- 32% of founders stop after the first question. 15% complete all nine.
- The verdict on question one predicted completion: 78% of "strong signal" ideas finished, against 2% of "let us find out".
- Half of all ideas (51%) were diagnosed as two products at once.
- Three frameworks out of nineteen account for 90% of strategy recommendations.
- The average recommended MVP was 31 weeks and 16 features.
- 73% of products were told to price as freemium plus tiers, and 91% were positioned mid-market.
- Only 12% of launch plans recommended paid advertising.
- Claude Code and Cursor were chosen for export at almost identical rates: 78 and 77 of 85.
- The engine gave a mean market score of 83.9 out of 100 and never once told a founder to stop.
Where this data comes from
ShipFit takes a startup idea through nine decisions in a fixed order: is it worth building, who pays, what hurts, how to win, what goes in version one, how to charge, will they pay, how to launch, and what to export to an AI coding tool. A stage only opens once the previous one is decided. That makes the product an unusually clean record of how founders behave when they are made to choose, and of what an AI engine recommends when it is asked the same questions about hundreds of different ideas.
This report covers every idea submitted from launch in October 2025 to August 2026 2026: 637 ideas, 84 founders and 3,237 completed AI runs. The query was run against the production database on 11 August 2026.
How it was counted
- One row per idea. Where an idea was re-run, only the latest run counts, so a single heavy user cannot skew the totals.
- Each figure has its own sample size. Most ideas stop early, so every later stage has fewer ideas behind it. A percentage is only meaningful next to its denominator, and every one below states it.
- Free-text answers were normalised. The engine wrote Jobs-to-be-Done in 15 different spellings. Those were collapsed into one bucket by hand, with nothing left unmatched.
What this data cannot tell you
- It is a self-selected sample. These are people who chose to run an AI validation tool, not a random sample of founders.
- It measures ideas, not businesses. ShipFit sees what founders submitted and what the engine concluded. It does not see what got built or whether it worked.
- The engine's scores are the engine's opinion. A mean score of 83.9 is a statement about ShipFit's model, not about the market.
Where founders stop
The steepest drop is immediately after the first question. Almost a third of ideas never get past "is this worth building", which is the cheapest moment there is to stop. The shape after that is less obvious: a trough around pricing, then a rise. Founders who decide how to charge almost always go on to finish.
Of founders who start a structured validation process, 32% stop after the first question and 15% complete all nine.
- Stopped after "is it worth building"
- 181 ideas 31.9%
- Stopped after "who pays"
- 89 ideas 15.7%
- Reached pricing
- 19 ideas stopped there 3.3%
- Completed all nine
- 87 ideas 15.3%
Sample: n = 568 ideas with at least one completed stage
What it does not say: Furthest stage reached, per idea. It measures whether people finished a process, not whether anything got built.
The first verdict predicts the last stage
This is the strongest finding in the data. Founders whose idea got a strong signal on question one were 39 times more likely to finish all nine stages than founders whose idea came back "let us find out". The verdict on the first question predicted behaviour on the ninth.
There is an honest second reading. Founders may simply drop ideas the tool was lukewarm about, which would make this the tool working rather than a prediction. In practice the soft verdict behaves like a no: we looked at that separately.
Founders told their idea showed a strong signal on question one were thirty-nine times more likely to finish all nine than founders told to go and find out.
- Strong signal
- n = 57, mean stage 7.75 78% finished
- Worth exploring
- n = 29, mean stage 4.79 37% finished
- Promising, needs focus
- n = 352, mean stage 3.77 8% finished
- Let us find out
- n = 82, mean stage 2.84 2% finished
Sample: n = 520 ideas with a first-question verdict
What it does not say: An honest alternative reading: founders may simply abandon ideas the tool was lukewarm about, which is arguably the tool working rather than a prediction.
Half of all ideas are two ideas
At the strategy stage the engine names the single biggest flaw in the idea. The most common answer was not weak differentiation or a crowded market. It was scope: a founder describing two products and calling it one. Scope confusion outranked weak differentiation by ten to one.
