Business Issue #127 ·

10,000 Signups a Day, But Revenue Won't Budge

The problem wasn't the email copy — it was who Railway was sending it to, and the fix was surprisingly simple.

Oswarld (Kwangseob Ahn) · GTM & AI strategist working the seam between technology and business

10,000 Signups a Day, But Revenue Won't Budge

Opening

Dear reader, last month an intriguing post appeared on the engineering blog of Railway, the cloud deployment platform. The title alone was provocative: “Kill your onboarding.” I recently built a financial-product analysis site as a side project, and that’s when I used Railway for the first time. Honestly, Vercel and Supabase had always been enough for me. I’d toyed with adding Sentry for anomaly detection, but even that felt like overkill — then a cost issue finally pushed me to try Railway, and it was better than I expected. And beyond being technically solid, the service turned out to come with a surprising number of interesting sales and marketing stories.

Railway is a service that takes in more than 10,000 signups a day. 2.9 million cumulative signups as of April 2026. But when they opened up their internal analysis, they had identified 21,000 company accounts matching their sales ICP1 — and had contacted less than 1% of them. To give you the conclusion up front: the problem wasn’t bad emails, it was the wrong recipients, and the fix was astonishingly simple.

🛢️ The Oil Was Already There: Companies That Collect Data and Never Use It

PLG2 companies face a structural dilemma. To maximize the user experience, they ask nothing at signup — and as a result, they can’t tell which users are individuals tinkering as a hobby and which are company teams building production infrastructure.

In other words, there are people like me building a site as a personal hobby, and there are people running company-level projects on the same platform — and from Railway’s standpoint, the B2B customers are obviously the ones to land. Railway was in exactly this state. It boasts an extremely smooth onboarding — deploy in seconds with nothing but a GitHub account — but deliberately never asks questions like “Which company are you with?” or “How many engineers do you have?” They go so far as to state that they “deliberately reject” the currently fashionable Granola-style onboarding, the kind that grills you on role, team size, and use case right after signup. Most products do ask those questions. (You’ve probably been subjected to them.)

The problem is the price they paid. A tiny sales team — 1 AE (account executive) and 2 SEs (solutions engineers) — had to face 21,000 potential company accounts with no way to know which ones to prioritize. Railway founder Cooper put it this way.

“The oil fields were all in place — nobody was drilling.”

What’s interesting is that Railway wasn’t short on data. They had been collecting product events through PostHog for years, and their internal system ‘backboard’ recorded each project’s service composition, instance sizes, database connections, even deployment outcomes in fine detail. Does this pattern sound familiar? It wasn’t that “we lack data” — it was that nobody was asking questions of the data already piled up.

This isn’t Railway’s problem alone. As of 2026, 58% of all companies have adopted the PLG model, yet only 25% have introduced a PQL3 framework that scores leads based on product usage data. That means the other 75% sit on user behavioral data while still relying solely on the marketing funnel.

🔍 No Hobby Developer Downloads a SOC 2 Report: The Art of Signal Selection

The approach Railway solutions engineer Des Conlon took was remarkably classical. He didn’t run an ML model; he pulled out every behavioral dimension one by one and computed the ratio at which it occurred in company accounts versus hobby accounts. He stacked dbt models on top of an analytics tool called Hex and iterated through dozens of GROUP BY queries to whittle down the candidate signals.

There was one crucial move here: he boldly threw away the signals that looked the most predictive.

For example, SSO usage was an indicator that answered “is this account a company?” with nearly 100% accuracy — but SSO is a feature only customers who have already finished a sales conversation can enable. It merely re-identifies people you already know about. This is what data analysis calls label leakage4 — the answer bleeding into the input. Conlon called it “1 = 1”: verifying a self-evident fact proves nothing.

One rule survived. Only “behaviors a hobby developer could theoretically do but in practice almost never does” were admitted as signals. The final five that made the cut are fascinating.

  • Trust Center document downloads: the strongest probabilistic signal. A hobby developer has no reason to download a SOC 2 report. Outreach based on this signal reportedly drew a response rate of about 50%
  • Seat count and its growth trend: after Railway abolished seat-based billing last year, this actually became a purer team-size signal
  • Credential resets: the behavioral pattern of organizations with internal security policies or password-rotation cycles
  • External DB connections: how fast real data gets connected is what separates hobby from work
  • Specific deployment-failure patterns: not “failed once” but “wrestling with production-grade problems” They weighted these 5 by uplift ratio and summed them into a score. Not gradient boosting, not random forest — a linear weighted sum. Why go this simple?

There were 2 reasons. First, most of the behavioral data covered only about a day, so a complex model was bound to overfit. Second, the person acting on this score is a single AE. A reason like “trust doc + 8 seats + DB connection” can be grasped instantly, but with nothing more than a model’s “0.87,” there’s no way to notice when the judgment is wrong. In Conlon’s words, “a model you can read almost always beats a model that wins on marginal accuracy.”

