Using Customers

One row is one TikTok buyer id

Buyers are grouped on TikTok's opaque, stable per-shop buyer id — present on every order in the book. It is deliberately not email and deliberately not nickname:

• Every buyer email is a TikTok relay address, and a buyer can carry more than one of them. It identifies nothing and is not stable.
• A nickname is a display name, not a key. Nicknames are not unique: in any book of reasonable size a share of buyers will share theirs with someone else, common first names and bare punctuation marks most of all. Buyers also change their nickname over time.

Nothing has ever been mixed up, because every row is keyed on the id. But it is why the name you see is built the way it is.

The name, and the smaller name under it

The big name is the buyer's nickname. The greyer line under it is the most recent unmasked shipping recipient name, and it exists purely so two rows with the same nickname are tellable apart — it resolves very nearly every nickname collision on its own.

It is shown on every row rather than only on collisions, on purpose: the grid pages 50 rows at a time, so a "is this name duplicated?" check on the page you are looking at would stay silent for a collision that straddles two pages. The second line is suppressed only when it is the nickname again in different punctuation.

The coloured circle is an avatar generated from the buyer id, so a face keeps its colour between sessions. Its initials are taken from the letters and digits in the nickname only — real nicknames are full of emoji and decoration, and a naïve first-character rule would put an emoji in the circle as often as a letter.

The money columns cover the whole order book, cancellations included

Orders, Lifetime and AOV are computed over every order that buyer has ever placed in your shop, at any status. Nothing is excluded for being cancelled, unpaid or refunded. That makes the figures a complete record of the relationship rather than a settled-money figure — which is the right read for "who is this person", and the wrong one if you were about to quote it as revenue. Finance is the screen for revenue.

AOV is simply Lifetime ÷ Orders. Shows counts the distinct live rooms their orders came from; orders placed outside a live show have no room and count toward nothing here.

"Usually buys" — and the little bar next to it

The tags are the buyer's most-bought size and top two brands. They come from the AI transcript of the live show, projected onto each order — which is to say they exist only where transcript capture actually ran. The short bar beside them is the coverage meter, and it is the honesty control for the whole cell: hover it and it tells you what percentage of that buyer's orders the profile was actually built from.

Coverage is counted against the buyer's total order count, never against the orders we happened to find a transcript for. A profile drawn from 30 of 124 orders reads 24% and must not look like one drawn from 101 of 113.

Last order changes colour only for VIPs

The recency cell turns amber at 22 days and red past 45 — but only for a VIP. That is deliberate. A one-time buyer who has not come back in a month is the normal case and there are thousands of them; tinting those would bury the hundred that matter. On a non-VIP the cell is always neutral, however old it is.

The six segment tabs

All · VIP (10 or more lifetime orders) · Repeat (more than one) · Lapsing (a VIP quiet for 22 to 45 days) · Return watch · One-and-done (exactly one order). Every threshold is in the reference table with the reasoning behind the number.

Lapsing, and the Lost flag that has no tab

Lapsing and Lost are VIP-only by construction: both ask "has a good customer gone quiet?", and the question is meaningless for someone who bought once. Lapsing is 22–45 days of silence — a window chosen because it is still early enough to act. Past 45 days the pill becomes Lost, which is rendered on the row but has no tab of its own. To find them, open VIP and sort by Most recent; the Lost ones sit at the bottom.

Return watch needs both halves of its rule

A buyer is on return watch when they have 3 or more lifetime returns and their own return rate is at least twice the shop's. The count is the evidence — three returns is a pattern, one is a Tuesday. The rate is the signal: three returns across two hundred orders at a 1.5% shop rate is exactly average, and flagging that buyer punishes them for being loyal.

The rate is always computed on distinct orders that came back, never on the raw count of return records. TikTok can mint several return records against one order — there is a real buyer with one order and seven of them — and dividing records by orders renders 700%, which reads as a bug rather than a fact.

Sorting, and the one thing search does that you would not expect

Sort offers lifetime value, order count, most recent and average order. Search matches the nickname, the shipping recipient name, or an exact buyer id.

