Lessons learned after 6 months of building a language learning startup

For the past six months, I’ve been working on LevelText, a startup that helps you learn languages by reading authentic native material. Actually, calling it a startup is a bit of a stretch because our app never produced any revenue and only a handful of users.

Nevertheless, I worked on it seriously for about half a year with a cofounder, joined an incubator, incorporated a company, iterated on an idea and launched several times. During this process, I learned a lot about the startup ecosystem, web development, talking to users, and many aspects of running a business. In this article, I will tell the story of how we got here and what we learned.

Long before LevelText, building a language learning app had been on my mind for a long time. I had always been passionate about languages: I studied several languages including French, Spanish and Japanese, and I was a PhD student researching linguistics and natural language processing. Thus, during my idle time, I liked to brainstorm ideas about improving language learning. Surely, language learning can be improved with my AI and NLP expertise, right?

There were a lot of language learning apps out there already, the most famous one being Duolingo, but there are many alternatives including Babbel, Rosetta Stone, Busuu, and dozens of others. Most people in the language learning community liked to criticize these apps: they offered similar features like language games and flashcards, teaching you words and phrases, but you could use them for a long time without achieving much fluency in the language.

According to many language acquisition researchers and educators, a more effective method of learning languages was “comprehensible input”: essentially getting a lot of exposure to authentic material (whether it’s books, articles, TV shows, podcasts, etc). The language is ideally at a level slightly above what you can easily understand, but not so difficult that it’s overwhelming. This was in stark contrast to all the apps that taught language through unnatural games and exercises, thus, despite the crowded marketplace, I was confident that teaching languages using authentic reading material would be a revolutionary app idea.

Iteration 1: Learning languages through reading

Personally, I like to learn languages by reading books. My process was the following: while reading, I would mark all the words that I didn’t understand, then I would look them up in a dictionary and add them to my collection to be reviewed later. I was improving my Chinese at the time, and found this process quite tedious because inputting unknown logographic characters into a dictionary was nontrivial.

So my first idea was an app where you take a picture of a printed text where you highlighted some words, and it would use some computer vision to look up these words for you in a dictionary. Then, if you liked, you could use this information to create Anki flashcards to review later. We called the tool VocabAssist and my wife helped me implement the prototype.

At this time, I started to interview all of my friends who were into language learning. A few people agreed with me that reading in the target language was a pain point, but most people did not have physical books lying around, and preferred reading on their phone or computer. And people who owned books wanted to keep them clean rather than writing all over them.

Therefore, we redesigned the concept: the user would instead paste text into the app. The app would then use some NLP to decide which words you probably don’t know, given your language level, and look up their definitions in a dictionary for you. Yet a problem still remained: how were users supposed to find material to read? Making them copy/paste text from web pages would have been too clumsy.

After doing a bit of market research, we found two existing apps with a similar concept already: LingQ and Readlang. These two apps did a decent job of helping learners study from the text, with fully featured dictionaries and flashcards, and they even supported video. However, they had to already have some material on hand; otherwise, both had libraries of user-contributed content to read, but they were small and you were unlikely to find what you’re looking for if you had niche interests. It would’ve been much better if they could leverage all of the content available on the Internet.

Both my wife and I had other full-time responsibilities, so progress was slow. Neither of us knew much about web development (I was a decent backend engineer but knew very little frontend), so the app was barely functional. As a result, we never launched the app to the public and only demoed it to some friends. Before we could launch it for real, I would need a cofounder with more experience building websites, and probably handle business responsibilities as well.

Iteration 2: Search engine for French learners

At the end of 2021, I had just moved to Vancouver and was wrapping up my PhD; I had more free time and was ready to give this project a more serious effort. So I set up a profile on the YCombinator cofounder matching platform, and soon I matched with Jonny Kalambay. He was working as a software engineer a Yelp and also had a passion for language learning. We worked on a toy project as a trial of working together and then decided to become cofounders.

Jonny was a talented full-stack engineer who could build and launch prototypes within days. Following the advice of the books “Running Lean” and “The Mom Test”, we started to interview users of our target demographic who were interested in language learning, asking them how they learn languages, what were the biggest struggles, and what tools they wish existed.

It turned out that finding interesting reading material of an appropriate difficulty was a fairly common problem. Thus, LevelText was born. We decided to build a search engine tool for French learners: we would take the user’s search query, translate it into French, search the internet for articles about the topic, and finally, return a sorted list of the articles ranked by estimated reading difficulty.

Our MVP assumed the user spoke English and was studying French. We decided to support only one language in our initial launch, since this would allow us to test out the concept with minimal development. We picked French because it was the most widely studied language in Canada and both of us were familiar with it.

After launching it on ProductHunt and conducting user tests on some French learners, the response was lukewarm. The tool was novel and cool, but didn’t provide enough value for them to keep coming back to the site. Intermediate and advanced learners already had lists of places to find reading material, and most articles on the internet were too difficult for beginners. We didn’t address the problem of motivation, one of the biggest struggles for language learners, and depended on the user remembering to use our site.

