We publish one idea-sourcing run unedited, with every rejection and its reason attached. That is the only reason this post can exist: you cannot itemise deaths you did not record, and almost nobody records them.
The run: 69 candidates sourced from 41 sources, 24 of which were opened and read in full. 36 were dropped. 33 survived the checks. 5 reached the shortlist.
The most common cause of death is not a crowded market, a competitor, or the price being wrong. It is that there was no documented problem behind the idea at all — 21 of the 36, more than twice the next cause.
Now the part I would have got wrong reading the chart. Add the bars and you get 57 reasons against 36 dead ideas. The census looks like it partitions the drops and it does not: 21 of them failed more than one check.
Implementation note on that number, because it is the kind of thing that silently inflates. We count DISTINCT checks per dropped idea, not raw reason tokens — two spellings of the same check on one idea is one failure. Counting tokens would print a count of two against a row and then contradict it with the multi-failure line beside it. On this run both countings agree, so the fix was latent rather than a moved public number, which is exactly when it is cheap to make.
What the overshoot means in practice: a weak idea is rarely weak in one dimension. The thing with no evidence of a problem behind it is usually also the thing that is hard to sell, because both come from the same missing piece of work. If you are hoping that answering the one objection you have heard will save an idea, the arithmetic says it probably will not.
On the biggest row specifically: it is a judgement about the evidence, not about the market. The sweep looks for a documented complaint the idea would answer and did not find one. There are two honest readings — the problem is real but undocumented, or the problem was invented by the idea — and the check cannot tell them apart, which is exactly why it drops the candidate rather than pronouncing on the market.
The second reading is the one that costs months, because nothing about it feels wrong from the inside. An idea assembled from a plausible-sounding gap looks exactly like one assembled from a real gap, right up until launch.
Three limits, stated before anyone leans on the counts. These are our checks rather than the market's verdict. The ratings underneath are model output sampled fresh each run, so a second sweep would not reproduce this list. And it counts only what the sweep could see — a problem that lives in somebody's inbox is invisible to this method, which is a ceiling rather than a caveat.
The whole census, and the run it is counted from: https://whittleos.com/guides/why-startup-ideas-fail
Top comments (0)