Zenkai Club blog

Love the problem, not the swords

Fri, 30 Oct 2026 00:00:00 GMT

I can make up product ideas all day. That is a trap.

On September 29, 2026, I wrote myself a note, and one line stuck: "I need to love the problem and think about truly defeating him, not coming up with swords and falling in love with the swords." Swords are solutions. The enemy is the problem.

Start from the sword and you go hunting for users. Start from the problem and they are already waiting.

Why are sudden startup ideas a red flag?

A sudden idea comes from your head, not your life. If you cannot point to a problem you hit many times, and talked about before the idea showed up, you are holding a sword with no enemy.

My note admits it: "Random ideas and imaginable possibilities are easy for me to generate." With AI tools, a "what if" becomes a demo in a weekend, so I can fall for a dozen swords a month. Now a sudden idea is guilty until proven innocent.

In November 2012, Paul Graham opened "How to Get Startup Ideas" with: "The way to get startup ideas is not to try to think of startup ideas. It's to look for problems, preferably problems you have yourself." In 1995 he spent 6 months putting art galleries online before noticing galleries did not want to be online. He called the bad kind "made-up" or "sitcom" ideas: friends could maybe see using one, and "Sum that reaction across the entire population, and you have zero users."

Clayton Christensen's milkshake story, written up by Harvard Business School Working Knowledge in February 2011, shows the same thing. A chain asked "milkshake drinkers" for a better shake and sold no more. A researcher who just watched for a day saw that 40 percent of shakes were bought early in the morning by commuters with a long drive. The job caused the purchase, not the profile.

Rob Fitzpatrick's The Mom Test (2013) adds three rules: "Talk about their life instead of your idea," "Ask about specifics in the past instead of generics or opinions about the future," and "Talk less and listen more."

Aha: The Mom Test works on you too. The best liar I know is me with a new sword. "I could totally see people using this" is the future-tense fluff the book says to throw away. So I ask myself: when did this really happen to me, and what did I do?

How often do startups fail from building the wrong thing?

In March 2026, CB Insights looked at 431 venture backed companies shut down since 2023. "Ran out of capital" led the list at 70 percent, but they call that "almost always the final cause of death, not the root problem." Of the 385 with known reasons, 43 percent cited poor product market fit. Skipping the problem check just moves the waste to the end, where it costs the most.

I have done it myself. I pictured a small legal consultant using my tool, and I know none. Pitching an idea to friends felt like proof. It was not, so I wrote: "Repeatedly discussing the same solution does not clear the flag."

Ten pitches leave the flag up. Ten run-ins with the pain clear it.

The sentence I aim for now: "Even for me, knowing the tools and how to combine them, this is still a real pain to do and keep working." If it hurts someone who can write the glue code, it is worse for everyone else.

Aha: being good at building makes this check more important, not less. Cheap swords mean more swords. When AI writes an app in an afternoon, picking the enemy is the hard part. My notes say it in three words: "Direction before speed."

How do I check an idea before building it?

Use a plain note, with no AI open.

Write the problem before the idea. Fill in "When I ___, I have to ___, and it costs me ___." If only the solution comes out, stop.
Count real times, in the past. Check your notes and messages from the last 3 months. Count each time the problem hit you, and each time you raised it before the idea.
Name real people. List people you have seen hit it and could message today. Empty means sitcom land.
Write what they do now and where it breaks. Note where it fails and what that costs.
Ask if you are in the group. If you are only visiting, go watch them first.

You come before them, and the past comes before the future.

To run it here, type your idea, then answer six questions: times it hit you in 3 months, times you raised it before, others you saw struggle, what they do now, what that costs, and whether you are in the group.

When unsure, round down.

A made-up example: an indie hacker wants to ship a tool "small accountants could use" by Sunday, scores zero, and asks two accountants about month end instead.

The clearest pass in my own work is the Earshot Teleprompter, an iPhone app that helps you write a script, read it while looking at the lens, and edits the take for you. It is mine, so weigh that. The pain came first: talking to the lens and cutting the take hurt every time I recorded. I met the enemy before the sword.

Love the problem. Keep the sword down until you have met it.

Sources

To make AI video look real, make it worse

Fri, 30 Oct 2026 00:00:00 GMT

Our best night shot was turned down in a single line: "still super AI, too high quality, face not dark/blurry enough." We had given it a lamp in view, hard falloff, two colours of light and named grain. On paper it beat every earlier try. On screen it looked more fake, because a phone filming a dark room at night does not make a lovely picture. It makes mush.