From the business side, the idea was difficult to monetize. There are several options for making money from a consumer-facing software product. Placing advertisements is the lowest effort option, but requires a massive amount of traffic to produce sufficient revenue: possible for a social network, but usually not for language learning apps. Most language learning apps make money by charging monthly or annual subscriptions. But we would need to provide a lot more value to justify charging a subscription – a simple tool wouldn’t cut it.

Iteration 3: Daily bite-sized reading

Our idea to fix the motivation problem was instead of the user needing to make searches, why not send them articles to their email? They would only need to indicate their interests once during sign-up and then they would receive a lesson every day in their inbox. With the lesson prepared and ready to go, the friction needed to start studying would be much lower.

Each lesson consisted of a paragraph of simple-to-read French text accessible to beginners, and a multiple-choice question to test their comprehension. We started prototyping this concept, using the GPT-3 model from OpenAI to simplify French articles down to a simple paragraph, and manually writing the multiple choice questions. We then enrolled a handful of French learners from social language learning sites to validate the idea.

The results were promising: most of the learners opened the e-mail and answered the question for three days in a row, although we never tried to charge money for it (the first real test of product-market fit), and the amount of manual work to prepare lessons meant that we could not get a large sample size of users to experiment with. Also, our mini-lessons could be completed in under a minute, which we felt was too little content for people to pay for it.

Founder troubles

It was around this time that we were invited to interview with YCombinator. This was a big deal because getting in meant receiving $500,000 USD of investment and connections to a network of experienced founders and mentors. We didn’t expect to get this far since our product hadn’t demonstrated traction yet; I guess it was our founder credentials that landed us the interview. We spent a day or two rehearsing our pitch and practicing answering questions for the rapid-fire 10-minute interview.

The next day, we received a rejection letter from YCombinator, basically saying that many great founders have tried to break into the language learning apps market, but it is hard to build a business in this space. Language learning (especially for non-English languages) was a hobby for most people with no urgent need, and they urged us to consider solving a must-have rather than a nice-to-have problem.

Our team morale took a hit after the rejection, and truth be told, we had gotten lots of hints at this point that language learning was not a great business direction. However, language learning was what we had been doing from the beginning, and we had no idea what else we could pivot to. Jonny suggested switching to making software for language teachers, because this seemed like a less competitive space than language learning apps. But I wasn’t thrilled with this direction: neither of us had connections or experience in education, so we had no good understanding of the problems that teachers faced and the realities of their daily working environment.

We started to fight more and more over product and business decisions: what level of language learners to aim at, how much to rely on automation versus manual content curation, how much we should care about monetization, etc. Neither of us were more experienced than the other in business, and neither of us trusted the other to take the lead on it. About two weeks after YCombinator, Jonny decided to leave LevelText and start his own company to make tools for language teachers. To be fair, he agreed to leave me with full ownership of the company and all its intellectual property (i.e., the code that we wrote), but I was on my own.

Iteration 4: Personalized courses from native content

Up to this point, we had talked to many users and ran some experiments, but we had never actually tried to charge money for our service. In this iteration, my plan was to take what I had learned thus far and build a product that at least had a chance of succeeding. By “success”, I mean making at least minimum wage, or roughly $3000 CAD/month.

Occasionally, you hear of startups raising investment from the founders’ credentials alone, without any revenue, but we talked to some investors and found this to be mostly not the case. Only incubators would invest on the order of $100k in early stage startups with no revenue, but we thought this was too little to unlock any options that we can’t already do, like hiring engineers. All the bigger investors expected to see monthly recurring revenue of at least $5-10k/month before our request for a $500k seed round would be taken seriously. So investment or no investment, I would need to start making that much revenue in the near future.

My business model was to bundle lessons into courses, which students could buy at a fixed cost. Several successful companies sold educational content bundled into courses, such as LinkedIn Learning and Chessable: the idea being that if most users drop off after a few weeks, we make more money by charging a one-time course purchase cost at the beginning than from a subscription (which they would quickly cancel). Another advantage of the course model was that I only needed to produce a fixed amount of content, whereas if I sold subscriptions, I would need to continuously generate new content for as long as they are subscribed, an unattractive proposition given that the app was not fully automated and a lot of work still needed to be done manually.

With a price point of $20 per course, I would need to sell just over 100 courses per month to make minimum wage: an optimistic, but not an impossible quantity. I built a landing page and a form where users input which topics they wanted to read about. The student then receives one trial lesson for free, at which point they would need to buy a course to receive more lessons.

To prepare a lesson, I would pick a French article of around B1 difficulty on your desired topic, and run an algorithm to guess the most difficult French vocabulary in the article and provide English translations. Additionally, the lesson included a computer text-to-speech reading and an English translation of the full article. For $20, I would prepare 7 daily French lessons matching your interests.

