We deleted our best marketing stats because they weren't true

While porting three sections from an older version of our site, I found some very good numbers. 2.1M+ kg CO₂ saved annually. $12M+ in compute costs saved. 99.99% uptime. 200+ apps.

They were the strongest copy on the page. We removed all of them, and the pages got better. Here’s the reasoning, because “be honest” is easy to agree with and surprisingly hard to apply.

The numbers described a company we aren’t

We’re in private beta. The waitlist hasn’t cleared 25 people. So what does “2.1M kg of CO₂ saved annually” describe?

Not us. It’s a projection — what the mechanism would save at a scale we haven’t reached, phrased in the past tense as an achievement. Same with $12M in compute costs. Same with 99.99% uptime, which is a claim about a service history we don’t have yet.

The “200+ apps” one was even simpler. The headline said 200+. The underlying data file listed 40. The real number in our product is 148.

None of this was anyone lying. It’s what happens when copy gets written aspirationally, then survives past the moment the aspiration was obvious.

The test that made the decision easy

A single question resolved every case: if a customer asked “where does this number come from?”, could we answer without flinching?

• 2.1M kg CO₂ — no. It’s a model with assumptions we’d have to describe, applied to a customer base we don’t have.
• $12M saved — no. Same problem.
• 99.99% uptime — no. We haven’t been running long enough to claim it.
• 1,000+ tools in the mesh — yes. That’s a count of rows in our capability mesh.
• 47 live Google actions — yes. Countable, verifiable, and we did.
• 148 apps — yes. It’s what’s in the product.

Notice the surviving numbers are less impressive and more specific. That turns out to be the pattern.

What replaced them

For the environmental claim, we kept the mechanism and dropped the outcome. The section now explains that a share of everyday LLM traffic is greetings and filler that never needed a frontier model, that we answer those locally, and then says plainly:

Independent analyses of production LLM traffic put the share of trivial or duplicate calls in the tens of percent. Velaris removes that class of call entirely — your own savings depend on your mix.

That’s weaker as a headline and stronger as an argument, because it’s checkable. A reader can reason about their own traffic. Nobody can reason about our 2.1M kg.

For uptime, we replaced the number with the architecture: fallback chains, and what happens when a provider degrades. You can evaluate a mechanism. You can only trust or distrust a statistic.

The waitlist counter we hide

The same rule produced a decision I like more than the removals.

Our waitlist counter doesn’t display a number until it clears 25. Below that it says “open”.

The tempting alternative is to show whatever the number is and hope it reads as momentum. But “Join 6 builders on the waitlist” is worse than saying nothing — it converts a neutral moment into a negative signal. And the alternative most companies pick, inflating it, is the thing we just spent a week removing.

So the honest version has a floor. When the number becomes real, it appears, and it’ll be a real number.

What this costs, honestly

I don’t want to pretend this is free.

The specific numbers were better copy. “2.1M kg CO₂ saved” is memorable in a way “your savings depend on your mix” will never be. Some readers who’d have been impressed will now scroll past. In a competitive comparison against a vendor happy to project, we look quieter.

That’s a real cost, paid now, against a benefit that arrives later: nobody ever has to walk one of these back. No customer discovers the 200+ was 148. No journalist asks for the CO₂ methodology. No one on the team has to decide whether to repeat a number they know is soft.

For a product whose entire proposition is let this software act on your accounts, that trade seems obvious. We’re asking for a lot of trust. Starting the relationship with numbers we can’t source is a strange way to ask.

The rule we kept

Every figure on the site now has to be one of two things: a count of something we can query, or an outside estimate, attributed as an estimate.

Anything else gets cut or converted into a description of the mechanism. It’s a small rule and it removed a surprising amount of copy.

The numbers that survived — 1,000+ tools, 47 Google actions, 148 apps, 100% owned stack — are all in the first category. If you ask where any of them came from, we’ll show you the query.

You can check most of them yourself: drive the OS in the browser, or browse the integrations and count.