More craft in the prompt made it worse. A real night clip is defined by what it lacks.

That note holds the main lesson from building realistic AI creators for brands. Realism comes from naming what a cheap camera really puts on screen, then letting the picture get worse. And a realistic AI creator is a craft tool, never a mask. Where a platform or a law asks for a label, you add one.

Why does AI video look fake even when it looks good?

Because it is too clean. Models learn to please, so by default they hand you even light, crisp focus, smooth skin and perfect exposure. Real phone footage is noisy in the shadows, its focus hunts, compression leaves blocks, and the face is a bit soft. Polish is the giveaway.

We read 81 transcripts from short videos about realistic AI clones from September and October 2026. The top tip: never ask for "realistic." Write the flaws in, like pores, oil on the nose and a camera hunting for focus.

This has a name. In 1970 Masahiro Mori described the uncanny valley: we like a robot more as it looks more human, until it gets very close and our liking falls off a cliff. His example was a prosthetic hand that passes at a glance, until "we could be startled during a handshake by its limp boneless grip together with its texture and coldness."

A rough sketch after Mori's essay. Most AI creators sit in the dip.

Aha: Mori's cold hand and our rejected night shot are the same failure.

Our frame passed every check we ran on purpose and flunked the one viewers run without thinking: did this come off a phone? So the fix was never in the prompt. In our pipeline a separate degrade step runs once the model is done. For video, our default preset shrinks the clip to 540 pixels tall, adds noise that changes every frame and starves the bitrate.

How do you write a realistic AI creator prompt?

Swap feelings for things a model can draw and check. One character took 8 rounds and 24 images, and for four of those rounds the angle words missed. Naming what that angle shows (nostrils, the underside of the jaw, a strip of ceiling) landed on the next round. OpenAI's Sora 2 guide says the same: "Verbs and nouns that point to visible results will always give you a clearer, more consistent output."

The angle only worked once we listed what it puts in the frame.

Here is the short version of our flow for one phone selfie shot:

Start from a real phone photo. Borrow the scene, light and camera feel. If it shows a real person, never borrow the face: cast a new, distinct person.
List the place in plain nouns. A kettle and a dish rack do more than any mood word.
Set framing with numbers. Ours: "face cheekbone-to-cheekbone spans 60% of frame width, eyes at 40% of frame height, chin at 82%." Words drift between rounds. Numbers hold.
Set the angle by what it shows, then fence it. Add a NOT after a visible target, like "NOT tipped far enough to show more ceiling than a thin strip." A bare NOT just plants the thing you banned.
Check, degrade, label. Tick off your own list, make the clip worse in a separate step, and switch on the AI label at upload.

To test your own wording, paste a prompt into the box below and it flags vague camera and lighting words, with a visible swap for each, plus framing numbers you can copy.

Aha: realism is mostly borrowed, not written. The night shot needed a phone's damage. The angle needed what a phone at chest height really sees. The most believable parts were taken from the real world.

I build Earshot, which makes realistic AI creators for brands, so weigh this with my bias in mind. One of our misses: a main reference caught mid-word, so every later clip kept talking with no sound. Every reference now has a closed mouth.

One place this pays off: a sleep app needing "2am, can't sleep" clips prompts one dim bedside lamp, then runs the heaviest degrade, and the shots finally look like a phone in a dark bedroom.

Do you have to label AI video?

Yes, in most places that matter. YouTube says: "We require creators to disclose AI-generated or altered content that seems realistic." TikTok said in May 2024 it had "required creators to label realistic AIGC for over a year." On Instagram, "you must manually label photorealistic video or realistic-sounding audio" made or changed with AI. In the EU, Article 50 of the AI Act, in force from 2 August 2026, says those who deploy AI for deepfakes "shall disclose that the content has been artificially generated or manipulated."

YouTube's rule for realistic AI content. Source: YouTube Help.

In the clearest backlash we saw, people were angry about a hidden AI clone, not about using one. Realism and honesty are separate goals, and you need both: name what a real camera sees, make the picture worse on purpose, and put the label on.

Sources

Our best customers wanted one video, not a daily habit

Fri, 30 Oct 2026 00:00:00 GMT

The users most likely to pay us looked the most likely to leave. It took a pile of charts to see why.

In October 2026 we ran ads for our teleprompter app and read every script new people pasted in. The serious users were small business owners, not would-be influencers: real-estate agents, a hair-care line, a sales recruiter, a small agency. Each brought a finished script of 100 to 500 words, often written with ChatGPT. They recorded it and left.