The three stats, then the profile

Lifetime with the date of the first order, AOV with the order count, and Last order in days. Below that, Usually buys — the same size and brand story as the grid cell, but stated in full: a coverage banner in words — "62% coverage. Built from the 78 orders that have a transcript", give or take whatever this buyer's real figures are — then size bars, then brand bars with each brand's share, then the colours mentioned on air. Under 50% coverage the banner turns amber and tells you to treat the profile as indicative; at 80% or above it turns green and says it is solid enough to act on.

Brand names are folded case-insensitively before counting, because the same brand is stored under several spellings — Figs and FIGS are two rows in the database. Unfolded, a buyer's top brand can be split in half and lose to a runner-up. The spelling shown is whichever occurs most, so you see the form you actually use.

Colours are counted the same way and are deliberately not mapped to a colour taxonomy. "Onyx", "Espresso" and "Candlelight Yellow" are the words used on air to sell the item; folding them into black, brown and yellow would trade the seller's own vocabulary for a guess.

Shows, returns, recent orders

Shows they buy from ranks the live rooms this buyer actually turns up to, by spend, up to twelve. Orders with no room are left out rather than bucketed as an unnamed show — they are ordinary non-live sales.

Returns shows the raw count, the amount refunded, how many distinct orders came back, and the resulting rate. When the buyer is on return watch, a line above it states the comparison in full — "4 returns across 29 orders — 13.8%, about 9x the shop average of 1.5%". That comparison is withheld when it would rest on fewer than three orders: at one order the only non-zero rate that exists is 100%, which against a 1.5% shop rate is automatically "67x" for everybody, before anything about the person is considered. The counts and the rate are still shown; only the multiple is held back.

Recent orders lists the last 25 with their show, item count, amount and status — returned orders are marked. Click one to open the full order.

Dates, shows, status, state, size, brand, segment

Every option in every dropdown carries the number of orders it would yield. Those counts are computed with every filter applied except the one you are choosing in — so picking one show does not make every other show read "0" and become unfindable. For the same reason, every option the shop has ever had stays listed even when the current slice reduces it to zero: a list that drops options on selection makes the thing you just picked impossible to find again.

The date presets are relative to the last order in your shop, not to today. A shop that has not broadcast for a fortnight would otherwise get an empty "Last 7 days" and read as broken.

The scope line and the eight KPIs

The line under the filters — "4,102 of 21,782 orders in scope · 2,731 buyers", to invent some numbers — is the one place that tells you how much of the shop you are looking at. Underneath, Buyers and Orders carry their share of the whole shop, and GMV carries the shop total to compare against.

First-time share counts orders that are the buyer's first ever order in your shop, not their first in the window you selected. That is what makes it an acquisition number rather than a windowing artefact. States reached is how many US jurisdictions this slice reached out of all fifty-one there are — fifty states plus DC, which is what the incl. DC line underneath is telling you.

Six metrics, five classes, two ways to cut them

The metric switcher sits above the ranked list: Buyers, Orders, GMV, AOV, Return % and Index. The colour ramp has five classes over the states that have orders in the current slice, and you choose how the cuts are made:

• Quantile puts an equal number of states in each class, so the map reads as a ranking. Its cost is real — your largest state and a mid-sized one can land in the same class.
• Magnitude uses equal-width bands on a compressed scale, so the map reads as absolute size and the top two or three genuinely stand out. Plain linear bands would put forty-nine states in class one, which is why the scale is compressed.

Neither is "the" right answer, which is why the screen offers both instead of picking silently. Grey means no orders in this slice — never "no customers"; it is a separate fill and never a step of the ramp.

Hatched states: what we refused to say

A ratio over a handful of buyers swings wildly — one buyer in a small state can read as 2.5x over-indexed, and a map that paints that dark states it as a fact. So the server applies two floors: a state needs 5 buyers before it gets an index, and 10 orders before its AOV or return rate is treated as a value.

Below those floors the row is still returned — the counts are real and you should see them — but it is drawn as a hatch, not a colour, and it sorts to the bottom of the ranked list regardless of value. A hatch cannot be misread as a value the way any fill from a ramp can. Among the hatched rows the ordering is by buyer count, so the ones nearest to qualifying come first.

Index vs all — the metric the filters exist for

Index compares this slice's state mix with your shop's overall mix. 1.00× is exactly your normal mix; 1.6× on Texas means Texas is 60% over-represented in whatever you filtered to. It is drawn on a diverging ramp with a neutral grey middle, so "normal" reads as nothing rather than as a value.

