Where we start
Analyst session 1 sized Kettle & Co's Japan market from the top: a government count, a government rate, two shares. Tonight we size the same market from the other end, the way a sales or growth team would: count the buyers, price the basket, count the baskets. The household count is real, read from the 2020 Census. The three behaviour inputs are constructed Kettle & Co pilot numbers, labelled as such on the page and in the widget. The two halves meet on the bench in Analyst session 6.
★ What you walk out with tonight. A bottom-up chain for Kettle & Co written as one sentence, each input with a low, a middle and a high, and a note for each input saying which table, column, filter and period it would come from in your own data, plus the one trap that source is known for. That chain is the second half of the triangulation bench.
Four links: households, buyers, basket, baskets a year
5 min liveBottom-up sizing multiplies a population by the share of it that buys, by what a buyer spends per order, by how many orders a buyer places in a year. The first link is the same Census count the top-down uses. The other three describe behaviour, and behaviour is what your own data measures best.
LiveThe four links, and what each one means3 min▶
Households: 34,553,907. The two-or-more-person households of the 2020 Population Census, the same count the top-down chain starts from. It is real, and it is shared: both methods lean on it.
Buyers: 4.5 percent. The share of those households that buy the category online in a year. Constructed from Kettle & Co's pilot, with a low of 3.5 and a high of 5.5 percent.
Spend per order: 2,400 yen. The average order value. Constructed, low 2,100, high 2,700.
Orders per buyer a year: 3.5. Repeat behaviour. Constructed, low 2.8, high 4.2.
Multiplied through, the middle is 13.06 billion yen a year. Carry all three inputs to their lows together and to their highs together and the range is 7.11 to 21.55. The widget below prints both, and so does the bench in Analyst session 6, where it meets the top-down's 19.63.
LiveWhy bottom-up scales with your actual go-to-market2 min▶
A top-down size describes the market. A bottom-up size describes the customers your channels reach, at the price you charge, at the rate they come back. That makes it the number a plan can be built on: every link is a lever someone owns. Conversion belongs to growth, the basket to merchandising and pricing, repeat to retention.
It also means a bottom-up grows only when a link moves. Doubling the size on a slide requires saying which link doubles and who will move it. If nobody can name the lever, the doubled number is a hope. Price is one of those levers, and how buyers respond to it is its own discipline: Learn Economics with Phoebe owns elasticity and pricing from data.
Self-studyBottom-up is not automatically more honest2 min▶
Bottom-up feels grounded because the inputs come from your own customers. That is also its weakness: your own customers are the ones who already found you. The chain is only as good as the claim that pilot behaviour describes the households you have not met yet. Part 2 is about exactly that claim, one input at a time.
The number the growth team could argue with. Kettle & Co's head of data showed the marketing lead two sizes for Japan: 19.63 from the top and 13.06 from the bottom. The marketing lead ignored the first and spent twenty minutes on the second, because every link in it was a number her team moved every week. Constructed Kettle & Co example; the household count is real.
Three inputs, three sources, three traps
5 min liveEach behaviour input has a natural home: buyers from a pilot or a purchase panel, spend per order from the orders table, orders a year from repeat behaviour by cohort. Each home also has a well-known way of misleading you. Name the source and its trap together, in the same line of the brief.
LiveBuyers: pilot conversion or a panel2 min▶
The buyer share is buyers divided by the households you reached. A pilot gives you both halves for a small, self-selected slice of the market. A purchase panel gives you a broader slice, measured by someone else, with its own definition of "buys the category".
The trap: early adopters are rarely typical. The households that convert in a pilot are the ones who were looking for you: tea enthusiasts, people who follow the brand, people your launch ads were targeted at. Their conversion rate describes them. Multiplying it by every household in Japan assumes everyone else will behave like the keenest buyers, the same mix of a local rate and a national population that Analyst session 1 caught with Shizuoka.
LiveSpend per order: the orders table, mean vs median2 min▶
Spend per order is an average order value: the sum of order amounts divided by the number of orders, over a stated period. The mean is what multiplies correctly through the chain, because buyers times orders times the mean order gives back total spend. The median tells you what a typical order looks like. Report both, multiply by the mean, and say so.
The trap: a launch-month average is inflated by promotions. Launch bundles, gift sets and free-shipping thresholds push the basket up in the first weeks. Take the average from a period after the launch offers end, and write the period next to the number.
Self-studyOrders a year: repeat by cohort, not first month2 min▶
Orders per buyer a year comes from repeat behaviour: group buyers by the month of their first order, then count how many orders each cohort places over the following months. A cohort view stops new buyers, who have had no time to repeat, from dragging the average down, and stops a burst of first orders from looking like loyalty.
The trap: first-year repeat from a 3-month pilot is extrapolated. Three months of a cohort's orders tell you about three months. Turning that into a year assumes the repeat curve keeps its shape, and that assumption belongs in the brief as a constructed link with a wide range, not as a measured one.
