The bottleneck moved twice, and the third position has no tooling

Ask a room of founders what changed since 2023 and you get a version of the same sentence: building got easy, so distribution is the hard part now. The first clause has exactly one randomized measurement behind it, and it points the wrong way. When METR ran 16 experienced open-source developers through 246 real tasks on repositories they already knew, the ones using AI tools took 19% longer. They had predicted a 24% speedup. Afterward, having been slower, they still believed they had been sped up by 20%.

That is not proof that AI makes people slower. METR themselves now think the method was biased: in the 2026 follow-up, developers refused to work without AI even at $50 an hour, and 30–50% of them deliberately withheld the tasks they expected AI to accelerate. The honest statement is narrower than either side of this argument wants: nobody has a trustworthy causal estimate of how much faster AI makes software get built. The most rigorous attempt produced a negative number, and its authors have disowned their own instrument.

Something did change, but not the thing everyone says changed. The constraint on shipping a software product has moved twice, and where it now sits has almost no tooling around it.

Adoption is well measured. Output is measured badly. Speed is not measured at all.

Adoption is the solid ground. DORA's 2025 report puts AI use at 90% of software professionals, up 14 points year over year, at a median of two hours a day. Stack Overflow's 2025 survey of 33,662 developers found 84% using or planning to use them. Microsoft told its January 2026 earnings call it has 4.7 million paid Copilot subscribers, up 75%: self-reported, but inside a public-company disclosure, which makes it the most reliable number in the category.

Both surveys carry a qualifier that rarely travels with the headline. Only 4% of DORA's respondents trust AI output "a great deal". In Stack Overflow's sample 3.1% trust it highly, 45.7% distrust it, the top frustration for 66% is code that is "almost right, but not quite," and 45.2% say debugging AI-written code takes them longer. Adoption is near universal; confidence is not.

Output is where the story gets thin. New App Store releases hit 557,000 in 2025, up 24%, the biggest year since 2016 — but 2016 was roughly a million, so the best year of the AI era sits 44% under a decade-old peak. Google Play went the other way: available apps fell from about 3.4 million to 1.8 million between early 2024 and April 2025, a 47% decline driven by Google's own quality enforcement, though new releases there rose 7.1%. More things are being made; fewer stay listed.

The defensible claim is narrow: the cost of a first version collapsed. Getting from nothing to something a stranger can open in a browser is far cheaper than in 2022, which is a real change and a different claim from "engineering throughput went up." If you have been comparing yourself to people who ship four products a quarter, you are comparing yourself to an unmeasured quantity.

The channel that used to carry small products for free lost two fifths of its clicks

A randomized field experiment by researchers at the Indian School of Business and CMU Heinz instrumented 1,065 US desktop users through a Chrome extension and observed 68,089 real searches in early 2026. When an AI Overview appeared, outbound organic clicks fell 39.8% and zero-click searches rose 34.5%; Overviews triggered on roughly 41% of queries. User satisfaction showed no significant difference, which is the part worth sitting with. The searchers are fine. The people on the other end of the link are not.

Work that through for an ordinary query mix. If 41% of your queries surface an Overview and those lose 39.8% of their clicks, you lose about 0.41 × 39.8%, or roughly 16% of your organic traffic, before a single competitor does anything. Pew Research got a compatible number a different way, tracking real browsing for 900 US adults across 68,879 searches: with an AI summary present, 8% of visits produced a click on a traditional result; without one, 15%. Only 1% clicked a link inside the summary. Google publicly disputed that study, which you should know before you quote it. Nor has a replacement channel arrived: weekly AI chatbot use for news rose only from 7% to 10%, with no increase at all in the US, UK, France or Germany.

Paid acquisition repriced in the same window. Meta's Q1 2026 results, an audited filing, report average price per ad up 12% year over year. Benchmarkit's 2025 B2B SaaS study puts the median new-customer CAC ratio at $2.00 of sales and marketing spend per $1.00 of new ARR, up 14% in 2024, with payback 12.5% longer than in 2022; it discloses neither sample size nor method, so treat it as directional only. On a $20-a-month product that is $240 of first-year ARR against $480 of implied spend, which a funded company amortizes over four years and one person cannot. In MicroConf's survey of roughly 700 independent founders, 57% of those running ads either wait seven months or more for a return or cannot tell whether the ads work at all — self-reported and self-selected, but a fair picture of the practice.

Note what is still missing. Nobody has directly measured "distribution is now the binding constraint"; five indirect indicators pointing the same way make an inference, not a finding.

Distribution didn't get solved. A filter got harsher.

The comfortable story is sequential: building got easy, then everyone figured out marketing, and now we are in the distribution era. What the evidence describes is simultaneous. The number of products rose at the same moment the free channel that used to carry small products degraded by roughly two fifths. Those two happening together is not a handoff from one problem to the next. It is a filter tightening while the queue in front of it lengthens.

That changes how you read every founder story you encounter. The products still standing in mid-2026 are not evidence that distribution got easier; they are the output of a harsher screen. Selection, not method. Which is why copying the tactics of a product that made it through tells you less than it did three years ago, and why "build in public" advice has quietly stopped generalizing.

