Is Your AI Startup Idea a Tarpit? How to Tell Before You Start
Founder Playbook · Beyonds Journal · 2026-11-05
On October 29, 2023, the founder of bunni.ai, a small app for chatting with PDF files, asked a question on Indie Hackers, a forum for founders of small, mostly self-funded companies: "Did OpenAI just kill my product?" OpenAI had released a "Chat with PDF" feature for ChatGPT that day, the founder wrote, "threatening to clobber the revenues" of apps like this one. Near the end, the post stayed calm: "if I'm focused and execute well, I think they can co-exist." In May 2024, OpenAI gave free ChatGPT users the option to "Upload files for assistance summarizing, writing or analyzing."
The idea had looked good to many people at once. There's An AI For That, a directory of AI tools, lists "78 alternatives" to ChatPDF, one of the best-known apps of this kind (as of October 29, 2026). Chatbase also began as a chat-with-PDF app, and its founder, Yasser Elsaid, described the next stage in one line in an Indie Hackers profile: "Within months, competitors flooded the space."
78 — Alternatives to ChatPDF listed by one AI tools directory. There's An AI For That, 29 Oct 2026
Was chat with PDF a tarpit? Y Combinator uses that word for an idea that many founders come up with, that looks like an unsolved problem and wins warm early feedback, and that keeps failing. Dalton Caldwell, a YC managing director, described it in those terms on Lenny's Podcast, a show about building products, in April 2024. His main example was an app for friends to decide where to go out at night, which founders have been starting, he said, since the 1990s. A crowded market is a different thing. Our reading: an idea turns into a tarpit when founders keep piling in on the strength of praise, without asking why the founders before them left, and that can happen in a crowded market or an almost empty one.
An idea turns into a tarpit when founders keep piling in on the strength of praise, without asking why the founders before them left.
If your own idea is drawing warm praise, and you can already name products like it, the 4 checks below take an afternoon and show whether you are in a tarpit.
What makes an idea a tarpit?
YC's Michael Seibel and Dalton Caldwell walked through the pattern in a 2022 YC video on ideas founders should avoid. Their examples were consumer ideas: an app to discover new things such as restaurants, yet another social network for some group of people, software for betting on things. In the video's own image, each one looks like a pool of fresh water that nobody is drinking from. Many founders want to build these products, and the real demand is smaller than the supply of founders. Their advice: study why the founders before you failed, and go in knowing how high the bar is.
Paul Graham described the same trap from the founder's side in "How to Get Startup Ideas" in 2012: "At YC we call these 'made-up' or 'sitcom' startup ideas." They are the kind a TV writer would give a character, plausible and bad. What makes them dangerous, in Graham's account, is the feedback they get. Run the idea of a social network for pet owners past friends who have pets, he wrote, and "They say 'Yeah, maybe I could see using something like that.'" That sentence feels like validation. Caldwell makes a similar point on the podcast: good early validation is part of what makes an idea a tarpit.
Is a crowded market a tarpit?
Often it is the opposite. In the same essay Graham wrote: "A crowded market is actually a good sign, because it means both that there's demand and that none of the existing solutions are good enough." His condition is that you have a thesis "about what everyone else in it is overlooking."
We look for the marks of a tarpit on the playground, as this issue's Lead calls it: every way one job gets done in a particular place, under that place's rules, with or without a product. The founders who tried to do this job better and gave up are part of its history, and that history is what a founder in a tarpit hasn't read.
What does the pattern look like in AI?
By our count, chat with PDF shows 2 of the marks. Many founders built it at once, and the job looked as if nobody had done it well yet. It is too early to say whether it keeps failing, the last mark in YC's description. 3 years after its founder's post, bunni.ai's site still sells chat with PDF on pay-as-you-go credits and says it has "8,500 users and growing" (as of October 29, 2026); it gives no revenue figures. The job itself, understanding a long document quickly, is old: people skim, read the summary, ask a colleague or pay someone to read it for them. The apps that arrived in 2023 joined that playground, and so did ChatGPT, as the opening shows.
Chatbase took a narrower road. Elsaid "quickly realized that focusing on PDFs alone wouldn't be enough," he wrote, so the company moved to AI agents that answer customers on behalf of businesses. By his figures, it grew from $64,000 in monthly revenue in May 2023 to "over $5M ARR," annual recurring revenue, by early 2025.
Source — Chatbase's revenue figures are its founder's own, published on Indie Hackers in February 2025.
Jasper, an early leader in AI writing tools, named a new CEO in September 2023. 3 weeks later it repositioned as "an end-to-end AI copilot for better marketing outcomes," to move AI assistance "from simply creating content faster to helping marketers perform better." Around the same time, according to The Information as reported by Maginative, it cut its internal valuation by 20%.
In AI, as we see it, the tarpit pattern can form within months around a feature. Many founders see the same demo at once, early users try it out of curiosity, and the player that ends up doing the job for most people can be a platform that adds the feature to a product people already use. Both companies in this section moved to a place where the job was more pressing: a business with customers to answer, a marketing team with results to show. These sources don't show how much of each change came from ChatGPT and how much from the crowd of similar tools.
My first users love it. Doesn't that prove it isn't a tarpit?