The pilot that converted its own fans. Kettle & Co's first pilot ran on the brand's newsletter list. Conversion looked superb, and the analyst nearly multiplied it by every household in Japan. The fix was to report the newsletter conversion as the high end of the buyer range and to source the middle from the paid-social cohort, who had never heard of Kettle before the ad. Constructed Kettle & Co example.
Every input gets a low, a middle and a high
5 min liveA single number per input pretends the pilot measured the market. A range per input says what the evidence actually supports. Carry the ranges through the chain together and the size becomes a range too, which is the honest form to put in front of a decision.
LiveCarry the ranges together, as corners3 min▶
The simplest way to carry ranges through a chain is by corners: every input at its low gives the low size, every input at its high gives the high size. For Kettle & Co: buyers 3.5 / 4.5 / 5.5 percent, spend per order 2,100 / 2,400 / 2,700 yen, orders 2.8 / 3.5 / 4.2 a year. The corners print 7.11 and 21.55 around a middle of 13.06.
Corners are wide on purpose. The low corner needs the pilot to have overstated conversion, the basket and the repeat rate all at once. That is unlikely, so the corners are the edges of the plausible, not the expected range. Say that out loud when you show them, or a reader will treat 7.11 as a forecast.
Self-studyThe alternative: sampling the inputs2 min▶
Some teams prefer to sample instead: draw each input many times from a distribution across its range, multiply each draw through the chain, and report the spread of the results (a Monte Carlo). The result is usually narrower than the corners, because all-lows-at-once is rare in the draws.
It is a choice, and teams disagree. Sampling needs a shape for each input and a view on whether the inputs move together; a pilot that overstated conversion probably overstated repeat too, and assuming independence quietly narrows the range. Corners need no such assumptions and anyone can check them by hand. Pick one, write down why, and keep it the same across both halves of the bench. Which input moves the size most, one at a time, is the tornado in Analyst session 4.
"Just give me one number." Kettle & Co's BD lead asked for a single figure for the board slide. The analyst gave her 13.06 with 7.11 to 21.55 underneath, and one sentence: the low needs the pilot to be wrong three ways at once. The board spent its time on which of the three inputs to test first, which was the right conversation. Constructed Kettle & Co example.
Build the chain from your own orders table
★ 22 min · everyone buildsThe widget below holds the Census household count and Kettle & Co's three constructed pilot inputs, each with its range. It recomputes the chain the moment you move a slider. Work the five steps in order and write down what it prints; the inputs are constructed, so once you move a slider your numbers will differ from the ones quoted here.
Step 1 · Read the default chain. All three sliders at their middles. Copy each link and its tag, real or constructed, into your notes.
Step 2 · Move the buyer slider, and only that one. Drag it up, then down. Watch the size: it moves in exact proportion to the buyer share, because the chain is a pure product. Raise buyers by a fifth and the size rises by a fifth. Then set it back to the middle.
Step 3 · Move spend per order, then orders a year. One at a time, back to the middle in between. Note that each moves the size the same proportional way, and that the range under the size moves with it.
Step 4 · Write where each input would come from in your own data. One line per input, in the same four slots every time, plus the trap that source carries. If an input has no home in your data, write "no source yet" rather than leaving it blank.
Step 5 · Write the chain as a sentence. The way it will go in the sizing brief: one clause per link, each with its source or the word "assumed", and the range at the end.
Keep the sentence next to last week's. The top-down printed 19.63, the bottom-up 13.06: a gap of 1.50x. Analyst session 6 asks which assumption explains it. Note that both sentences start from the same 34,553,907 households; if that count were wrong, both halves would be wrong together and the gap would not show it.
Before Analyst session 3
30 min- Find your orders table and write the query that returns three numbers: distinct buyers, mean and median order value, and orders per buyer, each over a stated period.
- Split one input by acquisition source: your earliest or most loyal buyers against buyers who came from a cold channel. Write down how far apart they are, and which one your pilot mostly contains.
- Re-run spend per order excluding the launch period and any promotion window. Note how much the average moves, and which period you will quote.
- Give each input a low, middle and high with one sentence of evidence for each end, then carry them through as corners and say in one line why all-lows-at-once is unlikely in your data.
Three questions
Clear all three and Analyst session 2 gets stamped in your passport.
1 · Which input in a bottom-up chain is most at risk from early-adopter bias?
The buyer share is a rate measured on a self-selected slice and multiplied by every household. That is the same local-rate, national-population mix as applying one city's spend to a country.
2 · Why is the household count called a shared link with the top-down?
Include one-person households in both and the top-down prints 31.64 and the bottom-up 21.06, yet the gap stays 1.50x. Triangulation only checks what the two methods do not share; Analyst session 6 shows it on the bench.
3 · Why does a range per input beat a single point?
A point pretends the pilot measured the market. A range per input, carried as corners or by sampling, tells the decision-maker how much rides on the assumptions and which ones to test first.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map.
Analyst session 2 cheat sheet · pin this
You can build a bottom-up chain from your own data and say where each input comes from. Analyst session 3 covers the market neither chain can reach: when no table measures it and you have no pilot yet, you size it by analogy and proxy.