Going further on day one went with coming back less. The people who reached the real finish line never returned.

People who only got a script ready came back on a later day 44% of the time. People who recorded a take came back 15% of the time. The six who saved their raw take came back 0% of the time. It looks like finishing the job makes people quit. I almost believed it.

Why does daily retention lie for low-frequency products?

The daily view said we were leaking. From October 12 to 23, 174 new people arrived, 41 recorded a take, and only 24% came back on any later day. Of 134 people from October 12 to 20, just 10 recorded again later: 7%.

The scripts told another story. These buyers had one listing, one product drop, one open role. They needed a new video when life handed them something new to promote, which happens monthly or a few times a year.

We were grading a monthly need with a daily ruler.

That ruler came from elsewhere. Andrew Chen writes that "DAU/MAU has been a popular metric because of Facebook, which popularized the metric," and he passes on the rule of thumb that apps over 20% "are said to be good, and 50%+ is world class." That benchmark grew up in a product people opened daily.

In 2016, Archana Madhavan at Amplitude wrote that "The product usage interval is the frequency (daily, weekly, monthly, etc.) with which you expect people to naturally use your product." Sequoia's data science team says the same: "Depending on the vertical of your product, a daily, weekly or monthly time frame may be most meaningful."

Brian Balfour's 2013 test for product market fit: "IF it flattens off at some point, you have probably found product market fit for some market or audience." Put that next to Madhavan's idea and you get this.

A curve can only flatten on the clock the need runs on.

Track a monthly need day by day and almost nobody has the need on any single day, so the curve slides toward zero and you decide there is no fit. Group the same people in 30-day buckets and it may hold steady. A daily dashboard can talk you into killing a product that works.

How do you measure retention on a customer's natural rhythm?

Name the trigger. One sentence on what makes the customer need you again. Ours: "a new listing, a new product, or a new offer to promote."
Put a number on the rhythm. Guess how often the trigger happens, then check it against what users say or do.
Count "within" a window, not "on day N." Someone back on day 6 is gone "on day 7" but retained "within 7 days." Pull returns within 1, 7, 30 and 90 days.
Choose the window before you look. For a monthly rhythm, the 30-day number is your retention. Deciding first stops you picking the flattering one.
Shrink the denominator. Where you can, count only people who had another need inside the window. Then you ask: when they needed this again, did they pick us?

The same pretend cohort gets four verdicts. Daily, it looks dead. Monthly, almost a third stayed.

To try it, pick how often your customer naturally needs your product, then type in your cohort size and how many came back within 1, 7, 30 and 90 days.

What does the "next one" moment look like?

We had also built for the wrong clock: a trends feed, a score on every take, stage emails. Daily features do not create daily need. An agent with one listing this month needs a finished video, not fresh trends.

Nothing in our app asked about video number two. Push notifications were on for about 10 people in total. The 0% return among savers did not prove they were done with us. It proved we never asked when they would be back.

On the left the product stops when the job stops. On the right, finishing one video sets up the next.

The fix is one question at the end of the first job: what will you promote next, and when? Then show up that day with a draft ready. The window you measure is the moment you design for.

This buyer is exactly who Earshot for brands is for. It makes done-for-you creator videos for businesses that need a promo. I build it, so weigh that.

Two made-up cases show the rule cuts both ways. A founder with a video tool for agents judges a 30-day window keyed to "finished a listing video" and asks when the next listing goes live. A meditation-app founder wants a monthly window because it looks nicer, but his need really is daily, so his 1-day and 7-day numbers stay the truth.

Our best customers did want us: one video, then another weeks later. Measure the weeks. Build the moment they return.

Sources

A paywall up front got us 12x more trials and 1 sale

Fri, 30 Oct 2026 00:00:00 GMT

We put a paywall inside onboarding and got twelve times more free trials. Two weeks later we had one sale, and that buyer switched off renewal about an hour after paying.

The app is Earshot Teleprompter. You write a script, read it on your iPhone while looking at the lens, and an AI edit turns the take into a finished video. In October 2026 we tested where the paywall should sit. I build the product, so weigh my view with that in mind.

The early wall traded makers for trials.

What happened when the paywall moved into onboarding?

From October 12 to October 23 we split new iPhone users into five arms. Two arms saw no wall during onboarding. Three arms saw one inside onboarding. Nothing else changed.

