Traction Got Cheap. Everyone Still Prices It Like Proof.

I spend between six hundred thousand and eight hundred thousand dollars a month buying attention for other people's products. It is the least glamorous line on my resume and by far the most useful, because it has permanently ruined my ability to be impressed by a growth chart. When somebody shows me a curve going up and to the right, I no longer see a business. I see a media budget and a question about what that curve does ninety days after the spending stops.

That reflex used to be a personality defect. This year it started looking like a methodology, because two pieces of research finally put numbers on something I had only been able to describe as a smell.

Before I go further, the disclaimer that this whole essay depends on. I am not an investor. I do not manage anyone's money, I hold no license to advise on where it goes, and none of this is advice about where to put yours. I am writing from the other side of that table, as the founder who has to sit in a room full of people with better spreadsheets than mine and explain why the thing he is building does not have a chart yet.

Traction is a purchase now, not a verdict

Here is what changed. Building a company that appears to work no longer requires building anything. The model is an API key. The product is a prompt, a thin interface and a payment page. A competent operator can go from idea to paying customers in a matter of weeks, and I say that as someone whose actual professional skill is compressing exactly that timeline for money.

Traction used to be expensive to fake because it was expensive to produce. You had to build something, get it to work, and convince a stranger to pay for it, and each of those steps filtered out people who could not do the previous one. That filter is gone. Customer acquisition is now a marketing exercise that runs largely independent of whether anything defensible exists underneath it, which means the number that used to certify depth now certifies only that someone knew how to buy an audience.

ChartMogul's analyst-in-residence Kyle Poyar put it better than I have managed to. He called it the curse of the AI wrapper: the downside of being easy to buy is being easy to cancel.

What the retention data actually says

Poyar and ChartMogul scraped roughly 3,500 software companies and sorted them into B2B SaaS, B2C SaaS and AI-native, then compared retention across the groups, looking only at businesses that had already cleared 250,000 dollars in annual recurring revenue. The AI-native cohort came to about 200 companies. Median net revenue retention for B2B SaaS was 82 percent. For the AI-native group it was 48 percent, with gross revenue retention at 40 percent, which is worse than consumer software.

Gross retention of 40 percent means the company loses roughly sixty percent of its revenue base every year before it counts a single new sale. And the damage sorts almost perfectly by price. AI products selling above 250 dollars a month retained 70 percent gross, essentially matching ordinary business software. Products under 50 dollars a month retained 23 percent.

Then comes the line from that report I cannot stop thinking about. Early-stage startups in the dataset were growing at more than 200 percent a year with net revenue retention below 40 percent. Growth and durability had come apart completely. Poyar also found that among low-retention companies, three times as many were shrinking as were growing quickly, which is the same fact viewed from twelve months later.

The pilot is the traction

The buyer-side research points the same direction. MIT's Project NANDA published a study in July 2025 called The GenAI Divide, built from 52 executive interviews, surveys of 153 leaders and an analysis of more than 300 public AI initiatives. It found that 95 percent of generative AI pilots produced no measurable impact on profit and loss. Only 5 percent of integrated systems created significant value, against 30 to 40 billion dollars of enterprise spending.

Underneath that headline sits the part worth copying into your notes. Sixty percent of organizations evaluated enterprise-grade AI tools, twenty percent got as far as a pilot, and five percent reached production. NANDA attributed the failures to brittle workflows, tools that never learned the customer's context, and misalignment with how the work actually gets done. Not model quality. The models were fine; the depth around them was missing.

Now put the two datasets next to each other, because they describe one machine from opposite ends. A pilot is a logo on a slide and a line in a revenue chart. It is, in every sense that a growth number recognizes, traction. And ninety-five percent of the time it converts into nothing at all, which is precisely why the vendor's retention curve falls off the table twelve months later.

The companies with nothing to show yet

The asymmetry that follows is the reason I wanted to write this down. Depth takes years. If you are building something that has to work at the layer below the model, you are looking at a long stretch of engineering during which you have no chart, no logo wall and no growth rate, because you have not shipped the thing yet. Meanwhile the wrapper that took nine weeks has all three.

I have been on both sides of this. I have shipped things fast that worked immediately and meant nothing, and I have spent years on architecture where the payoff arrived so late that nobody connected it to the work. The second kind is the only kind I have ever been proud of, and it is also the kind that is impossible to evidence while you are doing it.

Here is where the money makes it worse instead of better. The company with real depth is the company that spent its capital on engineers, and the one thing it therefore cannot afford is the market research, the paid acquisition and the category-defining content that would let it prove it is different. The evidence that would distinguish substance from surface is expensive. The signal that misleads is cheap. Those two facts point the same way, and the way they point is at the wrapper.

What i look at from the builder's chair

So I have stopped reading traction as an answer and started reading it as a question about what produced it. Gross retention before net, because net revenue retention lets expansion inside a handful of happy accounts paper over the sixty percent walking out the back. Revenue that survives the departure of the internal champion who bought it. Whether the thing still works when the underlying model is swapped for a different one, which is the fastest test I know for whether anything was built below the model layer.

The one that tells me the most is the least quantitative. What did you have to build that you could not have bought? If the honest answer is an interface and a system prompt, the retention numbers above are not a risk to that business. They are a forecast.

None of this is a complaint about wrappers. Some of them are good businesses and a few will grow into deep ones, and the operators running them are doing something genuinely hard, which I know because it is roughly what I get paid to do. The complaint is about the measurement. We are in a period where the cheapest thing to manufacture is the exact signal we have all agreed to treat as proof, and the companies most damaged by that arrangement are the ones quietly doing the work.

If you are looking at a traction chart this quarter, ask what it cost to produce and what happens to it when that spending stops. And if the company in front of you has no chart at all, that is not automatically a red flag. Sometimes it is just a receipt for building something.