It shows people are curious. In consumer AI, that curiosity can fade fast. RevenueCat, which runs subscriptions for mobile apps, studied more than 3,500 AI-powered apps and reported in August 2026 that they "churn at 30% higher rates than traditional subscription apps," while earning 41% more revenue per paying user in the first year. A consultant quoted in the study described apps that are "barely used long-term, because users came from a trending video showing one specific feature."
30% — Higher churn in AI-powered apps than in traditional subscription apps. RevenueCat, 18 Aug 2026
41% — More revenue per paying user in the first year. RevenueCat, 18 Aug 2026
The data covers mobile subscription apps, so it says little about tools sold to businesses. Our reading: in the first weeks, sign-ups measure curiosity, and the second and third months show whether people really have this job to do.
In the first weeks, sign-ups measure curiosity, and the second and third months show whether people really have this job to do.
How do you check whether your idea is a tarpit?
4 checks, each answered on paper. Together they take an afternoon, and most of the time goes on the first.
- List who has tried it before, and why each one stopped. Look for shutdown notices, founders' post-mortems and pivots. A long list of attempts that stopped for reasons you can't explain is the warning to take most seriously.
- Add the general tools to the ways people get the job done. In this issue's Lead, listing those ways is part of reading a playground. For a tarpit check, add the general tools your customers already pay for, ChatGPT among them. If one of them does the job well enough, write down what would make someone switch to you.
- Separate praise from commitment. "Maybe I could see using something like that" is praise. A pre-order, a paid pilot or an hour of someone's time with their real data is commitment. Count the second.
- Narrow until the job is pressing. Graham's image: you can dig a hole "that's broad but shallow, or one that's narrow and deep, like a well." Name the place where the job comes up every week and costs someone money when it goes wrong.
What can you do this week?
Do the first check. Spend 2 hours finding 3 products or startups that tried your idea before you, and write one line for each: why it stopped, or what it turned into. If you find none, ask 3 people who have to get this job done how they handled it the last time. Either list shows you the playground you are about to enter.
Commitment from the people who have to get the job done is where AI Startup Reality starts. From December 1, 50 founders spend 4 months building their AI companies with us. First, we find what people will pay for, and we aim to get the first payments before anyone writes code. Applications close on November 27.
Route — AI Startup Reality: market pull first, then the company. From December 1, applications until November 27.
FAQ
What is a tarpit idea?
A tarpit idea is Y Combinator's term for a startup idea that many founders come up with, that looks like an unsolved problem and gets warm early feedback, and that keeps failing. YC's examples include apps to discover things such as restaurants, and yet another social network for some group of people.
Is a crowded market bad for a startup?
Paul Graham called a crowded market "actually a good sign," because it shows there is demand and that existing solutions fall short, as long as you know what the others are overlooking. A tarpit is marked by repeated failed attempts and polite interest, and that pattern can appear in a crowded market or an empty one.
How do you know if your startup idea is a tarpit?
Look for earlier attempts whose failures you can't explain, feedback that is warm but asks nothing of the people giving it, and a job that general tools already do well enough. YC's advice includes studying why the founders before you failed.
Can an AI startup compete with a ChatGPT feature?
Chatbase is one example. It started as a chat-with-PDF app and, by its founder's figures, grew after moving to customer-support agents for businesses. A product that does what a general chatbot does, for the same people, has a weaker case.
What are examples of tarpit ideas?
In a 2022 YC video: apps to discover new things such as restaurants, yet another social network, and software for betting. Dalton Caldwell, a YC managing director, gives the example of an app for friends to decide where to go out at night, which founders have tried, he said, since the 1990s.
Sources
- Indie Hackers, "Did OpenAI just kill my product?" October 29, 2023: link
- bunni.ai, home page, viewed October 29, 2026: link
- OpenAI, "Introducing GPT-4o and more tools to ChatGPT free users," May 13, 2024: link
- There's An AI For That, ChatPDF listing, seen October 29, 2026: link
- Yasser Elsaid with James Fleischmann, "From viral side project to a $5M/yr B2B AI platform," Indie Hackers, February 2025: link
- Lenny's Podcast, "Lessons from 1,000+ YC startups: Resilience, tar pit ideas, pivoting, more," with Dalton Caldwell, April 18, 2024: link
- Y Combinator, video with Michael Seibel and Dalton Caldwell on tarpit ideas, November 23, 2022: link
- Paul Graham, "How to Get Startup Ideas," November 2012: link
- Beyonds Journal, "How a Blue Ocean Is Created: Before the First Line of Code, Look at the Playground Your Product Is Aiming to Play On": link
- Jasper, "Jasper expands to become an end-to-end AI copilot for marketing teams," PR Newswire, October 18, 2023: link
- Jasper, "Jasper Welcomes Former Dropbox President Timothy Young as CEO," PR Newswire, September 28, 2023: link
- Chris McKay, "Jasper Appoints New CEO and Cuts Internal Valuation as AI Growth Slows," Maginative, September 29, 2023 (citing The Information): link
- Margarita Loktionova, "We studied 3,500+ AI-powered apps to see why some retain users better than others," RevenueCat, August 18, 2026: link
Beyonds Journal, issue 003