Snap a photo, get paid: two high schoolers used this formula to make $30 million, and you can copy it

Open the App Store, search “AI,” and ten of the top fifty results are photo apps. The AI products that have actually been printing money over the last two years look pretty “low-rent.” They are not the all-knowing large models that discuss philosophy, write code, or generate movies. They are small, almost absurdly simple tools that make you wonder, “people pay for this?” You point your phone, snap a photo, the app spits out a one-line answer, and you hand it ten dollars a month.

The most important numbers first.

An app called Cal AI does exactly one thing: you snap a photo of your meal, and it tells you the calorie count. It was built by two seventeen-year-old high schoolers and launched in May 2024. Eighteen months later, it had crossed fifteen million downloads and thirty million dollars in annual revenue. In December 2025, MyFitnessPal acquired it, and the entire seven-person team came on board. Forbes projects Cal AI will hit fifty million in revenue by 2026.

Seven people. One photo feature. Thirty million dollars.

A replicable filtering logic sits behind this, and it has been validated more than once. I am going to break that formula down, and I will also work through the “I do not know how to code” objection with real numbers.

1. Users do not lack information; they lack the will to think

Most people building products go about finding pain points the wrong way. They assume users lack information, so they hand it over, and the user pays.

That is wrong. In 2026, information is the cheapest thing on the internet. Want to know the calories in a plate of pasta? A search engine spits out twenty answers in three seconds, free.

So why does Cal AI get to charge for the same information?

Because when someone on a cut eats out, they get a menu with twenty dishes. They can read every one of them, but they have no idea what to order. Is the fat too high? Will the carbs blow their limit? If they eat this today, do they need to burn it off tomorrow? In theory, they have to: identify the ingredients, estimate the portion size, look up the caloric density of each item, add it all up, and compare against their remaining daily allowance. That takes five minutes. Three meals a day, ninety meals a month.

Behavioral economics calls this decision fatigue: the more decisions you make, the worse the next one gets. What the user actually lacks is a definitive answer that tells their overthinking brain to shut up.

Someone built that. You snap a photo of the menu, the AI tells you what to order, and ten minutes of agonizing becomes a three-second decision.

Apps in this niche are doing well. Take MenuFit on the App Store: over one million downloads, 4.8 stars, 53,900 ratings, and a base subscription of $9.99 a month. In the subscription space, a million-plus downloads paired with a ten-dollar monthly fee is a real, lucrative business model.

2. Three filtering criteria, miss one and the model fails

Stack these cases and you can extract three screening criteria. These are for picking a direction, not a magic business formula.

High-frequency scenario × Decision paralysis × Photo input

Why is this only viable in 2026? Because the first two conditions have always existed. The third had to wait for multimodal models to mature. If you previously forced users to type “150g chicken breast, 100g broccoli,” the input friction was higher than doing the math themselves. The product would die on arrival. Now, snapping a photo is the only required action. The effort drops from typing thirty characters to a single tap. Only now are all three conditions aligned.

I will go through them one by one, because most people faceplant on exactly one of these pillars.

Condition 1: high frequency

It needs to happen at least once a day, ideally three times. Eating is the perfect scenario: three times a day, a thousand times a year. When a user opens your app three times a day, the psychological barrier to paying drops close to zero. On the flip side, a need like “help me calculate which mortgage plan to choose” comes up maybe twice in a lifetime. You can only charge once. And one-off revenue means you are stuck buying traffic forever.

Condition 2: decision paralysis

The user has to be in a moment where they must make an immediate decision but cannot pull the trigger. Standing in front of a menu, staring at store shelves, browsing a wine list. These are decision bottlenecks. “I want to learn about nutrition” is not a bottleneck. That is studying. Studying is not urgent, and non-urgent demands have bad conversion rates.

Decisions made under emotional stress have one defining feature: users are paying for relief, not information. A guy on a diet snaps a photo of his plate at a dinner, reads “320 calories” off his phone, and digs in without guilt. He did not pay for the data. He paid for the liberation of not having to agonize in that exact moment. That is the actual source of willingness to pay.

Condition 3: photo input

Multimodal models are the engine of this wave. TechCrunch reported that Cal AI’s recognition accuracy sits around 90 percent, not 100. For someone cutting fat, a three-second answer that is 90 percent accurate is far more valuable than a five-minute answer that is 100 percent accurate. Fat loss is a statistics game, not an exact science. A daily margin of error under 5 percent gets the job done.

3. Five steps from zero to launch

Follow these five steps in order.

Step 1: pick a scenario, not a feature

Grab a piece of paper and write down the moments in your own life where you have to make an immediate decision but always hesitate. Only write the ones that have genuinely stumped you. Do not write what you guess will stump others. A few directions for inspiration:

• Bar menus. Which drink has the lowest calories? A real pain point for the fitness crowd, and this demographic has spending power.
• Convenience store snacks. Snap the back of the packaging and tell me if I can eat this during my cut.
• Airport and train station dining. I have a two-hour layover. How do I eat without blowing my diet?
• Supermarket ingredient lists. Snap the label and tell me if it contains anything I am trying to avoid.

