or a digital human on screen was generated. Required wherever such a presenter is allowed at all.
The scorecard
7 min liveA scorecard has one job: make it impossible to praise a host for the platform's traffic or blame her for a quiet Tuesday. So it has two lines. The given line is recorded and never scored. The did line is scored, and every number on it divides the traffic out. Above both sits a compliance line, because a Risk label costs future traffic whatever the night sold.
LiveWhy Ren is the interesting row3 min▶
Read the scorecard column by column and Ren wins one metric outright: the longest average watch time of any host, 9.89 minutes. Read the compliance line and he is the only host with flags, 6.82 a stream. Read CTOR and he has the lowest of the four humans, 11.76 against Maya's 13.03. Those three facts are one fact.
Ren stretches the urgency. Session 3 showed what that does: the first two minutes of an urgency run convert at 15.12 and 15.44 percent, the third drops to 8.30 and it never recovers. A stretched urgency pitch holds people, watch time goes up, because nobody leaves while they think the price is about to move. It sells less, because the room stops believing the clock on the third minute. And every one of those minutes is a phrase the platform's misleading-content guide names. Ren's row is the urgency trade in one host: watch time bought with order rate and policy risk.
Without the compliance line, Ren's scorecard reads as a strong host with a soft close. With it, the row tells you exactly which habit to coach and why it is not optional.
LiveWhat goes on the given line, so a host is never scored on it3 min▶
The given line is everything decided before the host opened her mouth. It is recorded so the did line can be read fairly, and it is never scored because none of it was the host's call.
| Given | Who decided it | Why it stays off the score |
|---|---|---|
| Hour and day | The schedule | Session 4: 95 percent of human streams were evening, 26 percent of avatar streams. A per-stream number that ignores this is a schedule score |
| Giveaway opener | The brand or the MCN | Session 2: buys views (10,364 vs 7,546), costs GPM (1,547 vs 1,983). Scoring the host on views after a giveaway credits the gift |
| Traffic source and peak room | The platform | Session 1: GMV tracks peak room at 0.96. The feed is not the host's |
| Host type | The casting decision | An avatar's zero replies is a property of the format, not a performance |
The test in Part 1 of the shape is simple: could the host have changed it during the stream? If not, it is given. If so, it is did.
LiveReading the scorecard with a host2 min▶
A Lumen Glow monthly review used to open with a GMV league table, and the same host topped it every month because she had the Friday 20:00 slot. When the scorecard replaced it, the given line showed the slot, the did line showed her GPM in the middle of the humans, and the host who had been "quiet" on Mondays moved to the top on CTOR. The conversation that month was about scheduling for the first time, and about the Monday host's script for the first time. Nobody's numbers changed. What they were compared against did.
Self-studyThe platform's own labels map onto the scorecard3 min▶
TikTok Shop's LIVE Diagnosis stamps every stream Risk, Need attention, Can improve or Normal, and the three that matter sit on three different lines of the scorecard.
- Risk is the compliance line. Violation information attached; not a performance label. It outranks every did-metric because it costs the next stream's traffic.
- Need attention is the did line compared with the host's own past: a drop against herself. The scorecard's period-over-period column.
- Can improve is the did line compared with the category benchmark: a gap against the market. The scorecard's benchmark column, when you have one.
Keeping the scorecard in the platform's shape means a host sees the same diagnosis from you and from Seller Center, and the two never argue.
The test, in plain words
7 min liveTake one 120-minute stream and cut it into twelve blocks of ten minutes. Flip a coin for each block: cue on or cue off. Compare orders per thousand viewer-minutes in the on blocks against the off blocks of the same night. The feed, the day, the host and the platform's boost are identical on both sides, because they are the same stream. That is a switchback, and it is the whole test.
LiveWhy a switchback beats before-and-after, and what it costs3 min▶
Before-and-after compares Monday's room with Tuesday's. Everything that differs between the two nights, the feed, the boost, the day, the mood, lands on the cue's account. A switchback compares minutes 30 to 40 with minutes 40 to 50 of the same stream. Same room, same traffic, same night, same host. The platform's boost, which decides most of a night's GMV, is on both sides of the comparison and cancels.
The cost is carryover. Minutes are not sealed boxes. A CTA at minute 38 puts the product in a cart that checks out at minute 42, inside the next block, which was supposed to be "off". Left alone, that leak makes the off blocks look better than they are and shrinks the measured effect. The fix is a washout: drop the first two minutes of every block from the comparison. You lose a fifth of the data and keep the honesty.