I launched the platform, bought some traffic from Google Adwords, but to my disbelief, nobody bought the courses. To figure out what was wrong, I conducted another round of user interviews. Again, the responses revealed that we were far from product market fit.

Everybody complained that the pricing was too expensive for what we provided: other digital services like Spotify and Netflix cost a similar amount or lower, and you got a lot more. People also found the pricing model confusing and assumed it was $20/month for a subscription. They also complained that our app only helped you practice reading, with no exercises or anything to help you with speaking, listening or writing. Worst of all, users didn’t find much value in personalized lessons, which was supposed to be our unique selling point – they rated lessons catered to their personal interests about equally interesting as lessons about the default topics.

Reflecting on this feedback, I think what happened makes a lot of sense. Most users, when asked to provide a list of their interests, gave fairly generic responses like “sports” or “politics”: not very useful for personalization. I wanted to make at least minimum wage to justify all the effort I was putting into this project, so I worked backwards from the $3000/month revenue target and designed the pricing scheme to achieve this goal. But my users don’t care about me or how much money I wanted to make, they only saw that I didn’t offer much for $20, so of course they didn’t buy.

I gathered the user feedback and sketched out some ideas of how to improve the product for the next iteration. But honestly, I was running out of steam – after countless iterations and little to show for my effort, it became difficult to gather the focus and willpower to implement more features. Being a solo founder is not much fun: nobody else understood my problems or cared about my language learning startup. So I decided it was a good time to quit.

Why language learning is a poor business

Despite my passion for it, language learning is ultimately a weak business model, especially software-based apps. For English speakers, learning a language is mainly something that is done for fun; it is seldom a must-have. Learners typically get bored and quit after only a few weeks, when their initial motivation runs out and it becomes a chore. This makes it difficult to make a subscription-based model work (many language learning apps get around this by asking for payment of several months or a whole year upfront).

Most language learners don’t spend much money on their hobby – out of all the people I interviewed, hardly anybody had ever paid for an app. It’s hard to blame them – there are tons of free language learning resources on the internet, so there is rarely a compelling reason to justify taking out your credit card. Even when people do pay to learn a language, it is usually for human tutoring or physical books (which people expect to pay for), not software products.

Finally, the market is saturated with hundreds of language learning products offering similar features. There is no good way to measure results of language learning, so you have no way to prove that your method is better than the others in terms of educational efficacy. With no better alternative, most companies optimize for user retention, inevitably leading to an app filled with gamification features.

Was founding a startup the wrong choice for me?

There are several reasons why being a founder in this space was probably not the right choice for me personally. First of all, my main expertise is in AI: while useful, it is far from the biggest priority in the early stages of most consumer products. Language learning apps needed to be engaging and fun to use, and AI was at most a minor component. Before product-market fit, the main tasks in a startup are developing quick prototypes of web or mobile apps, networking on social media to find users, talking to those users, and aligning the product with their needs. As for AI, it is usually best to throw something together quickly with GPT-3 and improve it later if there is demand for it, because any more sophisticated machine learning would take too long relative to the benefits.

While I could probably learn web development, sales, and product management given enough time, I had no experience in any of these skills, meaning I have no competitive advantage over the thousands of other aspiring entrepreneurs, and would be at a disadvantage compared to founders with experience in these skills.

Another factor is personal risk tolerance: how long are you willing to work full-time on your startup without any revenue or investment? For me, I was willing to put six months. This short timeline drove me to only consider business models that would generate a couple thousand dollars a month of revenue right away, and kill the idea if this goal could not be reached. But in reality, it is common to iterate for several years before achieving this goal (eg: Readlang took about 3 years to reach $1000/month in revenue). Since I wasn’t prepared to risk several years, potentially with zero payout at the end, it would’ve probably been better to work on it on the side while working a regular job somewhere, instead of going all-in full-time with an overly optimistic timeline.

There is a common belief in the startup community that persistence in your startup is “good” and giving up is “bad”. Part of this is observation bias: you only read about successful companies that raise millions and get acquired, after a lot of failure and persistence, and never hear about the majority of startups that fail. This has led many talented young people to pursue the startup dream, often willing to fail at it for years. I don’t think this is a particularly noble act: there is nothing magical that happens when you incorporate a company and declare yourself to be a startup founder. Your “company” is only worth something if you can provide something of sufficient value to customers that they pay for it; if you have no revenue then you are basically no different from being unemployed (even if you write a lot of code and get a few people to use it).

Even though I didn’t succeed in my venture, this was a valuable learning experience. You learn much more by actually attempting a startup than just reading about them; the life of a founder is not as glamorous as portrayed in the media. Now I know what I don’t know, and have a better idea of what skills I’d need to have a better chance next time. That’s the end of my rather long post, I hope it will be useful to readers thinking about building a startup or a language learning app.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s