That last direction has already proven itself. Seed Oil Scout, an app that helps users avoid restaurants and foods containing seed oil, has public metrics of over one million downloads and 14,000 ratings at 4.8 stars. They built a store inside the app where users can buy compliant products, which means they cash in on two revenue streams at once: subscription fees and affiliate commissions.

Step 2: validate with niche data, not Google Trends

Once you have locked in your scenario, do not write a single line of code yet. Spend twenty minutes validating whether the demand is real or just a figment of your own imagination.

Here is the trap. Using Google Trends and search autocomplete dropdowns does not work for App traffic. Search behavior in the App Store is fundamentally different from Google, and the metrics for keyword popularity are not the same. Cal AI did not scale through Google in the first place. It rode TikTok video conversion rates and Product Hunt rankings to the top.

The right way to validate:

1. Open the App Store, connect to a VPN in your target market (the US, for example), and search your core English keywords (like menu calorie). Check if the top ten results include similar apps with over 100,000 downloads that are still getting regular updates.
2. Head to Sensor Tower, a third-party App Store data platform, and pull the download curves for the top products under that keyword. See if the trend is rising or falling.
3. Search relevant subreddits on Reddit and threads on Quora for phrases like “X decision” or “X help.” Gauge the complaint density. The more people venting, the more real the pain point.

Competitors that are still alive plus Redditors repeatedly asking the same questions means the demand is real, not just a hallucination.

Step 3: build it, do not start from scratch

Cal AI’s tech stack is no secret. TechCrunch laid it out. It taps into models from Anthropic and OpenAI, uses RAG (Retrieval-Augmented Generation, a technique that lets the AI pull answers from external knowledge bases), and feeds in a batch of open-source food calorie image databases. No proprietary models, no algorithmic breakthroughs. It is a product that recombines existing capabilities.

In 2026, your technical barrier to entry looks like this:

• Visual recognition. Tap into the multimodal APIs of GPT or Claude, the APIs that let the AI “see” and interpret images. Pay per request, so your startup costs are close to zero.
• App interface. Ship an MVP in 4 to 8 weeks. Add 2 to 3 more weeks for IAP (in-app purchase) approvals, privacy policy, and tax classification.
• Subscription billing. Use App Store or Google Play in-app purchases. The platforms handle global payments and refunds for you.

The real challenge is whether you have the guts to slash your feature list down to one. Cal AI has no social features, no recipe library, no workout plans. It does one thing, hits 90 percent accuracy, and cashes in.

Your first instinct with V1 will be to pile on features. Do not do it. Every feature you add dilutes that three-second moment of certainty.

Step 3.5: the playbook for non-coders

This question comes up more than any other, so I will call it out.

If you have zero coding skills, there is a viable path in 2026:

1. Build your first version with Glide, Bubble, Softr, or Adalo. These are no-code platforms where you piece together an app by dragging and dropping UI components. They can now tap into the GPT-4o vision API, so building an app that recognizes images and returns text is technically possible.
2. At the same time, hire a freelance Swift developer on Upwork at $30 to $50 an hour. Swift is the iOS development language. The freelancer rewrites your no-code prototype into a native app ready for the App Store.
3. Teach yourself Cursor and Claude Code, two AI coding tools, to handle future iterations.

This route is slower, but it spares you from learning to code from scratch. In 8 to 12 weeks, with a budget of $10,000 to $30,000 (no-code subscriptions plus freelance dev plus the $99 per year Apple Developer account), you can ship a version ready to launch.

Step 4: put the word “AI” in your App Store listing

One data point from a Sensor Tower report matters here. Cal AI crammed an average of 35 AI-related keywords into its App Store description every quarter. AI-related search terms then drove 25 percent of its search-driven downloads.

A quarter of all organic traffic came from putting AI in the title and description.

App stores index “AI” as a category tag. If your product title contains that keyword, it gets pulled into the AI search pool, which ties directly into ASO (App Store Optimization) keyword weighting. Right now, large numbers of users are searching the App Store directly for “AI food,” “AI scanner,” and “AI menu.” They are not searching for your brand. They are searching for the category. If your app name, subtitle, and description do not contain these keywords, you will not capture a drop of this traffic.

This is the cheapest one-off task you will ever do, and the payoff lasts a long time.

Step 5: price with a monthly subscription, never do one-time buyouts

$9.99 a month is the proven price anchor in this space.

Why not a one-time buyout? The value of a high-frequency use case compounds continuously. Users need it today and will need it tomorrow. A buyout model crushes your revenue into a single transaction, forcing you into an endless cycle of buying traffic. A subscription model means the user you acquire today is still paying you next year.

A crude math breakdown: 10,000 paid users times $9.99 times 12 months comes to about $1.2 million in annual revenue. But that is your ceiling, not the cash in your pocket. Deduct:

• App Store’s 30 percent cut. That leaves $840,000.
• Model API costs. At GPT-4o vision’s $0.01 to $0.03 per call, 10,000 users snapping 3 photos a day is 90,000 calls a day, or $27,000 to $80,000 a month. That leaves $760,000 to $810,000.
• Customer acquisition costs. US iOS health app eCPI runs $2 to $4. 10,000 paying users means you need 200,000 to 500,000 downloads, so a $300,000 to $2 million acquisition budget.
• The standard 8 to 25 percent refund rate for subscription apps. Take another 20 percent off.

