You're shipping more with AI and at least half of it can feel like slop. That feeling is correct, and it's also the reason your odds of a hit just went up.
You ship more now. A lot more.
Posts, features, variants, landing pages, whole products that used to take a season now take a weekend. And somewhere inside that speed a quiet worry showed up: the average dropped. Your feed looks like everyone else's feed. The volume is real, and so is the nagging sense that you're just adding to the noise.
So you reach for the old debate. Quantity vs quality. Pick a side. Make fewer things and make them great, or make many things and swallow the fact that most will be mediocre.
That debate is the wrong frame. Not because quality stopped mattering. Because finding a hit was never a quality problem to begin with.
(If you want to know how your specific situation stands, use the quick, interactive, free test at the end of this article.)
Hits were always a search problem
Years ago, well before any of this, I co-founded and co-ran a daily-deals business.
Every morning we shipped about five offers, one each for five different partners, a fresh set of partners the next day. We never knew which one would pop (although we had our educated guesses). A few were hits we'd ride hard. Plenty did fine. Some died on arrival. The volume found the winner, because we never could.
When an offer hit, the search didn't stop. It moved. For that one offer we'd run around ten Facebook ad campaigns, and usually one of the ten worked. The moment it did, we poured the budget into that single ad and pushed it as hard as it would go, because the offer was only live for a short time.
That whole business was a search engine with people inside it. Two layers, product and distribution, each running the same loop: make many different attempts, watch for the one that catches, drop the rest, bet everything on the survivor. No AI. Just a lot of shots and a short clock.
I didn't have taste good enough to pick the daily winner in advance (although it was good enough over time to make the odds much better). Nobody does. The whole apparatus existed because the hit is unknowable until the market touches it.
Why worse on average can still mean more hits
A hit is an outlier, and you don't craft an outlier into existence. You search for it. More attempts, faster feedback, more chances to match an idea with the right audience: that's what raises the odds. It's the portfolio logic VCs call shots on goal: most attempts miss, a few carry everything, and you can't tell which in advance. Quality per attempt can fall while your chance of landing one climbs, simply because you're taking far more swings.
The arithmetic is blunt. Say one in ten honest attempts lands. Ten attempts get you one winner, a hundred get you ten, even if the average attempt got worse along the way. Lowering the average doesn't cost you the hit. Skipping the attempts does.
An old story from a ceramics class catches this perfectly.
A teacher split his students in two. One half would be graded on quantity alone, by the literal weight of the pots they made. The other half on quality, a single flawless pot for an A. When grading came, the best pots in the room all came from the quantity side. While the quality group sat theorizing about the perfect pot, the quantity group made pot after pot and learned something from each one.¹
The group told to chase quality lost on quality. The group that just kept shipping found it, because volume is how you learn what good even feels like.
The same pattern is showing up at full scale. When large language models reached publishing, the number of new books roughly tripled in three years. The typical new book got worse, measured by how much readers actually engaged with it. And yet readers came out ahead: with three times as many books to choose from, more of them found one that genuinely fit, and the value readers got over and above what they paid rose about 7 percent in a single year. The authors who came before weren't pushed out.²
Triple the books. A lower average. More readers finding their match. That last part is the engine: more attempts means more chances for an idea and an audience to meet.
Volume only counts if you're learning
Most people get this next part wrong.
In one experiment, an indie hacker gave seven AI agents a hundred dollars each and told them to go build a business. One agent, running on Gemini, pumped out 104 near-identical blog posts and kept right on going. Its progress log read like a stuck record: another post, then another, then another. Traffic and revenue from all of it: essentially zero.³
That isn't a search. That's a photocopier with a thesaurus.
A search needs two things raw volume can't fake. The attempts have to genuinely differ, a new idea, angle, audience, or hook each time, or you've only reproduced the same miss in higher resolution. And you have to read what comes back. Speed only helps when speed buys learning.
The ceramics students didn't win by making clay faster. They won because each pot told them something, and they listened.
There's also a floor. Push quality too low and you don't widen your odds, you torch them. Below a certain bar the market won't touch what you make, and every bad attempt spends trust you don't get back.
Quantity and quality aren't enemies pulling in opposite directions. They sit in a balance. The faster you go, the more it matters that you can tell a promising rough thing from a worthless one, and that telling is a kind of taste.
Taste is the one part of this you can't hand to the machine doing the spraying. Going wider asks for more judgment, not less.
And taste is the one thing in this loop that compounds. Read the returns honestly and your eye sharpens, so the next batch starts from a higher base rate. That's why searching wide isn't gambling: the picker gets better every time it picks. The machine sprays at the same flat odds forever. You don't.
This is where my short clock earned its keep, though I didn't see it at the time. A limited offer window forced a verdict fast. The market answered before I could fall in love with my own idea. (Full disclosure: we had half a million subscribers by then, so the verdict came fast partly because there was a crowd standing by to give it.) The speed was never the prize. The forced feedback was.
That clock was a gift the business handed me for free. Building alone now, nothing expires. An attempt can sit there half-alive for months, never quite failing and never quite landing, while I tell myself the search is still running.