What nobody counts

No one publishes what share of AI-built products ever earn a dollar. Lovable announced 25 million projects created in its first year and 100,000 new ones daily, self-reported, with no revenue denominator, no active-project count and no survival rate attached, and a "project" includes every abandoned first prompt. Stripe's 2025 annual letter comes closest to a conversion signal: 20% of Stripe Atlas startups charged a first customer within 30 days, up from 8% in 2020. That also says 80% did not, and Stripe never publishes what share eventually do.

Survival rates for solo software products are measured nowhere. The BLS tracks establishments with payroll, which excludes almost every solo founder. The Census counts nonemployer businesses but not their survival. App stores publish no cohort retention. Product Hunt has never published a launch-volume series, so any figure you see came from a scraper counting featured launches against an undisclosed submission ratio. And the build-in-public evidence base is a voluntary, unaudited, survivorship-biased sample of people whose numbers went up.

The sharpest gap is in government data. US Census counts of nonemployer businesses with $1 million or more in receipts went 57,822 in 2021, to 116,803 in 2022, to 117,060 in 2023: a 102% jump, then a 0.2% increase, which is a dead stop. Average revenue across all roughly 30.4 million nonemployer businesses is $57,611, below the US average wage of $67,920. The 2024 and 2025 figures are unpublished, since Census runs a two-to-three-year lag. Anyone telling you AI is minting one-person millionaires is saying it without data, and the last observation before they started saying it was flat.

The third constraint: the survivors have revenue, and revenue is its own problem

A product that gets through the filter acquires a finance function with nobody staffing it, and it starts with money arriving later and smaller than the dashboard implies. Stripe's first payout is typically scheduled 7 to 14 days after the first successful payment, with settlement after that at T+2 business days in the US and T+3 in the UK and supported EU states. A UK seller taking a £1,000 annual plan from a US customer pays 3.15% plus 20p on the international card and 2% for conversion, £51.70 on that one charge, and Stripe Billing adds 0.7% on top, so £1,000 of "revenue" lands as about £941. If that customer refunds in month two, the fees from the original charge are not returned.

Then there is what that £1,000 is. On the day it lands it is cash and not revenue. Under ASC 606 and IFRS 15 you hold a contract liability and release a twelfth each month as you deliver, so eleven months of that balance is an obligation sitting in your account looking exactly like profit. Bessemer puts it plainly: committed MRR "can very often be disconnected from the 'cash health' of the business". That gap is the subject of why MRR and cash are different numbers, and it is widest when growth is fastest.

Tax starts at the first foreign customer, not at a revenue threshold. A seller established outside the EU gets no version of the €10,000 Article 59c allowance, because that provision requires establishment in a Member State. A UK sole trader with £5,000 of turnover who sells one €10 subscription to a German consumer owes EU VAT on it, and HMRC gives them exactly two ways to settle it: register for the non-Union scheme in one member state, or register in every member state they sell to. The UK's own registration threshold has no bearing on either. The rest of the picture, including US state nexus and what a merchant of record does and does not absorb, is in the guide to VAT and sales tax when you sell software.

And nothing is withheld on the way in, so the whole income and self-employment tax bill on a first profitable year lands at once, months later, on money already spent. What that calls for is spend control and cash visibility sized for one or two people rather than a finance department, the gap a Ramp alternative sized for one person addresses; and if you sell client work alongside the product, the effective hourly rate calculator shows which half funds the other.

The three positions, side by side

Read the middle column downward: the loudest opinions in this market sit on the least data. The third row is where that costs the most, because the confident advice runs out exactly when real money starts moving, and because it is the row where nothing is built at the right size, which is the premise behind Worklyn's page for product founders.

Which constraint is binding for you right now

Constraints are ordered. Only one is limiting you at any moment; working on the others is recreation. Take the first question you answer with a no.

1. Can a stranger use the thing today, without you present? If no, building is binding, and it binds on scope rather than tooling. Nothing in the last two sections applies to you yet.
2. Have at least 100 people you do not personally know tried it in the last 30 days? If no, distribution is binding and your conversion rate is not measurable yet, whatever your analytics claims. If yes and almost none of them paid, the constraint is the price or the product, not the channel.
3. Without opening a dashboard, can you say what landed in your bank last month, what share of that is tax you have not set aside, and what share is prepaid service you still owe? If no, the third constraint is binding, and it is the one that ends products that are otherwise working, because it fails while every growth number looks good.

The test separating the second from the third is one subtraction. Take last month's collected cash, subtract the tax owed on it, subtract the unearned portion of every annual plan, and divide by monthly fixed costs. Under three months, distribution is no longer your most expensive problem, because you will run out of time before you solve it. Do that in Worklyn's runway calculator rather than in your head, since the two figures people leave out are the tax reserve and the deferred portion and both make the answer smaller. If it comes out under three in a quarter that felt good, plan the slow one now.

Worklyn's 13 Weeks Out projects cash forward from your real invoices and bank feed, so the month an annual cohort comes up for renewal is visible before you have spent it.