The no-wall arms had 64 people. 30 of them got a script ready (47%), and together they started 1 trial. The wall arms had 83 people. 26 got a script ready (31%), and they started 12 trials. So the early wall cut script making by about a third and sent trials way up.

On a dashboard, that is an easy winner. Then the trials began to end.

Why did more trials not bring more buyers?

Most of those trials were not buying decisions. They were people unlocking the app for free and cancelling at once so they could never be charged. Since October 12 we had 16 real trials on annual plans ($59.99 for creators, $79.99 for business, with 7 free days first). 11 were cancelled, 7 of them within 10 minutes. 4 were still running. 1 was charged $79.99 and turned off renewal about an hour later.

One charge out of sixteen, and it already had renewal off.

Not one trial user saved an AI-edited video, the very thing we sell. The money did not work either. One annual sale nets about $51 to $68 after Apple's cut, and ads cost us about $17 to $35 per trial. That needs 1 in 2 to 1 in 4 trials to pay. Among trials that had ended, 1 in 7 paid.

Why do people cancel a free trial right away?

Because on the App Store, cancelling early is free insurance. For most apps, a trial keeps running after you turn off renewal. Apple's subscription page says: "Their subscription begins immediately, but they won't be billed until the offer duration ends." So a careful person cancels at once and keeps the full trial.

This is common, not just our bad luck. In a May 2026 RevenueCat post, Daphne Tideman wrote: "A staggering 55.4% of 3-day trials are cancelled on Day 0."

Every trial length spikes on the first day. Ours ran for seven.

The connection: a trial count measures a free click, not a buying decision.

Seen that way, the early wall did not make twelve times more people want to pay. It gave twelve times more people a free button that costs nothing to press or undo. Back in 2006, Fred Wilson warned: "Asking for a credit card even though you won't charge anything to it is not a good idea."

Is a hard paywall better than freemium?

Often, yes, if you count money. In an April 2026 RevenueCat post, David Barnard reported that "hard paywalls convert downloads to paid at a median of 10.7%," against 2.1% for freemium. That is downloads to paid, not trial starts. Compare that with Superwall's May 2026 study of more than 40 million onboarding paywall opens, which counts a trial start or a purchase as a conversion. A scorecard like that rewards anything that makes the free button easier to tap.

The connection: the common paywall scorecard rewards the exact number that lied to us.

How do you measure free trials honestly?

Here is the method in five steps.

Pick your value event. Name the one action that proves someone got what they came for. Ours is saving a finished AI-edited video.
Tag the instant cancels. For each trial, take the start time and the time renewal was turned off. Mark anything switched off within 60 minutes.
Judge only trials that have ended. A trial still running has not voted yet. Ours: 1 charge from 7 ended trials.
Count four things side by side. Free yes, reached the value event, paid, and still renewing after paying.
Score each arm by revenue per 1,000 visitors. This catches the arm that wins on trials and loses on dollars.

The real gap is box two: nobody reached the value.

To check your own test, type your visitors, trials started, trials cancelled in the first hour, paid conversions and price into the tool below, with an A and B column for two paywall placements.

One use case: a growth marketer buying install ads with a "start trial" goal. Cost per trial looks fine. Once he strips out instant cancels and works out cost per paying customer, the cheapest ad sets turn out to bring the most unlock and cancel trials.

What we are changing

The paywall moves off onboarding and goes after the first finished video, where it asks a harder question: was that worth paying for? We are dropping forced sign-in before purchase, since 3 of 10 people who tapped "Start trial" quit at that screen and Apple's ID is enough to buy.

Our sample is small, 147 people across five arms, so read it as a story, not a benchmark. Still, the lesson holds. What matters is not how many people take something free. It is how many pay once they have seen what you made.

Sources

Rank videos by spike, not views

Fri, 30 Oct 2026 00:00:00 GMT

On October 7 we nearly scheduled a batch of smart-glasses videos with five hashtags picked because each matched a word in our script. Then came a plain question: which tags did the winning smart-glasses videos actually share? We pulled 49 videos that had clearly beaten their creators' usual numbers and counted. Two of our five tags appeared on none of them.

That is the quiet danger of research. It can hand back exactly what you already believed. The fix, here and in four other traps, is to stop trusting the easiest number and ask what it is compared to.

Same two videos, opposite order. Views reward the big audience. Spike rewards the video.

Why are raw views a bad way to find what's working?

