The triangulation bench

Top-down here is households times survey spend times two shares. Bottom-up is households times the share that buys online times spend per order times orders a year. They share one link, the household count, and nothing else. Every other link comes from a different source: a government survey on top, Kettle & Co's own pilot underneath.

That is what makes the gap informative. If the survey spend were wrong, only the top-down would move. If the pilot's repeat rate were wrong, only the bottom-up would move. But if the household count were wrong, both would move together and the gap would not notice. Try it on the bench: switch the population to "all households" and both sizes rise, to 31.64 and 21.06, while the gap stays at 1.50x.

So the honest claim is narrow: "the two methods agree within 1.50x on everything they do not share". The shared link needs its own check, which for a Census count is reading the table carefully, as Analyst session 1 did.

Practitioners disagree, and both sides have a point. One camp prefers bottom-up for any plan: it is built from things you can observe, customers, prices and repeat orders, and it scales with your actual go-to-market. Its critics answer that a bottom-up built from a pilot inherits the pilot's early adopters, who are rarely typical. Another camp trusts top-down because it rests on large official samples; its critics answer that the shares that narrow a top-down to your segment are usually guesses, as they are here.

The position taken on this bench is a working one, not a verdict: neither method is the answer alone, and the useful output is the gap plus the assumption that explains it. A team that only ever builds one method has no gap to learn from.

The top two inputs are both top-down shares: share bought online closes 76 percent of the gap, premium share 68 percent. That is a finding, not a coincidence. Those two shares are the links nobody measured; the survey spend under them is real. The tornado in Analyst session 4 showed the same order: share bought online swings the gap from 1.10 to 2.00, the widest bar.

No input closes 100 percent on its own, but the ranges already overlap from 10.90 to 21.55, so the methods are consistent. The sizing brief should say so, and should name the share bought online as the input to measure before anyone commits a budget.

"Every tea drinker is our customer" removes the premium and online filters. The top-down becomes households times all leaf spend: 198.24 billion yen a year, 10.1 times the triangulated top-down. The bottom-up does not move, because Kettle's pilot customers did not change. The gap becomes 15.18x, the ranges no longer overlap, and the best single input closes 7 percent.

That pattern, a gap no input can close, is the signature of two methods sizing two different things. It is what a TAM looks like when somebody puts it next to a bottom-up built from real customers. Business session 2 reads the same hero assumption on a pitch deck.

Switch the category on the bench. "Tea, all of it" adds bottled and canned tea drinks: top-down 148.96, gap 11.40x, and the best single input, share bought online, closes 13 percent. "Wellness drinks" adds mineral water, sports drinks and lactic acid drinks: 271.00, gap 20.75x, best input 10 percent. Premium loose-leaf: 19.63, gap 1.50x, ranges overlap.

The bottom-up never moved, because it is built from the customers Kettle actually has. So the category toggle is a test of the category: a frame whose top-down cannot be reconciled with your own customers by any input in range is a frame your customers are not in. That is evidence for marketing's positioning call in Business session 1, not a replacement for it.

Press the break button and the top-down is replaced by a constructed 50.00 billion yen from a report that does not publish its method. The gap against the bottom-up is 3.83x. The bench cannot say which input explains it, because one side has no inputs: no population, no rate, no shares. Nobody can tell whether the report sized a wider category, a different year, a different channel, or made a mistake.

Analyst-firm market sizes are contested, and the disagreement is worth teaching straight. Some teams use them as an outside anchor: a figure from people who looked at the market for longer than you did. Others treat any figure without a published method as unusable, because category definitions vary between firms and the firm that sells the report has its own reasons to make the market look large. On the bench the practical rule is narrower and both camps accept it: a number without a method can sit beside a triangulated range as context, but it cannot be one side of the triangulation.

Set the population to "all households, one-person included" and read the bench: top-down 31.64 (range 17.58 to 51.13), bottom-up 21.06 (range 11.46 to 34.74), gap 1.50x, share bought online still explaining 76 percent. Everything bigger, nothing different. If the one-person assumption is wrong, the bench will never tell you, because both chains multiply by the same households. The defence is upstream: read the population of every table you use, as Analyst session 1 did.