What company values are actually for
tl;dr I tested whether values tell you anything about a company. The answer is yes, just not about the company.
Somewhere in the operating manual of modern tech is a rule that a company must publish its values. Not live them, necessarily, but publish them: on the careers page, between the mission and the benefits, four or five of them, each a few words long, each the product of more meetings and debate than seems reasonable.
I collected 343 of these lists from technology companies in the US, UK and Europe, 1,687 values in all, and built a small tool (for anyone as weirdly-interested as I am about these things) to explore them. I wanted to test something specific: whether a candidate reading a values page, taking it at face value but reading it carefully, could learn anything true about the company behind it: what the culture is like, where the power sits, the things you actually want to know and can’t ask in the interview.
The short version: the page won’t tell you what the company is like. It will tell you something better, and I don’t think it means to.
First test: does the content describe the company?
If values were chosen to fit the company, those choices should leak. Fintechs should lean on integrity, hardware companies on craft, consumer companies on the customer. I’d have bet on all three, so I tagged every company by sector and tested every concept that appears at twenty or more companies.
Nothing survives. Fintechs name integrity at 30%, against a corpus average of 29%. Hardware companies name craft at 8%, which is the corpus average again. And consumer companies, whose entire business is customers, turn out to be the second-least likely of the six sectors to name the customer at all: 30% against 37%. Shuffle the sector labels at random 20,000 times and 88% of the shuffles produce a more organised table than the real one, so the lists are less sector-shaped than a random relabelling.
The ideas don’t travel in groups either. You’d expect some of them to arrive together, that a company naming “speed” might be likelier to also name “quality” to hold (and recognise) a healthy tension. So I dealt them all out again, 20,000 times over, keeping every company’s hand the same size and every idea exactly as popular as it really is. The only thing that changed was who held what. Eight pairs looked like they clung together. A typical random deal produces about six, and 15% of deals produce more than eight. I had also just run 231 separate tests at once, which is an excellent way to discover things that aren’t there. Correct for that and none of the eight was left standing. Speed never brings quality. Nothing brings anything.
So when it comes to selecting values, there’s a common set, and nearly everyone draws from it: 308 of 343 companies name at least one of the top five ideas, and some version of “one team” appears at 46% of companies whether they make meal-boxes or satellites. But the median company draws only a third of its list from that pool. A couple of well-used values to make the set recognisably ‘values-y’, and the rest from anywhere, related to nothing, predicting nothing, signifying nothing…
The first test results are: which values a company picks carries no information about what kind of company it is. Read the content all you like. There’s nothing much in it.
Second test: does anything else leak?
Here’s where it gets interesting, because it looks like the lists do carry a signal. It just isn’t about the company.
I ran a test to see whether a company’s values sit closer together than five values pulled at random. Not closer in topic, which would only mean companies pick related themes, but closer than a random list matched on the same topics which is the trickier test.
They do. Out of 5,000 shuffles, not one produced lists as internally coherent as the real ones. To be clear, the effect is modest: pick a real list and a matched fake at random and the real one is more coherent 62.5% of the time, where 50% would be nothing, and about a third of companies write lists no more coherent than a random draw. But it’s there, it survives every control I could think of.
But you don’t need a model to see it, because 15% of companies show it in the letterforms. Paddle’s values are “Paddle for others”, “Paddle together”, “Paddle simply”. Zilch: “Do it”, “Do it brilliantly”, “Do it the right way”, “Do it for our customers”. Perplexity publishes five values and they are “Craftsmanship, Ownership, Entrepreneurship, Scholarship, Partnership”, which I choose to believe was discovered rather than written. Twenty-two companies template the entire list[1].
I know how these get made, and it isn’t one person at a desk. It’s a workshop: forty Post-it notes, everyone contributing, genuine mess. But that’s not what ships. Nobody publishes the stickies. Between the workshop and the careers page, one voice goes back over the material, culls it, and rewrites what’s left until it reads like it came from somewhere. “Paddle for others / Paddle together / Paddle simply” is not what forty Post-its look like; it’s what one person does to forty Post-its to get it ready for sign-off. The coherence in the data isn’t shared conviction, it’s finish, and finish has an author even when the ideas had a committee.
(Even the weird ones belong. The statements my coding couldn’t or hasn’t yet classified, “Enlightened Despots”, “Be Awesome”, fit their own lists exactly as well as the normal ones fit theirs. They didn’t escape the edit. They’re what the edit looks like when the editor is having a good day.)
What I think’s going on
I originally decided that this exploration meant values were a “genre exercise” and moved on. But that niggled, as it never explained why the genre has the shape it has, but I think there’s a way of thinking about them which does.
