"Show me a random one" is a few lines of code. Take a list, pick an index, return it. Any developer can build a random link button in an afternoon, and a great many have, and several of the results are genuinely good things that people use happily for years.
For the broader context behind this business guide, read How The Stumble Picks What You See.
That is worth saying at the start because the argument here is not that shuffling is a poor idea. It is that the shuffle is not where the value sits, and a tool that consists only of a shuffle inherits whatever was in the list it was handed.
What this is, and who it is for
For anyone who has just been told there is another random link button and would reasonably like to know what makes this one different from the ones they already have bookmarked.
The honest answer is a boring one: the randomness here is not different at all. It is a uniform draw, the same as everyone else's, and it is meant to be. What differs is the list, and the list is the expensive part.
Scope: the register and this article specify a process, not a deployed populated surface or selection result.
What a random link button actually does
It picks uniformly from an index it did not compile. That means the tool's entire character is inherited from wherever the index came from, and index provenance is usually one of three things.
- A crawl. Broad, cheap, and includes whatever the crawler found, which is a great deal of parked domains, abandoned projects and pages that were never for anyone.
- A submission queue. Everything anyone bothered to add, which skews towards people who promote things.
- A snapshot of a curated list from somewhere else. Good on the day it was taken, and progressively less true afterwards, because nobody is re-checking a list they did not compile.
None of these is dishonest. They are just cheap, and cheapness at the intake stage is what the shuffle then samples from. A shuffle cannot improve a list. It can only present it in a different order.
Why a shuffle stops being interesting quickly
The failure mode of an unfiltered shuffle is not that it shows you bad things. It is that it shows you things that are not for anybody.
Press it a few times and the pattern usually appears. A domain that has lapsed. A launch page for a product that never launched. A blog that stops abruptly a few years ago. A page that loads a spinner and nothing else. Each individual result is unremarkable and the cumulative effect is that pressing the button stops feeling like discovery and starts feeling like sampling.
That is a liveness problem rather than a taste problem, and it is not solvable in the draw. There is no clever weighting that turns a stale index into a fresh one. The only fix is upstream, which means someone has to look at things, and looking at things is slow.
What a human adds before anything reaches the shuffle
Everything that makes an entry eligible here happens before the draw, and every step of it involves a person. The bar every candidate cleared sets out the criteria in full. What the work consists of, in order:
- Confirming it is a real thing you can act on today. Not a waitlist, not a landing page, not a coming-soon, not an idea with a mailing list attached.
- Confirming independent ownership as a checked fact rather than an assumption, which is often the slowest single step.
- Ruling out the vehicles. Not primarily an advertising, affiliate, dropship, lead-generation or content-farm site, several of which look like small independent businesses at a glance and are structurally something else.
- Actually examining it. A named reviewer uses or examines the thing and has to be able to describe it in our own words, which is a much higher bar than reading the homepage, because you cannot paraphrase what you have not understood.
- Running a dated liveness check by the published method, producing a date and a row in the check log.
- Writing the description. In our own words, without repeating the operator's marketing claims except as quoted and attributed material.
- Re-checking it, forever. This is the part that never finishes and the part that most lists skip.
Then, and only then, the entry becomes one item in a pool that gets sampled at random with no preference whatsoever.
The refusal rate as the measure of curation
The number that tells you whether a curated list is curated is not how many entries it holds. It is what share of candidates were turned away, and that number only means anything if it comes from a log rather than from a feeling.
This edition supplies no review log, so the refusal rate is not measured. The refusal file defines the calculation and anonymisation rules without presenting a projected value.
We state the absence rather than estimate, because an estimated refusal rate is exactly the kind of number a site like this should never publish. What the bar turns down explains how patterns and shares must be anonymised. For this edition, the answer to "how selective are you" is not measured.
What curation costs, per entry
Per entry, the cost is a person's attention twice: once at review and once at every re-check, for as long as the entry exists.
The second half is the part that makes curated registers rare. Intake is a one-off cost with an obvious payoff, and everyone budgets for it. Re-checking is a recurring cost with no visible payoff at all, because a successful re-check produces no new content, no new page and nothing to announce. It just moves a date. The cost of intake follows that asymmetry through the economics of listing sites generally, and it explains most of what is wrong with the category.
The structural consequence is that this register will be small. Not small as a launch phase, small as a permanent property. A list that one set of people can keep true has a size limit, and pretending otherwise is how lists become graveyards. Why every revival stays small makes the same point about discovery tools generally, and concludes that staying small is the feature rather than the constraint.
Where randomness is still doing the work
Having spent this long defending curation, it is worth being clear about how much of it stops at the pool boundary.
Inside the eligible pool, the draw is uniform and takes no input from anything. Not from us, not from the business, not from you. There is no editor's pick, no rotation of favourites, no "we thought you might like", no ordering by anything. Two people drawing at the same moment are running the same coin toss over the same pool.
That is deliberate and it is the other half of the argument. The curation is a filter, not a ranking. A filter answers in or out against a published bar. A ranking answers "which of these deserves more attention", and once that question is being answered by a tool, its answer becomes a position, and positions get sold. The rotation rules set out why the draw is kept deliberately dumb.
So: a human decides what is eligible, and nothing decides what you see.
What you lose by curating
A real list, because the trade is real.
- Breadth. A crawled index will always be far larger. Most of what exists on the web will never be in a reviewed register, including a lot of it that is good.
- Surprise from outside our taste. This is the serious one. Every entry here was examined by a person, and people have preferences they cannot fully see. The set of things we recognise as worth listing is not the set of things that are worth listing. A shuffle over a crawl has no taste at all, and that is genuinely a feature: it can hand you something nobody would have chosen.
- Speed. An index can grow by thousands overnight. This grows at the rate a person can look at things.
- Whole categories. Several verticals are excluded outright at our entry bar, which means good businesses in those areas are absent for reasons that have nothing to do with them.
- Recency. Something launched this week is not here, because nobody has looked at it yet.
The claim being made is narrow, and it is worth stating exactly. Curation here does not produce better results than a shuffle over a crawl. It produces checked ones: things that a named person confirmed were real and trading on a stated date. Whether checked is what you want depends entirely on why you are pressing the button. If you want the strangest corners of the web, an unfiltered shuffle is the better tool and the independent web points at several places to find them.
In Plain English
What it means. The random part of this tool is ordinary and deliberately so. The work is in deciding what is allowed into the pool, which is done by a person, one entry at a time, and repeated forever.
Who it helps. Anyone deciding whether a discovery tool is worth using, and anyone considering building one who has not yet costed the re-checking.
When to use it. When someone describes a tool as curated, ask what share of candidates were refused and where that number came from.
What to do next. Read our entry bar for what an entry has to survive, and the refusal file for the rejection taxonomy and evidence rules.
Common questions
Is the randomness weighted at all?
No. Uniform across the eligible pool, with nothing about the business, the reader or us entering the calculation.
Does curating mean you think the things you list are good?
No, and this is the distinction the whole site rests on. The bar answers whether something is real, independent, working and describable. It does not answer whether it is any good, and nothing here is ever rated, scored or ranked.
If curation is so expensive, why not automate the checks?
Parts of the liveness check are automated and always will be, because a machine is better at noticing that an address stopped resolving. What cannot be automated is deciding that something meets the bar, and that decision has a named person against it every time.
Apply the distinction
The stumble specification shows the selection protocol. Its claim is about uniform selection from an eligible set, not about the size or availability of a live register.