Everything the map encodes is repeated as text at the bottom of the page under The map as a table — buyers, share, orders, GMV, AOV, return rate and index per state, with the same hatching rules expressed as em dashes. Rows are clickable there too.

Prefer ZIP3 over the city field

ZIP3 is the first three digits of the shipping postcode — a real postal area, tagged with the state it most often ships to in the current slice. It is the trustworthy sub-state unit.

When they buy

Orders by weekday × hour, Monday first. The clock is US Central for every buyer — a single approximation for a US-wide buyer base, used consistently across the whole product so two screens can never disagree about which day a 1am order belongs to. It is daylight-saving aware. Switch Heat to Values to read the counts; hovering any cell gives the count and its share.

Size, brand and colour — read the coverage bar first

These three panels are the only thing on the screen that resembles a demographic attribute, and they exist only where transcript capture ran. Each card carries a coverage bar — "Size known · 1,240 / 4,102 · 30%", to keep inventing numbers — and that bar is the first thing to read, because the shares below it are computed over the orders that have the dimension, not over all orders. Bars that summed to 53% because half the orders were never transcribed would read as a data error.

Size is the one ordinal panel: XS→4X has a natural order, so the rows keep that order and the colour ramps along the scale rather than along the ranking. Buckets that are not points on that scale — Bra, Numeric, Volume, Kids', One size — sort after it and are drawn neutral, because they have no position on it. Numeric covers any size given as a plain number, waist sizes among them, so it is not read as a measurement of one particular thing. Brands and colours are nominal, so every bar takes the same hue and length carries the value alone.

Clicking a size or brand bar filters the page to it. Colours are display-only.

Frequency buckets read lifetime orders, always

The purchase-frequency buckets (1, 2, 3–4, 5–9, 10–24, 25+) count a buyer's lifetime orders, not their orders inside your filter. A five-plus buyer stays a five-plus buyer when the screen is scoped to a single show — which is what lets you ask "how much of this show's revenue came from my regulars?" and get an answer. Toggle between counting buyers and counting GMV.

First-time vs returning plots one column per day that actually had orders. Days with no orders are absent rather than drawn as zero — a shop that broadcasts on half the calendar days would otherwise render as a comb of empty columns.

Audience by show

One row per show in the slice: buyers, first-timers, % new, orders, GMV, AOV, the top state with its share, and unique viewers. Click a row to scope the whole page to that show — the map then answers "where did this show's buyers live", which is the question the index metric exists for.

First-time counts buyers whose first order in your shop ever landed in that show — the acquisition read, and the reason a show can report far more buyers than first-timers without anything being wrong. AOV blanks to an em dash on a show with too few orders to average honestly — the same sufficiency floor the map applies, set and applied on the server. The show's counts stay exact either way; it is only the ratio that is withheld, because an average over a handful of orders swings on one of them.

The four figures across the top

Value at risk adds up the lifetime value of every VIP currently in the win-back window. Read it as an upper bound on what is in play, not a forecast: it prices each buyer at everything they have ever spent, and a buyer who drifts away rarely takes all of that with them.

Lost within a week counts the ones past day 39 — they leave the window on their 46th quiet day and stop being counted as winnable. Comes back in N days is the median recruit rate across every show old enough to judge; the card names the window in its own heading, so read the number there rather than assuming one. In the room is the selected show's buyer count, with the milestone breakdown underneath. Pick All shows and the same card becomes Customers — every buyer you have, split by where they stand today.

In the room — who to name on air

Pick a show and this lists the buyers in it. It opens on your most recent show, and a show that is currently live carries a LIVE badge in the picker. Worth a shout-out is the default and the point of the panel: it narrows to buyers who either hit a milestone in this show or are VIPs. Everyone drops that filter.

Three milestones can appear. 2nd order — first repeat is the cheapest win in the book: the single hardest gap to close is one order to two. 10th order — new VIP marks the crossing. Back after N quiet days carries the gap because the gap is the point — “back” alone gives you nothing to say, where “back after 61 quiet days” does.

The strip above the table splits the room by kind — first-time, second order, repeat, VIP — with the spend each group put through tonight. A kind showing zero is drawn greyed rather than hidden, so the shape of the room is legible at a glance.