Raw views mostly tell you how many people already follow a creator. A huge account gets a floor of views on nearly anything, so its average post looks like a hit. Spike, views divided by followers, takes that floor away. A 40,000 follower account with 900,000 views sits at 22.5x. A 5 million follower account with 1 million views sits at 0.2x. The first traveled far past its own crowd. The second was an ordinary day.

TikTok said something close to this in 2020. A video may get more views from a bigger account, but "neither follower count nor whether the account has had previous high-performing videos are direct factors in the recommendation system."

TikTok's 2020 explainer. Its last paragraph is why spike works.

Aha: followers are a head start, not a verdict. Divide by followers and you remove the starting line and keep the race. A list sorted by spike shows which videos strangers kept watching, and that is the part a newer creator can copy.

The platforms walked this road first. YouTube added watch time to its recommendations in 2012, and Cristos Goodrow, a VP of Engineering there, later wrote that the change cut views by about 20%. YouTube picked a smaller number that meant more. Tools like vidIQ now score a video against the average of other videos on the same channel. Spike is the rough short-form cousin, built from the one number every scraper gives you free: followers.

One honest catch. A 500 follower account can hit 100x by luck, so keep a follower floor by default and lower it on purpose to look for small-account explosions. And be plain about limits: no scraper shows real watch time or click-through rate. Spike is the best public clue that hidden watch signals were good, not a copy of them.

How many comments do you need to read before you claim buying intent?

Hundreds, not twenty. Buyers show up in comments as "where do I get this?" or "how much?" But page one holds about 20 comments, and a video that spiked often has 600 or more.

Twenty out of six hundred. Most days page one hides every buyer.

If 1 in 100 comments is a buy question, all 20 on page one miss about 82% of the time. You write "0.00% intent," and across a dozen niches every niche looks dead. Doctors solved this in 1983. Hanley and Lippman-Hand's rule of three says that zero events in N tries puts the true rate somewhere below 3 divided by N.

Aha: zero out of twenty is not zero. By that rule, zero in 20 allows a rate up to 15%. That is not a dead niche. It is no answer at all. To say a niche sits under 1%, you need about 300 comments with no buyers.

Then read the quoted questions yourself. In travel and relocation niches, "how much?" often means rent, which is not product intent. A word filter counts it anyway.

How do you run a spike research pass?

A short version of the routine we run before anything gets scheduled:

Search the words real videos rank for, like "how to stop saying um," not a category name.
Cut by date and topic. One run this month went from 58 queries and 6,744 results to 2,520 recent posts and 1,536 on topic.
Compute spike for every row and roll up by creator. Two or more hits from one person suggests a format, not luck.
Read the shortlist in full and page the comments. That run picked 87 videos and read 81 transcripts end to end.
Count hashtags across the spike-ranked winners, then judge frames, not covers. Covers are one picture the creator chose, and their links often expire within hours.

One real run. Most of the work is throwing rows away for good reasons.

To try steps 3 and 4, paste one video per line as name, views, followers in the first box, then enter your comment counts in the second.

Two people this fits. A skincare founder with budget for one account compares "barrier repair" with "acne scar journey." Views favor scar journeys, but spike favors barrier repair, and its comments ask which product it is while the scar comments mostly say "this is me." She picks barrier repair. A small agency counts tags across 50 spike-ranked winners and finds four of a client's ten never show up once, so they swap them.

I build Earshot, which turns spiking hooks in a brand's niche into videos from realistic AI creators, so I need spike to be true. Every rule above came from a number I believed that turned out hollow. Our five hashtags felt like research and were really a mirror. Next time a number makes you feel sure, divide it by something first.

Sources

Useful tools that actually work

Fri, 30 Oct 2026 00:00:00 GMT

Most software promises a lot and asks for a lot. We try to do the opposite: small tools that do one useful thing well, and keep doing it.

What's in the lab today

Constella is a memory for your whole team. It pulls in notes, email, Slack and meetings, and lays them out on a smart canvas you can ask anything.
Earshot gives your brand realistic AI creators. It finds the hooks going viral in your niche, pairs them with your product demo, and makes the videos for you.
Earshot Teleprompter lets you read your script while you look straight into the lens, then coaches each take and turns it into a finished video.
Fullbloom helps you calm down fast when big feelings hit. It paces your breathing with a vibration you feel in your hand, then walks you through the DBT skill that fits.
Neko is a personal assistant inside LINE. You text it errands, and it asks before anything goes out.
Smart Glasses put your second brain in your line of sight, on Even Realities G2.