But a search with no deadline isn't a search. Nothing forces the market to answer, so the answer never comes, and I mistake the open loop for an open mind.
So here's the part nobody mentions: if your attempts don't come with a clock, you build one. Pick the date the verdict lands, and decide ahead of time what a yes looks like: someone who comes back without being prompted, uses the thing, pays, or sends someone else, not a spike of likes that never turns into a second visit. Otherwise "spray to find signal" slides into spray forever, and that isn't searching. It only feels like it.
The search runs on distribution too
Most of the AI conversation stops at production. Make more things, faster. But the daily-deals business spent as much of its search on distribution as on the offers, and that's the half people forget. The seven-agent experiment hit the same wall: building is maybe ten percent of the problem now, and getting a stranger to care is the other ninety.
AI can be brutally good at the first part of that distribution work. Ten versions of a hook, twenty thumbnails, one demo cut five ways, translated into six languages, aimed at four audiences, all before lunch. The raw labor of trying many angles to the market, the thing that used to gate small teams, has mostly collapsed.
What AI still can't do is manufacture the things that actually convert. It can't make someone trust you. It can't conjure attention out of nothing. It can't fix bad timing, and it can't supply the taste that separates a hook that lands from one that makes people wince. It floods the top of the funnel and goes quiet exactly where judgment lives.
So the distribution move is the same move as everywhere else. Let AI generate the ten hooks, then spend your judgment on the two things it can't answer: which one a real person would actually stop for, and which channel they're already standing in. The spray is free. The aim isn't.
When spraying is the wrong move
All of this rests on one condition: attempts have to be fast. Take that away and the whole approach falls apart.
Some things can't be sprayed. When a single attempt takes months and real money before it gives you any signal, you don't get many swings, you get a handful, and the game becomes getting them right instead of running them wide.
You can't train a foundation model from scratch fifty times to see which one sticks. You usually can't ship five versions of a physical device this quarter and keep the one people liked. A medical or fintech product that has to clear regulators before anyone can legally touch it, an enterprise platform with a year-long integration before the first real verdict: these are slow, expensive swings.
Big upfront cost, a long road before any feedback, or a public failure that's fatal the first time, and you're in a different discipline, the one where you plan, simulate, and finish before you ever ship.
So this is no universal law. It's the right move for a specific situation, the one most indie builders are actually in. If your attempts can be made cheap and fast, and most of yours probably can, searching wide is how you reach quality fastest.
Spray to find signal, then slow down to finish
For that situation, the move runs at two speeds.
Phase one is fast and wide. Fast attempts, real variance, a short clock, and one job only: find signal. Most of what you make here will be mediocre, and that's fine. Mediocre is the cost of searching, and sometimes mediocre is just a hit you haven't finished yet.
Phase two starts the instant something catches. Stop spraying. Slow all the way down and turn that promising, half-decent thing into something genuinely good. This is the phase the volume crowd never reaches and the craft crowd starts with by mistake. A hit is usually a rough thing someone caught early and then refused to leave alone.
And something quietly got expensive. Making things is cheap now, so making things stopped being the bottleneck.
The hard part is deciding what's worth trying, reading the early signal honestly, knowing which feedback is real and which is vanity, and having the taste to tell a true signal from a flattering number.
Production got commodified. Judgment didn't. That's the same shift I've called the new bottleneck, arriving through a different door.
Quantity vs quality was always a false race (in most cases). The real shape is a search. Spray to find the signal, slow down to finish the winner. AI didn't change that logic. It made the spraying almost free, which hands the whole game to the one thing it can't do for you.
Let the machine do the spraying. Do the seeing yourself.
Want to find out what your actual specific situation is?
Answer a few quick questions to get a clear analysis and personalized next steps (with more than 68,000 possible results, including one that matches your case), all here, no gate, no signup, completely free.
Rabbit hole
If this clicked, a few neighbors are worth wandering into:
- When production gets cheap and judgment gets scarce, where exactly does the constraint land? That's The New Bottleneck.
- Phase two, the slow finish that turns a rough survivor into a hit, has its own strange logic: If You Did It Right, Nobody Will Ever Know.
- If hits come from searching across many attempts, the value moves to the idea you're testing rather than the build. More on that in The Idea Is the Product Now.
Footnotes:
- The ceramics-class story comes from David Bayles and Ted Orland's Art & Fear (2001); its real-world seed was a gambit the photographer Jerry Uelsmann used on his beginning students. The quantity group outproduced and outclassed the quality group because making pot after pot let them learn from each attempt, which is the whole mechanism in miniature.
- Reimers and Waldfogel, "AI and the Quantity and Quality of Creative Products," NBER working paper w34777 (2026). Large language models roughly tripled new book releases from 2022 to 2025 and average quality by reader usage fell, while the value readers captured beyond what they paid (what economists call consumer surplus) still rose about 7 percent in 2025, with no displacement of pre-AI authors. It's the cleanest evidence so far that the average can drop while total value rises.
- From the Indie Hackers experiment "I gave 7 AI agents $100 each to build a startup." The Gemini agent pumped out blog posts by the hundred, near-identical and unread, with no traffic or revenue to show for it, while the series' recurring lesson was that distribution, not building, is now the hard ninety percent.