A company’s values page was never a description. It’s an instrument. Values are that rare corporate text that gets quoted back at you, unlike the mission or the strategy deck. When you lose an argument it arrives as “disagree and commit”. At appraisal time it arrives as “not yet demonstrating ownership”. In the budget round, “scrappy” and “high agency” let you know the hoped-for headcount isn’t coming. Nobody was ever stack-ranked against a mission statement.
Once you see the page as an instrument, every odd property the data found stops being odd.
The content is generic because it has to be. A rule that will be applied later, to anyone, in disputes nobody’s had yet, must be portable and unfalsifiable. “Ownership” is a beautifully designed tool precisely because it describes nothing and can be cited against anything. Asking why the values don’t resemble the company is like asking why the rulebook doesn’t resemble the players.
That workshop-to-rewrite-to-sign-off funnel is the whole point, a ritual that looks like co-creation. Many voices in the workshop, one voice in the edit, one signature to seal the deal. These are the criteria everyone will later be judged by, and the part of the process where everyone participates and hands over their consent is the part that never ships. What ships is the selection, and the selecting hand doesn’t sign its work. The whole process is built to produce something that reads as co-created and lands as endorsed, which is a useful property in a document that will later be applied to the people who think they wrote it.
And the silences are the tell. Of 1,687 published values, the number that mention profit, revenue, margin or money at all is three, and one of the three is against it. Work-life balance, in a corpus built from the careers pages of companies actively recruiting, is named plainly by one company (Dataiku, “Work-Life Balance”, speaking with their whole mouth). You don’t put terms in the rulebook that might be cited against management. A published value is a stick, and everyone involved knows which way the stick will point, so the ones that would point upward never make it onto the page.
Which also means those rare deviations are the only expensive words in the genre. Everyone can afford “teamwork”, but somebody at Dataiku decided to publish a value their own leadership could be beaten with, and somebody at Hopper wrote “Grow Revenue” out loud. Whatever else those choices are, they cost something, which in this genre makes them the only words worth reading at face value.
Who signs the rulebook
One more thing from personal experience rather than data: I’ve spent twenty years inside companies like these, and values don’t ship until the most powerful person in the building is willing to stand behind them. The drafter varies. An agency, a copywriter, a head of people, occasionally the founder at 1am. The signature doesn’t.
So everyone contributes, one voice edits, one authority signs. And sign-off is the stage that makes the silences deliberate. A CEO initialling the values page knows exactly where (and when) that document gets used, which means the absence of “balance” and the absence of “profit” aren’t oversights, they’re choices made by someone who understood the instrument they were approving. The data seem to show one hand finishes list. Experience tells me one authority endorsed it. Whether those are the same person or a pen and a signature, what’s published is what power was willing to be quoted on, and what’s missing is what power declined to be held to.
What you can actually use
So, what’s left for the person the page is nominally for? I’ll suggest three rules, none of which require believing a single word of the content.
- Ignore anything from the common pool. “Teamwork”, “ownership”, “customer focus”: these are habits, chosen by nobody, meaning nothing. This is most of what you’ll read.
- Read the form instead of the content. A templated or eerily coherent list is telling you one voice runs to the end of the process unchallenged. Whether that’s clarity or a warning depends on the voice, but it’s the closest thing to an org chart the page will give you.
- Price the deviations. Anything that breaks a genre silence, money, balance, the buyer, was expensive to publish because they have to own their words. Expensive words are the only ones that carry information, and there are about a dozen of them in 1,687.
What you’re actually holding when you read a values page isn’t a window into the culture; the data is clear that it describes nothing. What you’re holding are the terms you’ll eventually be judged by, drafted by one sensibility, countersigned at the top, published with the awkward clauses removed. It fails completely as a portrait of the company, but it can’t help succeeding as a portrait of whoever approved it: their taste, their tics, and the silences are a precise list of things they’d rather not be quoted on.
The corpus, every value, the concept coding and all the overrides are browsable at values.stach.uk. The coding is my interpretation, not a definitive one; about a tenth of the statements resisted being, or just haven’t been coded yet, and I’d genuinely like to hear where you think it’s wrong. At some point I’ll add in a way to make it easy to let me know if you think something’s wrong, but if you want to add a company you can[2].
All fifty were removed before the statistics above, since repeated words would fool a similarity measure, and the signal holds on the remaining 290. So the templates show the hand in the companies that make it obvious, and the embeddings find the same hand in the ones that don’t. ↩︎
As of 21 August, 2026 this is now working. Values and Companies can now be flagged in the UI and it’ll give you a form to fill out. ↩︎