Two of these five pricing pages are not one page. Retool’s hands out more than one count of itself, reading after reading, and the controls our fetches disagree about are the same ones every time — Contact sales, Book a demo and Request trial. Our own fetches of Notion’s page disagree with each other too, though far more rarely. Linear’s, Airtable’s and monday.com’s pages have handed the same count to every fetch of them, on every reading we have ever taken — and that is the half that makes the first sentence worth reading rather than a complaint about our own instrument.
Five pricing pages, fetched and counted by code with no model in it — no buyer, no decider, nobody reading the page — each one fetched several times in a sitting, and in sitting after sitting since the middle of October. The question underneath every other thing we publish is this one: a reading of a stranger’s page is evidence about that page only if a second fetch agrees with it. On three of these pages it always has. On Retool’s it mostly has not: seven of the readings we have taken of that page watched its control count take more than one value, against eight readings that reached it at all. Notion’s page has done the same thing on two of them, against eleven readings that reached it. Our own single fetch of Retool’s page for the table below counted 230 controls, and the sitting behind this page counted a control more than that on most of its fetches.
Nobody was in the loop. No ghost user walked these pages, no model decided anything, and not one sentence or number here was written by one — which is why this is a reading and not a run, and why there is nothing on it about where a person got stuck.
- The instrument
- Each page was opened in a real browser, watched until its own list of controls stopped changing, and then counted — and that was done several times in a row, in one sitting, so that the page had the chance to answer differently. Nothing was signed up for, nothing was pressed, and no sentence and no number on this page was written by a model. A page is said here to hand out more than one count if any one of our sittings watched it do so, and said to hand out one only if some sitting established that over at least three fetches that reached it and no sitting we have ever taken disagreed. Those two tests are deliberately not mirror images, because the evidence is not symmetrical: fetches that disagree settle the question, and fetches that agree do not.
- When
- The counts here are counts of our own readings, not of anything either company publishes. They say how often a page answered differently when we asked it again; they do not say how often it serves each version to anybody else, and no share, percentage or probability is drawn from them, because that allocation is theirs and is not visible from outside. A page that has agreed with itself on every reading so far may disagree on the next one, and the asymmetry cuts one way only: the moment one reading sees two values, no number of later readings that agree takes it back.
- Pages read
- monday.com/pricingairtable.com/pricingretool.com/pricinglinear.app/pricingwww.notion.com/pricing
Counted on the pricing pages themselves.
We opened each pricing page, waited until its own list of controls stopped changing, and counted the controls that send a reader somewhere else in a new tab — recording, for each one, whether it goes to material of that company's own on a separate site or to somebody else's. These are a fact about these pages on 22 October 2026 and nothing else, and anyone can open the same pages and count them again.
| Pricing page | Controls on the page | Open a new tab | …to its own material on a separate site | …to somebody else's site |
|---|---|---|---|---|
| monday.commonday.com/pricing | 169 | 32 | 8community.monday.com, ir.monday.com, support.monday.com | 13 |
| Airtableairtable.com/pricing | 99 | 17 | 5community.airtable.com, support.airtable.com | 7 |
| Retoolretool.com/pricing | 230 | 10 | 4docs.retool.com | 2 |
| Linearlinear.app/pricing | 64 | 6 | 0 | 4 |
| Notion AIwww.notion.com/pricing | 73 | 0 | 0 | 0 |
What we read and are not publishing.
- Why either page serves more than one version of itself
- An A/B test, a rollout, a cache, a late-loading element, a logged-out variant — each is a claim about what a company meant to do, and nothing we fetch reaches intent. All we can say is that the page answered our fetches differently, which is a fact about the answers rather than about anybody’s plan.
- How often either page serves each version to anybody else
- Our counts are counts of our own fetches. The allocation is the company’s and is not visible from outside, so a share, a percentage or a probability drawn from our readings would be our sampling published as their configuration. The reading that produced this page refuses to derive one for the same reason, in its own records.
- That the three agreeing pages will keep agreeing
- Nothing here is a prediction. The claim is about every reading we have taken and stops there, and it is the direction a single disagreeing fetch would end permanently — which is exactly what makes it worth stating rather than assuming. If one of them disagrees with itself tomorrow, this page is the thing that has to change.
- Which of Notion’s controls come and go
- Our fetches of Retool’s page disagree about the same controls, reading after reading, so naming them is a fact the records carry. Notion’s page has answered differently on far fewer readings, and a set of controls named from that little is a set we would be guessing at. So the counts are here and the names are not.
- A price reading, or anything about what these products cost
- This page publishes no reading of what anything costs. It is about whether the number of controls on a page is the same number on a second fetch of it, and the price readings of these same pages are published, with their records, on the pages that compare them.
- Anything a person did
- Nobody walked these pages. There is no buyer here, no model decided anything, and no sentence on this page was written by one — so there is nothing here about where somebody got stuck, and these numbers are not evidence about what a reader would do.
Anyone comparing these five products can fetch all five pricing pages this afternoon. What they cannot do is find out that two of the pages they fetched would have answered differently if they had asked again — that takes having asked, repeatedly, for weeks, and it is the reason a single screenshot of a competitor’s pricing page is a worse piece of evidence than it looks. It is also why this page accuses nobody: serving more than one version of a pricing page is an ordinary thing to do, we do not say why either company does it, and a reason would be a claim about intent that no fetch of ours can reach. What the split does decide is everything else we publish. A count from a page that answers the same way every time can be compared with the same count a fortnight later, and a difference is theirs; a count from a page that does not cannot, and a difference there may be nothing but which fetch we got. So every sentence we have published about one of these pages changing, or holding, is about the three pages on the left of this split — and the two on the right are the pages we have always refused to characterise, with this as the reason.
← Everything we have published, including the runs with a buyer in them