In July, George Hotz published a post called I love LLMs, I hate hype. Most of it is about two things that irritate him: the constant doom talk about windows closing and people falling behind, and the strawman leap from "it is a fancy autocomplete" to "it will own the whole light cone". Read it, it is short and it is funny.
But there is a question near the end, thrown in almost as an afterthought inside a parenthesis, and it has been bothering me for two months:
where's all this new magical software that the productivity improvements should imply?
His answer, in the same sentence, is blunt. All the vibe coded stuff is still slop.
Now, I agree with half of that. The other half is what I want to argue with, because I think he is asking exactly the right question and looking for the answer in exactly the wrong place. At Licheo we spend our days reading other people's websites the way a machine reads them, so we see the missing half of his answer every week.
How much software is actually being built with AI?
More software is being built than at any point in history, and the volume is not in dispute. In June 2026, Lovable told TechCrunch that more than 50 million projects had been created on its platform, and that usage had reached one million new projects a week. Those are the company's own numbers, reported without independent verification, so treat them as a claim rather than an audited fact. But even if you cut them in half out of suspicion, and even if you remember that Lovable is one tool among Replit and Bolt and v0 and Cursor and a dozen others, the picture stays the same.
Millions of things exist now that did not exist two years ago. Somebody sat down on a Tuesday evening, described a tool for tracking which of their houseplants they had watered, and by Wednesday it was live on a real address with a real database behind it.
So where is it all? Why has the internet not visibly changed?
Here is my answer, and it is less satisfying than "it is all slop", which is perhaps why nobody says it. To find it you have to stop looking at what was built and start looking at what happens to a thing after it is built.
Why does nobody find the things people build with AI?
Nobody finds them because building a thing and being found were always two separate jobs, and AI made only the first one cheap. Ahrefs published a study in December 2023, by Tim Soulo, looking at 14 billion pages in their index. The finding was that 96.55 percent of those pages get zero traffic from Google. Another 1.94 percent get between one and ten visits a month.
Think about what that means for a moment. That study was done in a world where making a website still required effort. Somebody had to buy a domain, hire a developer or fight with WordPress for a weekend, write the copy, pick the photos. Even with all of that friction acting as a filter, 96.55 percent of the output was invisible.
Now remove the friction. Make the building step cost one afternoon and eleven dollars. What do you think happens to the share of things nobody ever sees?
The truth is that the bottleneck is attention, and attention is the one resource that no model has managed to generate more of. There is the same amount of human curiosity in the world this year as last year, divided now among many more things asking for it.
We point the Get Found Check at other people's sites all day, and the pattern repeats until it becomes boring. The product works. The product is good. And the file the server sends out contains no sentence anywhere that says what the product is. The owner has been waiting for months for an audience that was structurally incapable of arriving.
Did coding agents change discovery at all?
Coding agents changed nothing about discovery, for the same reason compilers changed nothing about discovery. Hotz makes a comparison in the same post that I keep coming back to. He brings up a remark he attributes to Linus Torvalds, that coding agents make programming ten times more productive while compilers made it a thousand times more productive. Hotz himself says both numbers are extreme, and he is right to say so, but the shape of the argument is sound.
Compilers were an enormous gift. They turned a month of assembly into an afternoon of C. And yet nobody, in the entire history of computing, has ever been discovered because they used a good compiler. The compiler solved the part of the problem that was inside the machine. Everything outside the machine, the question of whether any other human being would ever learn that your program existed, stayed exactly as hard as it had always been.
Coding agents are the same gift with the same limit. They compress the inside job and leave the outside job untouched. If anything they make the outside job harder, because everyone else got the same gift on the same day.
How do people find a new product in 2026?
There are two doors into a new product in 2026, and both of them want the same thing underneath. One is that they search for something and your page appears in the results. The other is that they describe their problem to ChatGPT or Gemini or Perplexity, and the assistant names your product in its reply.
The second door is the interesting one, and it opens faster than the first. Licheo wrote up a case in August about a small app built by describing it to an AI, live since 27 July 2026, with no backlinks and no Google Business Profile, which after 30 days of focused work was being recommended by ChatGPT in an incognito window while Google still had it nowhere near page one. One case proves nothing as a law, of course. But it does suggest that a new project does not have to wait years to be recommended, which is the assumption most builders are quietly operating under when they decide the effort is hopeless.