The most common flaw in a startup idea is not that it is a bad idea. It is that it is two ideas.
- Trying to build two products at once
- 123 ideas 51.2%
- Aimed at the wrong buyer
- 43 ideas 17.9%
- High-risk approach chosen
- 37 ideas 15.4%
- Weak differentiation
- 13 ideas 5.4%
Sample: n = 240 ideas that reached the strategy stage
What it does not say: The diagnosis is the engine’s, generated from the founder’s own description of their idea.
Three frameworks do nine-tenths of the work
The engine can pick from nineteen strategic frameworks. Six have ever been picked, and three account for 90% of selections: Jobs-to-be-Done, Blue Ocean Strategy and 7 Powers. The choice tracks the buyer. Ideas for individual prosumers pull Jobs-to-be-Done, while ideas for small businesses and consumers flip to Blue Ocean.
Given nineteen strategic frameworks to choose from, the engine picks three of them nine times out of ten.
- Jobs-to-be-Done
- 94 ideas 39.2%
- Blue Ocean Strategy
- 82 ideas 34.2%
- 7 Powers
- 41 ideas 17.1%
- Product-Led Growth
- 18 ideas 7.5%
- Playing to Win
- 3 ideas 1.3%
- Sales-Led Growth
- 1 idea 0.4%
Sample: n = 240 ideas that reached the strategy stage
What it does not say: This is what ShipFit’s engine selected, not what worked. It observes no outcomes, and the sample is people who chose to run an AI validation tool.
This one is unflattering to us, and we are publishing it anyway. A tool that only reports the numbers making it look good is not reporting numbers.
Who founders are building for
Almost nobody submitted an enterprise idea. Prosumer tools, bought by one person with their own money, were the largest group by a distance, and nearly every product that reached pricing was SaaS.
Prosumer buyers outnumber enterprise ones roughly forty to one among ideas run through a validation tool.
- Prosumer
- 155 ideas 43.3%
- B2B, small and mid-size
- 94 26.3%
- B2C
- 93 26.0%
- B2B enterprise
- 4 1.1%
Sample: n = 358 ideas with a classified buyer
What it does not say: Three separate classifiers in the flow disagree with each other about business model. This cites the earliest and largest only.
Across 2,888 problems surfaced from 280 ideas, founders named about ten problems each, and rated the average one 70 out of 100 for intensity.
- Rated "must solve"
- 1,096 problems 37.9%
- Occurring weekly
- 41.4%
- Occurring daily
- 22.1%
- Mean intensity
- median 75 70.4 / 100
Sample: n = 280 ideas, 2,888 problems, 10.3 per idea
What it does not say: Intensity is scored by the engine from the founder’s description, so it inherits both their optimism and its own.
The seven-month minimum viable product
A tool built to cut scope recommended, on average, a 31-week first release with 16 features. The lean option averaged 14 weeks and was never once the recommendation. That is a fault in our engine, and it is published here because a report that hides its own bad numbers is not worth citing. What an MVP is supposed to be is a useful corrective.
Across 208 AI-generated MVP plans, the average recommended build was 31 weeks and 16 features. The lean option, which the engine never once recommended, averaged 14 weeks.
- Features proposed per idea
- 19.8 mean
- Features in the recommended build
- 16.0 mean
- Recommended build time
- median 30.5 31.4 weeks
- The lean alternative
- recommended 0 times out of 208 14.0 weeks
Sample: n = 208 ideas that reached MVP scoping
What it does not say: A tool that exists to cut scope recommended a seven-month first release, and recommended the same package every single time. Both are worth knowing.
This one is unflattering to us, and we are publishing it anyway. A tool that only reports the numbers making it look good is not reporting numbers.
Whatever the product, the pricing advice converges
Asked to price 173 different products, the engine landed on freemium plus tiers nearly three times in four, and called nine in ten of them mid-market. Some of that reflects a sample that is almost entirely SaaS. Some of it is an AI reaching for the most common answer. If you are pricing, a method such as Van Westendorp asks your buyers instead.
