The claims this site will not make

Most marketing copy is checked by whoever is in the room when it is written. Ours is checked by a file. site.test.ts imports the strings these pages are built from, holds them to a list of rules, and fails the build when one breaks. The test beside it does the same for the pricing table, against the backend's own plan rules.

That is an odd way to run a website, and the reason is that the failure it catches is invisible. A comparison page that quietly drops its section on when to buy the other product still renders. An integration page with no limits block still reads well. The cost lands months later, on a reader who believed us about something else.

This post is bound by every rule it describes, which is the joke and also the point.

No outcome we cannot measure

The first rule is a list of things a sentence may not contain: a percentage, a price in somebody's currency, a multiple, or a promise about the hours you will get back. It runs over every string the autonomy pages and the workspace panels are built from, with no exception for a figure somebody feels confident about.

The reason is not modesty. A number inside a claim either came from somewhere a reader could check, or it came from a slide. Ours would be the slide: the platform is new and holds no customer AI usage, so a figure about what Sell does to your close rate or your week is one we would have decided in a room and then set in a serif font.

What that costs is not the wrong figure. It is what the wrong figure does to the sentences beside it: once a reader works out that one number is decoration, the true parts become decoration too. Cutting the figure is cheaper than earning that trust twice.

Every credit figure says where it came from

There is one place we do publish numbers, and it is where they carry the most weight: the credit cost of each AI feature we have measured, which the pricing configurator multiplies by figures you type in. The test makes every one of them name the feature, the model that ran, the date it was measured, and whether it was measured at all. A figure that arrives without those is a figure somebody guessed, and the configurator would multiply it just as willingly.

The figures also have to reproduce. Each publishes the token counts and the model multiplier behind it, and the test recalculates the credits with the same function the billing code uses, so a figure that does not reproduce fails the build. Each key is checked against the product's own list of AI features, so a price cannot exist for a capability that does not. What the numbers mean, and what a single run does and does not tell you about your month, is the subject of the essay on credits.

Nothing that says money moves while you are asleep

This is the rule most likely to break by accident, because the sentence is the natural one to write. A product that reads your workspace overnight and drafts your follow-ups sits one sentence away from one that settles your invoices overnight, and that sentence would be false here. Card collection is still being switched on in this deployment. The Payments panel says so, the Stripe page says so, and a test asserts that the panel still says so. When it is switched on, the charge will be made on your own connected account rather than ours.

The overnight round has two jobs that come near money, and neither acts. One finds work a client accepted with no invoice raised against it and tells you, because raising it is yours. The other reads the invoices past their due date and writes the reminder, then waits for you whatever your other settings say.

What the rule protects is not our modesty. It is the reader who believes the overclaim, switches something on, and learns the truth from their client.

Autopilot is off until you turn it on

Every part of the site that mentions Autopilot has to say that it is off until you turn it on. The test reads the whole autonomy section, and the workspace panel separately, looking for one of three phrasings, so a draft that has quietly lost the sentence does not ship.

The same rule bans the register that promises software with nobody in it. The plain description is better anyway: one round a night, and what it produces is drafts. It writes them, you press Send. Letting it send on its own is a second switch with a daily limit that starts at zero.

The rule cuts against us in the other direction too. A second test requires every job in the round to carry both a ceiling and a refusal, and fails a card that states only what the job can do, so the section cannot be made shorter by dropping the half that bounds it. That is what keeps the refusals printed rather than implied.

When to buy theirs, and what ours cannot do

Comparison pages are the ones most likely to embarrass us, so they carry two extra rules: each opens with what the other product is genuinely good at, and each names real situations where you should buy it rather than ours. A page that gives no reason to buy theirs fails, and so does one whose generous opening is too short to be generous.

So the pages say things like this. If you want meeting notes, you want them free, and you want them today, buy Fathom. If your CRM is HubSpot or Salesforce and native sync matters more than anything else, buy the notetaker that has it. If your process is genuinely unlike anyone else's, build it yourself: you will own the code outright, and we do not offer that. The stack comparison has rows where our own column says no, because a specialist beats us in its specialty.

No competitor's price appears anywhere on this site. Someone else's pricing is not ours to keep current, and a stale figure is the fastest way to lose an argument in public. Anything that moves is marked as something to check on their own site, under the date we last reviewed the category.

Integration pages are held to the matching rule: none of them ships without its limits. The browser extension is desktop Chrome and Edge only, and it cannot record a phone call on the same phone, because the operating systems forbid it. Google calendar and mail are separate grants, and meetings run in our own browser room rather than in Meet. HubSpot, Salesforce and Asana have no integration, and we have not tested ChatGPT's connector, so we do not claim it works.

Why these are tests and not a style guide

A style guide is a document people agree with and then forget on a deadline. A test fails the build. That is the entire mechanism, and none of it is clever.

One test counts the jobs registered in the backend's Autopilot module and fails when the page walks through a different number, so the section cannot become a description of last quarter's product. One reads the Python that decides what each plan may do and fails when a cell in the pricing table disagrees with it, so the table cannot advertise a limit the product does not keep, or keep one it does not disclose. One checks that no plan card sells a capability in its bullets and crosses it off four lines below.

The price of all this is that some pages are less punchy than they could be. An unfair comparison is worth less than the visitor who catches it, and a figure nobody can check is worth less than the sentence it pushed out. The one test we cannot run is you: sign up, and hold the product to this page.

Questions this raises

Does refusing to publish numbers just mean the numbers are bad?

It means we do not have them. The platform is new and holds no customer AI usage, so a figure about what Sell does to your week would be invented. When there is real usage worth reporting, it will arrive with its basis attached, or it will not arrive.

Do the tests actually catch anything, or are they decoration?

They catch the well-meaning edit. The failure mode is not somebody writing a lie. It is somebody making a comparison page punchier and dropping the section that says when to buy the other product, or adding a credit figure without the token counts behind it, or writing a lock card with a percentage in it. All three read fine in review. All three fail the build.

What if I find a claim on this site that breaks one of these rules?

Tell us and it gets cut. That is not a courtesy: the comparison pages carry a line inviting correction for exactly this reason. Rules that live in a test still depend on somebody having written the test, so the rules we have missed are the ones you will find first.