What both doors want, underneath, is the same boring thing. They want text they can read, which says what the product is, who it is for, and what problem it removes. That is it. That is the whole secret, and it is so unglamorous that people building beautiful interfaces refuse to believe it.
Why is my AI built app invisible from the first minute?
An AI built app is usually invisible because of four decisions that nobody consciously made. Let me describe the pattern, because it repeats with a consistency that stopped surprising me a long time ago.
The app is a single page. The interface is drawn by JavaScript after the page arrives, so the file the server actually sends is an empty container and a script tag. A human with a browser sees a polished product. A machine reading the file sees nothing whatsoever. The title of the page is the product name, which is an invented word nobody on earth has typed into a search box, and there is no second page anywhere that explains, in ordinary language, the problem the thing solves. Nowhere outside the site does the name appear at all, so there is no forum thread, no directory listing, no comment on a discussion, nothing an assistant can point at and say: other people have heard of this.
How do I fix an app nobody can find?
You fix an app nobody can find by making the page describe itself in words a machine receives. Every one of those four failures is a decision, even when nobody made it consciously, and every one of them is reversible in an afternoon.
You write the pages so that the words arrive in the file rather than being painted on afterwards. You put in the title what the thing does rather than only what it is called. You write one honest page that describes the problem in the words a frustrated person would actually use when complaining about it, because that is the sentence they will type. You answer the eight or ten questions somebody would ask before trusting you, in full sentences, on the page, so that an assistant has something it can lift and attribute. And you get the name mentioned in one or two places you do not control.
All of it is the digital equivalent of putting a sign above the door and telling a few people where the shop is. What is strange about this era is that we have automated the construction of the shop and left the sign entirely to chance.
How do I check whether my app can be found?
Three checks will tell you, and together they take ten minutes without any tool at all. If you have built something this year and you suspect it is sitting in the dark, start here.
Check one: can a machine read what your product does?
Open your own site, right click, and choose "view page source". Use view source rather than the inspector. The inspector shows the finished page after the browser has done its work for you, while view source shows the file a machine receives. Now search that text for the sentence that explains what your product does. If it is not in there, then as far as every crawler and every assistant is concerned, your product does not describe itself. We wrote a longer walkthrough of this test in How to Tell If ChatGPT Can Actually Read Your App.
Check two: does ChatGPT name you when somebody describes your problem?
Open a fresh window where you are not signed in to anything, go to ChatGPT, and describe your product without naming it. Say the problem, the way a customer would say it: "I need something that reminds me which plants I have watered". Read the answer. If your product is absent, you now know exactly which competitors the machine trusts and, more usefully, you can go and look at what their pages contain that yours does not.
Check three: does your product exist anywhere you do not own?
Search your product name in quotation marks and see how many results exist that you did not publish yourself. Zero is a common answer and it is the most fixable of the three. One decent thread on a forum where people are complaining about the problem you solved, one directory entry, one comment from somebody who actually used it, and you have moved from a closed system to something the rest of the internet can verify.
The results of those three checks explain most of the silence people blame on the algorithm.
Is AI built software mostly slop?
Hotz is half right, and it is the half that stings. I do not want to be too comfortable here, because there is a version of his argument that survives everything I have written.
If somebody cannot be bothered to write four honest paragraphs about what they made, is the thing worth finding?
Building used to be the filter. It was slow and painful and it separated people who wanted the thing to exist badly enough to finish it from people who were entertaining themselves on a Sunday. That filter is gone now, and something has to take its place. Judging by what I see, the new filter is the willingness to do the unexciting work after the fun part is over: to name the thing clearly, describe it in plain words, publish the description, and tell somebody.
So Hotz asks where all the magical new software is. I would answer that a good portion of it is exactly where it was deployed, on a live address, functioning correctly, waiting, and that it will wait there indefinitely because its creator believed the hard part was finished on the day it worked.
The hard part starts on the day it works. That was true before any of this, and it is the one thing the models have not touched.
If you want the ten minute version of those three checks done for you, put your address into the Get Found Check. It reads the page the way a machine reads it, tells you what an assistant would find, and shows you who is being named instead of you.