All shows is the top row of the picker, and it answers a different question: not what one broadcast did, but who your customers are. The heading changes to Across all shows, the Tonight column is replaced by Shows — how many broadcasts that buyer has bought in — and each row is labelled by standing rather than by milestone. The shout-out filter greys out while it is selected, on purpose: a milestone is something one show did to a buyer, so with no show picked there is nothing for it to narrow to.

Standing counts lifetime orders the same way the Buyers tab counts them, so a VIP here is a VIP there. Cancelled and unpaid orders are excluded from both. A show’s roster deliberately does not exclude them — it describes what happened in the room, and an order that was placed and later cancelled still happened.

Slipping away — the win-back queue

Every VIP between 22 and 45 days quiet, longest silence first. That sort is deliberate and it is not the same as sorting by value: the row at the top is the one closest to falling out of the window, not the one worth the most. Sorting this queue by lifetime value would bury the deadline, which is the only thing that makes it a queue rather than a list.

The Window column bands each row — freshest, most winnable through halfway to lost and lost within a week. Export CSV writes the rows you can see, with the buyer's name, days quiet, lifetime value, orders and what they usually buy.

Windows still open

The table below it ranks shows whose win-back window has closed — a finished result you can do nothing about. This one is the opposite: recent shows whose recruits are still inside their window, so there is still time to bring them back. Each row is one show, and the column that matters is Still to win — the first-time buyers it recruited who have not bought again yet.

A show whose recruits have all come back is not listed. That is a result, not work, and it belongs in the table below once its window closes. The list is ordered by deadline, soonest first, and it scrolls inside the card rather than being cut short — what you see is all of it.

Which shows recruit

A show recruits when someone places their first ever order in your shop during it. It counts as a keeper when that person orders again in a later show, inside the win-back window. The rate is keepers ÷ recruits, and the bar shows it against the shop median rather than against 100% — the median is the only honest comparison, since no shop converts every first-timer.

A show only appears once it clears the recruit floor and its window has closed. Both matter, and the second is the one people forget. At three recruits the rate can only be 0, 33, 67 or 100%, and a show that recruited three people and kept all three would otherwise top the table over one that recruited sixty-eight and kept fifty. The floor is set where a rate stops moving in steps too coarse to rank, and the window is long enough to cover several weekly cycles — so a show is judged on more than its first weekend.

Every profile says "no size data"

Size, brand and colour come from AI transcripts of your live shows. A shop that has never run transcript capture has none, and every profile is legitimately empty. Capture and transcribe a show and they start filling in.

A buyer's lifetime total looks too small

Check whether the search box is still filled in. A search filters the underlying orders, so a matched row's totals cover only the matched orders. Clear it, or open the buyer — the drawer is never search-scoped.

The map is almost entirely grey

Grey is "no orders in this slice". Either the filters are tighter than you meant, or you are on a money metric without finances.view — in which case the metric switcher drops GMV and AOV entirely rather than painting a grey map, so if you can still see them, it is the filters.

A state's index looks enormous

If it is drawn as a value at all, it cleared the five-buyer floor. If it is hatched, the screen is telling you it will not stand behind that number — read the buyer and order counts in the table instead.

Flags on a buyer row

Buyers grid columns

Returns arithmetic, precisely

Demographics: sufficiency floors and segments

Retention thresholds

Where each part of the profile comes from

What this screen cannot see

Stated plainly, because several of these are things a screen called "Customers" is normally expected to know:

Read the coverage bar before you believe a profile

"Usually buys M" from 4 of 60 orders and "usually buys M" from 55 of 60 look identical until you look at the bar. It is on the row, in the drawer and on every Demographics dimension card for exactly that reason.

Work Lapsing before it becomes Lost

The 22–45 day window exists because that is where contact still changes the outcome. Past 45 days the pill turns Lost, which is a record rather than an opportunity.

Treat return watch as a prompt, never a verdict

Three returns at twice your shop rate is worth a second look at how the item was sized and described. It is not evidence of anything about the person, and the screen deliberately does not tell you what to do about it.

On Demographics, set a filter before reading the index

Unfiltered, the index compares the shop with itself and is 1.00× everywhere. Scope to a show, a size or a brand first — that is the comparison the metric was built for.

Quote ZIP3, not the city field

If a concentration finding is going to be acted on — a regional shipping decision, a targeted show — take it from ZIP3. The city column contains counties and cannot be cleaned up honestly.