Setting up a shop online has never been easier or cheaper. Being found once you have done it has got harder. Those two facts sit awkwardly together, and most explanations of the second one collapse into blaming a single company, which does not survive contact with the details.
This piece takes the paradox apart into four causes. None of them requires anybody to have decided anything, which is what makes the pattern so persistent.
Editorial note. This piece is argument rather than measurement. Where a figure would ordinarily carry the point, we have written the mechanism instead and said plainly what a number would be needed for, because we would rather publish a mechanism a reader can check than a statistic we cannot source. Our sourcing rules are at our source ledger.
What this is, and who it is for
For two readers. One is trying to buy from, hire or use something small and specific, and keeps landing on the same aggregators. The other has built something small and specific and cannot understand why the internet appears not to contain it.
The analysis below is our reading of how the mechanisms interact. It is inference from things anybody can observe rather than a measurement of anything, and where it is inference this piece says so in the sentence rather than in a footnote.
The paradox, stated precisely
State it carefully, because a loose version of it is easy to refute.
The cost of publishing has fallen. Hosting, a storefront, payment handling and a usable design are now bought the way a phone contract is bought: from off-the-shelf services, on published self-serve pricing, without writing code and without anybody's permission. We are deliberately not quoting a price, because any figure we printed would be wrong within a year and would differ by country anyway. The check is one a reader can run in five minutes: open the pricing page of any hosting provider, any hosted storefront and any payment processor, and note that the price is published, that the plan can be started without a salesperson, and that nothing on the page requires technical skill. What matters for the argument is not the amount but that the gate is gone.
The cost of being encountered has risen. Not the cost of being reachable — a person who knows your name will find you in seconds. The cost of being encountered by someone who did not know you existed.
So the tail is longer and denser than it has ever been, and simultaneously less traversable. Both halves are true at once, and any explanation that denies either half is wrong.
One honest counterweight before we go further: for a person who knows what they want and can name it, the modern web is dramatically better than what preceded it. Naming a specific maker and reaching them is close to instant. The failure is confined to encounter without a name, which is a narrower claim than "everything is worse" and a more defensible one.
Cause 1: the cost of being visible rose faster than the cost of being online
Publishing and visibility are two different products with two different price curves, and they have moved in opposite directions.
Publishing became a commodity, so its price fell toward the cost of the underlying infrastructure. Visibility, by contrast, is positional. There is a fixed amount of attention on any given query or surface, and the number of parties competing for it has grown along with the ease of publishing. A positional good with rising demand and fixed supply gets more expensive. That is not a market failure; it is what positional goods do.
The consequence for a small operator is specific. The work required to be encountered — content production at volume, paid placement, sustained presence on several surfaces — has a floor that does not shrink with the size of your business. A one-person workshop faces roughly the same visibility cost structure as a company with a marketing department, while having a fraction of the capacity to pay it.
This is the cause with the least controversy attached and the most consequences flowing from it.
Cause 2: the intermediaries changed what they optimise for
Between a small operator and a potential customer there is almost always an intermediary: a search engine, a marketplace, a social platform, a maps product, a review site. Each of these optimises for something legitimate, and none of those things is "surfacing the least-known adequate option".
- A search engine optimises for answering the stated query well. A small unknown supplier is rarely the best answer to a question phrased in general terms, because it has less evidence attached to it in every direction.
- A marketplace optimises for a completed transaction inside the marketplace, which favours sellers whose listings convert reliably.
- A social platform optimises for held attention, which favours whoever produces frequently and in the platform's preferred format.
- A local platform optimises for a usable set of nearby options, which requires a threshold of data about each option that a small business may never generate.
Notice that every one of those optimisation targets is defensible and that none of them is hostile to small operators. The exclusion is a side effect of the target, not a policy about the excluded. Attributing intent here is both unfair and analytically useless: change the company and keep the target, and the same outcome appears. We take this apart in more detail in why every result looks the same.
There is a second-order effect worth naming. Because these intermediaries need evidence about an option in order to rank or display it, the businesses that generate the most evidence — reviews, mentions, structured data, transaction volume — become the most visible. Evidence generation correlates strongly with size. So the ranking is not of quality; it is of legibility, and small things are systematically less legible.
Cause 3: the curation layer was never funded and quietly stopped
For a period, a meaningful share of the encounters described above did not run through an intermediary at all. They ran through people: link lists, rings, editorial directories, forum recommendations. That layer had a specific property the intermediaries lack — it pointed away from itself, at things that had no other route to an audience.