📧 Don’t Drip, Trigger: Reinventing the Email

The score flows down two paths: mid-range scores go to automated emails via Customer.io, and high scores go to direct human outreach.

But these automated emails are structured completely differently from a conventional onboarding sequence. Instead of a time-based drip — “day 1, day 3, day 7 after signup” — they’re event-triggered, sent the moment a specific behavior occurs.

The send conditions are strict, too. Two things must hold at once.

  • The event must correlate with the likelihood of being a company account
  • The event must mark a moment when the user is slightly stuck or making an important decision And the email content is not a meeting request. It doesn’t pitch the enterprise tier either. It goes out under an actual sales rep’s name, and the message is roughly at the level of “it looks like you’re working on something like this — anything I can help with?”

The results were dramatic. The generic onboarding emails had a 27% open rate; per-trigger open rates climbed into the 50–70% range. Out of 300–400 emails sent on day 1, 2 real replies came back.

Of course, these numbers come with important context. Apple Mail’s privacy protection can inflate open rates, and since the emails went to the most interested people at the most fitting moment, the selection bias is large. Railway itself acknowledges this and says it’s designing an A/B test.

But the direction itself is meaningful. Most PLG email marketing guides recommend “behavior-based triggers”; what Railway did differently was layer an ICP filter onto the trigger condition itself. Not simply “email users who connected a DB,” but the double condition of “email only users who connected a DB and are also likely to be a company account.” That difference compressed the send list from thousands a day down to 300–400 — and it was precisely that compression that created the quality.

Oswarld’s Take

Honestly, when I first saw this case I slapped my knee. It laid bare the most common failure pattern I’ve seen countless times while building GTM strategies.

The pattern goes like this. When a company’s growth is sluggish, the first thing it touches is the email copy or the landing-page design. Because that’s what’s visible. But where the problem actually lives is the targeting layer — “who are you talking to?” If, like Railway, 10,000 people walk in every day and you’ve contacted only 1% of 21,000 company accounts, no amount of copy tweaking will help. Sending the perfect message to the wrong person is a perfect waste.

And what I find most striking in this case is something else. The problems Conlon experienced at Looker (later acquired by Google) and at Railway are perfectly symmetrical. At Looker, customers had reasonable questions but no data infrastructure to answer them; at Railway, the infrastructure was in place but nobody was asking the questions. Neither case was a matter of “not having the data.” The link between data and questions was missing.

It’s no coincidence that “GTM engineer” job postings passed 3,000 this year. Instead of a sales team guessing at “enterprise-ish signals” by gut feel, this sequence of work — computing uplift ratios, filtering out label leakage, building a readable scoring system — is the capability whose market price is climbing fastest right now. And it’s not a capability AI replaces; it’s one that matters more in the AI era. However good the model, deciding which signals go in is still a human’s call.

In Korea, GTM-related roles remain oddly unpopular, but I keep writing these steady posts believing a boom will come someday… In the end, the money has to be made. You can’t keep selling dreams forever…

Closing

To sum up.

  • PLG’s real bottleneck isn’t the absence of data but the absence of data use. 75% of PLG companies hold product behavioral data yet don’t use it to score leads
  • A simple, readable score beats a sophisticated AI model in the field. A model the person acting on it can’t understand can’t be corrected when it’s wrong
  • Email performance is decided by timing and targeting, not copy. Changing “to whom, and when” alone can lift open rates more than 2x Next time you audit your organization’s onboarding email performance, check one thing before touching the copy. Is that email actually reaching the people it should?

Is your organization wiring post-signup behavioral data into sales? Or are you still at the “generic welcome email” stage? Let me know in the comments.

References & Further Reading

Primary sources

Background

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). At the university he teaches statistics and data analysis, including business data management and business analytics, while in the field he leads GTM strategy and AI strategy consulting, designing the interface between technology and business. He has published academic research on memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a project curating global AI papers every day. He completed a master’s program at Korea University’s Graduate School of Management of Technology and a KMBA. He is the author of Those Who Outsource Their Thinking: Homo Brainless.

Footnotes

  1. ICP (Ideal Customer Profile): The standard that defines “which organizations would get the most out of our product,” built by combining company size, industry, region, funding stage, and so on. ↩

  2. PLG (Product-Led Growth): A strategy that acquires and converts customers through the experience of using the product itself rather than through salespeople. Slack, Notion, and Figma are the canonical examples. ↩

  3. PQL (Product Qualified Lead): A lead whose purchase likelihood is scored on product usage behavior. Unlike an MQL, qualified for something like downloading a whitepaper, it’s grounded in actual product use — so conversion rates run 2–3x higher. ↩

  4. Label Leakage: When the answer you’re trying to predict seeps into the input data. It’s like the answer being printed on the exam paper — the model turns out to be useless in the real world. ↩