MenHair App

A few months ago I made a small web app over a weekend, to show my friend how easy it is. He told me about how he looked for an AI hairstyle try-on app, and he just found a bunch of mobile apps which: 1. require you to install the app 2. you need to sign up and go through long onboarding 3. hard paywall - you can't even try the app for free 4. super expensive, like $12.99 /week

But the biggest problem was that they're all focused on something else, like photo filters, or female styles. The men's hairstyle was just a tiny feature that didn't have a lot of options and it didn't work well.

I was like "well that's a simple AI wrapper I could make by myself quickly", and probably in a better way than the competitors (lol).

I made it in two days (like 5h of work total for the mvp), shared it with my friend and launched in November, free to use. No marketing, no social media, and then I forgot about it for almost two months. I'm working on some other stuff and never thought about participating in this industry in any way.

I looked at Google Trends to see how people search for this kind of apps and then named it MenHair App.

The PMF?

Then in late January I got an invoice from Google Cloud for the Gemini API cost (I've been using Nano Banana 1 for the image generation back then) and saw a number I wasn't expecting.

I quickly checked the analytics and realised that for a month I had about 10 people a day with high engagement rates, like 92% of them engaging, and ~75% successfully generating at least one image successfully.

I had no clue where they came from and why.

I quickly added limits (first to 3 free images, then to 1), and then added the paywall.

They still kept using it and people started paying for additional images.

I saw it as a good sign of product-market fit and then invested some time to add more features and hairstyles.

In early March the number of users started going up sharply, as well as the sales.

I finally found out what drove the growth - it was the LLMs. To be precise - ChatGPT and Copilot. They brought over 70% of that traffic.

It was incredible - people asking "how do I try a hairstyle before the barber?" were getting sent directly to my app.

Sales notifications kept coming.

The problem

So I decided to put some extra effort in SEO and the performance improvement. I looked for the best practices in both, and the app worked better by each day.

Then suddenly, the number of users went down from ~100 /day on average, to 40, then 30, and then 20 in a few days.

The sales curve became flat.

Here's what happened - while improving the app, I read some articles and spoke to several AI agents that convinced me to switch from llm.txt to llms.txt format, claiming that it's the standard and it's what the agents are looking for, and that keeping both files just makes them confused and forces them to move away from my app.

So I deleted llm.txt and kept llms.txt...

The traffic dropped to near zero overnight.

Once I switched to llms.txt the traffic went down and the LLMs stopped sending people to my app. The new file was well structured, improved version of the previous llm.txt, but it didn't matter.

I kept it for a while, hoping the new file would get picked up by them, but nothing, I got to like 3 visitors /day.

Did I kill the product that was on a great trajectory?

The solution?

About 10 days ago, I brought back the llm.txt file as the source, and kept llms.txt as a carbon copy.

Nothing happened for the first 5 days, but then - ChatGPT started sending people my way more and more.

The numbers are slowly rising again.

Is it safe to think it will go back up - probably yes. But there is hope :)

The conclusion

First, and most important - if you're building an indie product and getting AI referral traffic - don't touch what's working until you understand exactly what's pointing at it.

Of course, it sounds obvious, but when you notice the traffic and sales increase, you get tempted to make it better, thinking it would move the needle further towards having a successful and profitable product.

Well, lesson learned the hard way.

Second - as it turns out, different AI crawlers use different conventions.

Some look for /llms.txt. I don't know which ones, cause I didn't notice any new traffic.

Others look for /llm.txt. From my experience - ChatGPT and Copilot.

There's no single standard yet.

The fix? Serve both. They're identical files.

/llm.txt ✓ /llms.txt ✓

One file, two paths, traffic back.