It also had no revenue model. The work was recurring, unpaid, and invisible when done well. What the blogroll used to do is the long version of this, and the economics of not pruning explains why the maintenance half specifically is what fails.
The important part for this argument is what its absence removed. It did not remove a nice thing. It removed the only discovery route that did not require the discovered party to be commercially legible. If you are small and specific and generate little evidence, the human recommender was the mechanism that worked for you, and it worked precisely because a person could see quality that a system had no data about.
Cause 4: volume made human review economically impossible
The final cause is arithmetic, and it is the one that closes off the obvious remedy.
If the answer to intermediary bias is human curation, then somebody has to look at things. Looking at things costs time per thing. The number of things has grown enormously, and the time available per reviewer has not changed at all.
This produces a hard ceiling. A human-reviewed collection can be accurate or it can be comprehensive. It cannot be both, and the gap between the two widens every year as the population of candidate entries grows. Any collection that chooses comprehensiveness has to drop review, at which point it becomes an index with a curation-flavoured description. Any collection that keeps review has to accept being a rounding error against the size of the category.
That is not a temporary condition awaiting a technical fix. Automation can widen the funnel — it can check whether a URL responds, whether a page changed, whether a business appears to have stopped trading — but it cannot make the judgement that something is a real independent operation rather than a shell, because that judgement rests on reading the thing.
What this means for someone actually trying to buy small
Four practical consequences, stated without a workaround that pretends to fix the structure.
Absence from a surface is not evidence of absence. The small supplier you are looking for may be entirely real and entirely missing from every list you consulted, because they never generated the evidence those lists need.
Ranking position tells you about legibility, not quality. A first result is the option about which the most is known and can be computed, which is a genuinely useful signal for some purchases and close to useless for others.
Query-shaped searching finds named things. If you can name it, search is excellent. If you cannot, you need a route that does not start with a name: a person who knows the field, a community, a curated collection, a market stall.
The routes that still work are mostly human. They are slower, they are partial, and they are not organised for your convenience. Where people still find independent work is a list of them, most of which have nothing to do with us.
More on how the intermediary layer makes its money, and how to read a listing sceptically, is at how directories work.
What would have to change
Three things, in ascending order of difficulty.
Someone would have to fund accuracy. Not intake. Accuracy — the recurring work of re-checking that a listed thing still exists and is still what the listing says. The obstacle is a cost shape rather than a shortage of goodwill. Intake is a one-off cost per entry that leaves behind a permanent, countable asset. Accuracy is a cost per entry per cycle that produces no visible change when it succeeds, because an entry that was checked and is fine looks exactly like an entry nobody looked at. The revenue in this category attaches to entries existing, not to entries being true, so re-checking has to come out of somebody's margin and is the first thing cut when the margin tightens. The economics of not pruning runs that arithmetic through each of the usual funding models. We are not asserting that nobody anywhere pays for this work; we are saying that a collection which does is paying for it against its own incentives, and that is a fair thing to ask any collection about.
Discovery surfaces would need a use for illegible things. As long as a surface requires evidence in order to display an option, small operators will be under-represented on it, no matter how well-intentioned the operator of the surface is. A route that works differently has to be built differently, and it will be much smaller.
Readers would need to accept partial collections. A curated collection is a sample of somebody's attention. It cannot be a category's coverage. If a reader treats partial as failure, every honest collection loses to a comprehensive dishonest one, and the incentive points the wrong way for everybody.
The register specification requires a named reviewer, a visible check date and a bar defined independently of any candidate. That is a model for partial, auditable curation rather than a claim of comprehensive coverage or a solution to the structural problem described here. The rest of this cluster is at the independent web.
In Plain English
It costs almost nothing to put a small business online, and a great deal to be encountered by someone who did not already know about you. Both things are true, and that is the paradox.
Four causes. Visibility is a positional good, so its price rises as publishing gets cheaper. Every intermediary between a seller and a buyer optimises for something reasonable that is not "show the least-known adequate option". The human curation layer that used to route around all of this was never funded and mostly stopped. And there are now far too many candidate businesses for human review to be comprehensive, which closes off the obvious fix.
None of that requires a villain. Keep every company exactly as it is and change nothing about anyone's motives, and the same outcome falls out of the structure.
What helps a buyer today is knowing that missing from a list is not the same as not existing, that ranking measures how much is known about a business rather than how good it is, and that the routes which still surface small things are mostly people rather than systems.
Browse the register to see what a small, dated, human-checked collection actually looks like, including how little is in it.