Actual net profit might land around $300,000 to $600,000, and that is only a number you can reach at Cal AI’s scale. It is a small business that supports one or two people, not $1.2 million dropping cleanly into your pocket.

4. The real bottlenecks on this path

I am ranking these by how hard they will choke you out, not by how likely they are to happen.

Bottleneck 1: marketing is the wall, tech is not

This deserves its own section. The previous steps broke the technical barrier down so low that you will get the illusion marketing is also a breeze. That is wrong.

App Store keywords can get you some organic traffic, but to hit a million downloads, you almost certainly have to pay for user acquisition. In the US iOS health app category, CPI (cost per install) typically runs $4 to $8. Blended with organic traffic, eCPI sits around $2 to $4. A million downloads means a paid acquisition budget of at least $2 to $4 million. That is not a high schooler’s budget. That is a Series A startup raise.

The Cal AI crew poured far more firepower into marketing than into code. Building the thing is the starting line, not the finish line. That is not a platitude. That is the situation.

If you have run an e-commerce business before, here is good news. Your ability to run paid traffic on Meta and TikTok, and your keyword expertise in Amazon ASO (listing optimization), can be directly ported to App Store ASA (Apple Search Ads). That is your real entry point. Not knowing how to code does not matter.

If you have never run paid traffic, you need to master that game first before touching this path.

Bottleneck 2: the formula is necessary, not sufficient

Those three filtering criteria only explain why people are willing to pay. They do not explain why you are the one who gets to pocket that cash.

Cal AI’s moat is not “fewer features.” It is the data flywheel generated by 15 million scans. Every photo gets corrected by the model and human reviewers, making its food recognition database far more accurate than any newcomer’s. A newcomer taking 1 million scans would still not be enough. This is a data asset built over time, and no formula can hand it to you.

You will think, “I can clone Cal AI too.” The formula drastically underestimates the difficulty of going from 0 to 10,000 paying users. Three conditions can help you avoid the most obvious pitfalls, but whether you actually break out depends on the combined force of product experience, marketing, and team execution.

Bottleneck 3: the moat is shallow, but it has a deeper side

You are building a product whose features can be summed up in one sentence, which means a competitor can copy it in one sentence. The standard playbook is speed and vertical depth. Whoever becomes the default choice in a narrow scenario first eats up that keyword’s search traffic. But there are no silver bullets. Being first to 1,000 reviews, owning keyword placements, building brand search volume. These are all assets that demand time to accumulate.

The deeper moat lies in your user data. After 1 million scans, your recognition database becomes 5 percent more accurate than a newcomer’s. That 5 percent determines whether you rank number one in search results.

Bottleneck 4: fragmented word of mouth is the norm

MenuFit holds a 4.8-star rating on the App Store, but its written reviews average a dismal 2.7, flooded with complaints about subscriptions and billing nightmares. The 4.8 is the global star average. The 2.7 reflects the actual experience of users who bothered to write a review. A bad subscription and billing experience is an epidemic across this niche.

There is nothing wrong with making money via subscriptions. Building your revenue model on making it a nightmare for users to cancel, though, is a recipe for disaster. That path ends one way: a barrage of one-star reviews and getting delisted by the platforms.

5. The 30-day test sprint

If the analysis above holds, stop theorizing and run this 30-day test sprint.

Days 1 to 3. List five use cases. Pick the one that frustrates you the most, and validate it on the App Store and Reddit using the method from Step 2.

Days 4 to 7. Build a no-code prototype with Glide, Bubble, or Softr, and wire up a single photo recognition flow (even if it only recognizes one dish). Cost: $0 to $50. Decision gate: can the recognition accuracy hit 70 percent? If not, pivot.

Days 8 to 14. Find 20 target users on Reddit, Discord, or Pinterest for a free trial, and gather raw feedback on whether they will actually pay. Decision gate: if three or more of the twenty say “I would pay $9.99 a month,” push forward. Otherwise, pivot.

Days 15 to 30. Hire a freelance Swift dev to rewrite the native version, write App Store assets, and prep IAP for launch. Budget: $3,000 to $8,000. Decision gate: by day 30, are you confident you can land 100 real paying users within 60 days? If not, kill the project.

An e-commerce veteran’s instinct is to scale if ad spend ROI breaks even and kill if it does not. That rule applies to apps as well. Validate the direction in 30 days. If you cannot, pivot. Stubbornly grinding a dead end is the biggest cost in this game.

Grab that piece of paper with five use cases now.

The biggest cost for an ordinary person building a product is not money. It is waiting for a more grandiose idea, and then doing nothing. Two seventeen-year-old high schoolers did not wait. They built a photo feature, and eighteen months later, MyFitnessPal acquired them. You are not worse than they are, but you will be slower than they were, unless you start snapping your first photo today.

Pick up your phone.