The formal version, how long the blocks should be, how to randomise them, and how carryover biases the estimate, is worked out in Bojinov, Simchi-Levi and Zhao, "Design and Analysis of Switchback Experiments", Management Science 69(7), 2023. You do not need the paper to run the test. You need it when someone asks why the blocks are ten minutes and not one.
| Design | Compares | Holds constant | Weakness |
|---|---|---|---|
| Before and after | Two nights | Nothing | The feed and the boost take the credit |
| Host A vs host B | Two people | The night, if scheduled together | Every difference between the hosts |
| Switchback | Blocks of one night | Room, traffic, night, host | Carryover; needs a washout and several nights |
LiveWhat the bench already showed, and why a real room still needs the test3 min▶
Session 5's replay bench ran tonight's stream, Jules on a Friday at 20:00, 40 times under each policy and averaged. Against the host's own plan at 18,198 dollars: the playbook, all three good cues together, 23,166 dollars, +27.3 percent, watch time 9.90 minutes, peak room 885. Stretching the urgency when viewers drop: +0.5 percent, with 3.18 policy-risk flags a night. A mid-stream giveaway when viewers drop: -11.2 percent, even though views rose to 8,732 from 8,519.
Those are means over 40 seeded nights, computed live in the browser from the same generator that produced the 60 streams. They are the strongest possible evidence that the method finds what the data contains, because the data's causal structure is known. They are not evidence about your room. Your hosts, your feed and your viewers are not the generator's, and the size of every effect will differ. What transfers is the shape of the finding: the cues that pay per viewer, the one that pays nothing and costs flags, the one that buys views and loses money. The switchback is how you get your own numbers to put next to them.
Self-studyWhen you cannot randomise3 min▶
Some cues cannot be switched off for ten minutes without the room noticing: a price card that appears and disappears at random will read as a glitch, and a host told to stop replying to names for a block will not do it. For those, the fallback is the within-stream comparison the builder track uses: each stream compared only with itself, minute by minute, which is what produced the +6.50 CTA and +0.96 Q&A numbers. It is weaker than a switchback because the host chose when to do the thing, and a host who says the CTA when the room looks ready is not choosing at random. It is still far stronger than two nights side by side.
Randomisation, power, how many nights you need for a given effect size, and what to do when the blocks cannot be independent are the subject of the experimentation course. Tonight you need the plain-words version and the washout.
The claims that end an account
5 min liveA cue that adds half a percent and three policy flags a night is not a trade; it is a countdown. Everything the track has said about urgency and claims sits here in one table, with where each one is banned and what a host can say instead. This page goes at the back of the playbook and at the top of every host's first week.
LiveThe compliance list, in one place5 min▶
Two readers, one list. The platform reads it as a Risk label and, repeated, a traffic limit or an account removal. The regulator reads it as a deceptive commercial practice. The phrases are the same on both lists.
| Claim | Where it is banned | What to say instead |
|---|---|---|
| "100% effective", "Guaranteed results", "Works for everyone", "Permanent effect", "Instant results", "No side effects" | TikTok Shop Misleading Content guide (banned absolutes); Content Policy (eff. 2026-09-17): no "exaggerated promises about a product's performance, effect, or price" | What the product is and does, in words the label carries; what it did on the host's own skin, on camera |
| "Was 59 dollars, tonight 29" when 29 is the everyday price | Fair Pricing Policy (eff. 2026-09-01): reference prices must "reflect a genuine and established selling price"; no "Artificially inflating the original price" | "29 dollars, the price it always is. The two-pack is 49 tonight" - and the 49 is real |
| "Sale ends today", "Only 2 hours left" when it does not and there are not | Misleading Discount guide (eff. 2026-08-31): must be genuine; FTC "Bringing Dark Patterns to Light" (2022): False Limited Time Message, Baseless Countdown Timer; CMA207 guidance (18 Nov 2025) countdown-clock example | Use the LIVE Flash Sale tool with the real end time. The platform draws the clock and can check it |
| "While stocks last", "Don't miss these big discounts", "We are having our largest sale!", "This sale is for selected customers only" | Misleading Content guide flagged phrases; FTC: False Low Stock Message | The real count in the bag tonight, said once, and updated when it changes |
| "We stop selling this after tonight" when you do not | UK DMCC Act 2024 Sch. 20 para 7 (in force 6 April 2025): falsely stating a product "will only be available for a limited time ... in order to elicit an immediate decision"; EU UCPD Annex I item 7; UK Content Policy (eff. 2026-01-09) | Nothing about stopping unless it is true; the two-minute urgency cue from session 3 works on a real bundle price |
| Urgency run past its second minute, every cycle | The same flagged phrases, repeated; the bench counts 3.18 flags a night for +0.5 percent | Two minutes, then move to the CTA or the Q&A. Session 3's numbers: 15.12, 15.44, then 8.30 |
| An AI-voiced presenter hosting the LIVE, or a synthetic face undisclosed | Requirements for High-Quality Videos and LIVEs (US, eff. 2026-09-17): real-time verbal interaction, no AI-generated voices; AI-Generated Content Restrictions (US, eff. 2026-09-02): disclose synthetic faces, voices, digital humans; EU AI Act Article 50 from 2 August 2026 | A human voice, live. Where a synthetic element is allowed, the disclosure label on |
| Generated footage that makes the serum look more effective than it is | AI-Generated Content Restrictions: "Do not use AI to make a product look more powerful or effective than it really is" | Real product, real skin, real light. The close-up ratio from session 4 does the persuading |
Read the list as "when", not "never". "Only two hours left" is fine when the flash sale ends in two hours. "Was 59" is fine when it was. Every line on this list is a true statement that becomes a violation the moment it stops being true, which is why the playbook says when a cue applies and never says always.