Not sure which one is for you

Tap "Recommend me a product" on the home page. It asks where you want to go next in life and what you need to get there. Only if something in the lab can truly help will it suggest it.

Have an idea for a tool?

Tell us the problem you wish someone would fix. If your idea becomes a successful tool, we'll give you up to $100k. You can share it here.

174 people came from our ads. Zero left with the video we sell.

Fri, 30 Oct 2026 00:00:00 GMT

For twelve days, our app looked busy. Then I counted the one result it exists to deliver, and it was zero.

The app is Earshot Teleprompter. You read a script on your iPhone while looking at the lens, and an AI edit turns the raw take into a finished video with cuts and captions. That finished video is what our ads sell and what the paid plan unlocks. From October 12 to October 23, 2026, 174 new people opened the app, mostly from our first paid Meta ads. None of them saved or posted an AI-edited video.

Five bars of motion, then one bar of value, and that last one is empty.

I build this app, so weigh my view with that in mind.

Why does a busy funnel hide an empty result?

Because most funnel steps count effort, not reward. In those two weeks, 114 people got through onboarding, 59 had a script ready and 41 recorded at least one take. That reads like a small product with a leaky top.

But writing a script is work. Recording yourself is work too, and scary work for many people. None of it is the moment someone watches a clean, captioned video of themselves and feels like a pro. That moment is why anyone pays or comes back.

Our cliff sat exactly where the product should shine. Of the 41 people who recorded, only 13 started an AI edit, and only 5 ever saw a finished one in the app. The edit ran on our server for 5 to 12 minutes. Most people were gone within a minute.

People gave us about a minute. The reward needed five to twelve.

The server logs made it sting more. There, 14 people started an edit, 11 got a finished video, 6 ever opened it, and 0 saved it. Five finished videos sat ready and unwatched. Each person who recorded a take cost us about $9 in ads, and most walked straight into that wait.

Can going deeper into a funnel make people leave?

Yes, when the step where they stop is a dead end. We took the 134 people who arrived from October 12 to 20, sorted them by the furthest step they reached on day one, and checked who came back on a later day. People with a script ready returned 44% of the time. People who recorded a take returned only 15%. The few who saved a raw take returned 0%.

Small groups, so trust the shape: return rises, then falls hard after recording.

Aha: going further into a funnel can make people less likely to return if the step they stop on is a dead end.

Recording without seeing the edit is the worst spot to quit. You did the awkward part and got back a plain phone clip of your own face. A deeper step with no payoff is not progress. It is a sharper letdown.

Growth teams spotted this long ago. Dave McClure's pirate metrics defined activation as users who "enjoy 1st visit: "happy" experience." Facebook's growth team chased 7 friends in 10 days. Slack saw that once a team sent about 2,000 messages, 93% of those teams were still customers.

Aha: McClure's activation, Facebook's 7 friends and Slack's 2,000 messages are the same thing measured in three different products.

Each one counts people who reached the result. Our "recorded a take" counted effort just before the payoff, where a number looks best and means least. Appcues calls the missing piece time to value: "the gap between signup and the moment users experience your product's core benefit."

How do you measure the promised outcome?

Here is the check in five steps.

Write the promise in one sentence. Use your ad's words. Ours became: the user saves or posts an edited video. Not signed up, not recorded.
Make that moment a tracked event. For us that meant edit_saved and edit_posted.
Count people, once each, at their furthest step. Drop test accounts, and join anyone your analytics split in two at sign-in.
Divide by starters, not by the step before. Ours, against starters: 0 / 174 = 0%.
Time the gap, then rule out bugs. Compare minutes to value with how long leavers stayed. Then check error logs for the same dates. We found outages and a sideways camera.

Type each step of your funnel and how many people reached it into the tool below, then mark which step is your moment of value.

If it warns that fewer than 10% of starters reach that step, hold off on buying traffic.

Who else falls into this trap?

Any product whose reward arrives after effort or a wait. Take a solo founder with an AI headshot app. Uploads look great, but the value is downloading a headshot someone actually uses. If training takes twenty minutes, most people never see one. The fix is a rough preview in the first minute, then counting downloads, not uploads.

What we are changing

Every person who records should see a finished edit. The edit will start on its own after the first take, the result will appear inside the app, and we will ask to send a notification right when people are waiting. The paywall moves to after the first finished video, so people pay for something they have seen.

A bad attempt is not a bad idea. What broke was the road from tap to reward. Next time we run the numbers, I will count the last row first.