Asked to price 173 different products, an AI recommended freemium-plus-tiered 73% of the time and positioned 91% of them as mid-market.
- Freemium and tiered
- 127 ideas 73.4%
- Hybrid, multiple mechanisms
- 36 ideas 20.8%
- Usage-based
- 5 ideas 2.9%
- Positioned mid-market
- 158 ideas 91.3%
Sample: n = 173 ideas that reached pricing; 98% were SaaS
What it does not say: Almost the entire sample was SaaS, which limits how far this generalises. It is the engine’s recommendation, not an observed price.
This one is unflattering to us, and we are publishing it anyway. A tool that only reports the numbers making it look good is not reporting numbers.
Launch advice is overwhelmingly organic
Across 135 launch plans, community and organic channels made up four recommendations in five. Paid advertising was recommended 12% of the time.
Only 12% of AI-generated launch plans recommend paid advertising. Community and organic account for 81%.
- Community
- 122 recommendations 45.5%
- Organic
- 94 35.1%
- Paid
- 33 12.3%
- Partnership
- 18 6.7%
Sample: n = 135 launch plans, 268 recommended channels
What it does not say: A recommendation, not a result. Nothing here observes whether any channel worked.
Claude Code and Cursor are in a dead heat
The last stage exports the plan as configuration for an AI coding tool. Founders could pick several. Claude Code and Cursor were chosen at near-identical rates, and most founders took both. It is a small sample, but it is a direct record of tool choice by people about to build, which is rare. Setup notes for each are on the Claude Code and Cursor pages.
Among founders exporting a build spec for an AI coding tool, Claude Code and Cursor were chosen at almost exactly the same rate, 78 to 77, and most picked both.
- Claude Code
- 78 of 85
- Cursor
- 77 of 85
- Tool-agnostic spec
- 32 of 85
- Replit
- 1 of 85
Sample: n = 85 ideas that reached the export stage; multi-select
What it does not say: Founders could pick more than one tool, so the counts do not sum to 85. It records what they exported a spec for, not which tool they went on to build with.
The engine grades almost everything a pass
This is the finding that matters most for anyone using an AI to judge an idea, ours included. The engine has a verdict that tells a founder to stop. In 520 ideas it never used it, and it rated exactly one proposed solution in 238 as weakly differentiated. The full breakdown, and what to do about it if you rely on any AI validator, is in do AI idea validators ever say no?
Across 520 ideas, an AI market-research engine gave a mean opportunity score of 83.9 out of 100 and killed exactly none of them.
- Mean score
- 83.9 / 100
- Median score
- 82
- Scored below 50
- 7 of 520 1.3%
- Told to kill the idea
- across 3,151 completed runs 0
Sample: n = 520 ideas with a numeric market score
What it does not say: This is a statement about ShipFit’s model, not about the market. A kill verdict exists in the code and has never once been assigned.
This one is unflattering to us, and we are publishing it anyway. A tool that only reports the numbers making it look good is not reporting numbers.
Asked to score how differentiated 238 proposed solutions were, an AI rated exactly one of them five or below out of ten.
- Mean differentiation score
- 7.59 / 10
- Median
- 7.5
- Scored 5 or below
- 0.4% 1 of 238
Sample: n = 238 recommended solutions
What it does not say: The interesting finding is not that the ideas were differentiated. It is that the engine appears incapable of saying otherwise.
This one is unflattering to us, and we are publishing it anyway. A tool that only reports the numbers making it look good is not reporting numbers.
How to cite this data
The figures are free to reuse under CC BY 4.0. Please link to this page and keep the sample size next to any number you quote.
Vallée, X. (2026, October 22). Startup idea validation data: what 637 ideas show. ShipFit. https://shipfit.ai/research/startup-idea-validation-data
Ready to make your next product a success?
9 decisions between your idea and a product worth building.