The one-page playbook
★ 20 min · everyone buildsSix pieces, all of which you already have or will write in the next twenty minutes: the say card, the shot list, the three live cues, the scorecard, the test plan for the next four streams, and the compliance list. Assemble them in that order. The rule for the whole page: every instruction carries a per-viewer number, a "when", and a date it will be tested.
Top third: the say card and the shot list, side by side. Session 3's 12-minute cycle with its CTOR value per segment; session 4's camera columns with their CTR values. One line under both: "the words move the order rate, the picture moves the click rate."
Middle: the three live cues from session 5, each with its bench value and its trigger. CTA on a dip, +14.6 percent. Product in hand on every CTA, +9.7 percent. Answer the room every ten minutes, +4.4 percent. Then the two the room begs for and the bench refuses: stretched urgency, +0.5 percent and 3.18 flags; mid-stream giveaway, -11.2 percent.
The scorecard, blank, in the shape from Part 1. Given line at the top and greyed. Did line: GPM, CTOR, watch, replies. Compliance line above the did line. Write next to the compliance line what a flag is: a phrase from the list in Part 3, or an urgency run past two minutes.
The test plan for the next four streams. One paragraph, plain words. Which cue. Switchback in ten-minute blocks, order drawn at random before the stream. Washout two minutes at the head of every block. Metric: orders per 1,000 viewer-minutes, on blocks minus off blocks, per night, then averaged over the four. Who holds the block order, and who tells the producer.
Bottom: the compliance table from Part 3, unchanged. Do not summarise it. The phrases are the point; a host needs to recognise them in her own mouth.
Read the page for the word "always". Every one you find is a claim that will be false on some night. Replace each with the condition under which it is true. Then read it for "GMV" and replace each with the per-viewer number it was standing in for.
The playbook is a hypothesis with a review date, not a policy. Every number on it came from 60 streams of one brand and 40 replays of one night. Yours will differ. The test plan at the bottom is what turns the page from something you believe into something you measured, and the review date is what makes finance stop asking.
After the track
30 min- Fill the scorecard for your hosts' last month from Seller Center. Given line first. Then rank them by GPM and by flags, and write one sentence on any host whose two ranks disagree.
- Draw the block order for your next stream before it starts: twelve blocks, coin flip each, written down and given to the producer in a sealed message. After the stream, compute on-minus-off with the washout. Do not look at GMV until you have.
- Take your current host guidelines and count the word "always". Rewrite each one as a "when". Bring the before and after.
- Read the Misleading Content guide for your own market in Seller Center and add any flagged phrase not already in the Part 3 table. Then listen to one of your own streams for them.
Three questions
You told the hosts to hold the serum up on every CTA from Monday. Tuesday's GMV was 30 percent higher than Monday's. Why is that not evidence the cue worked?
The giveaway in session 2 looked worth 5,576 dollars a night across streams and lost money per viewer. Before-and-after has the same flaw: everything that differs between the nights lands on the cue's account.
In a switchback with ten-minute blocks, a CTA at minute 38 leads to an order at minute 42, inside an "off" block. What does that do, and what fixes it?
Carryover leaks the treatment into the control. Dropping the head of every block costs a fifth of the data and keeps the comparison honest. Bojinov, Simchi-Levi and Zhao (2023) work out how long the blocks and the washout should be.
Three lines from a draft playbook. Which one would earn a Risk label from the platform and a deceptive-practice reading from the regulator?
A false low-stock message and a baseless price deadline are named in TikTok Shop's misleading-content guides, in the FTC's dark-patterns report, in UK DMCC Sch. 20 para 7 and in EU UCPD Annex I item 7. The other two lines are the +9.7 and +4.4 percent cues, with no flags.