The word "directory" now mostly means a page full of listings that somebody paid to appear on. That is a late meaning. The original idea was different, more ambitious, and worth understanding properly, partly because it explains what went wrong, and partly because anyone building a curated collection today is signing up for the same four problems.
For the broader context behind this business guide, read What We Lost When The Blogroll Died.
Editorial note. Every claim here about how the model worked is either quoted from a named source with a retrieval date, listed at the end and at our source ledger, or written so that it does not rest on a figure. Claims about submission volumes, waiting times and numbers of editors circulate widely and we could not source one, so no such figure appears in this piece.
What this is, and who it is for
For anyone who has heard that directories used to be good and wants to know what that actually meant, and for anyone who is building a curated collection and would like to know in advance which walls they are going to hit.
The account below tries to avoid two easy framings. The editorial directory was not a golden age that was taken away, and it was not a naive failure that deserved to lose. It was a set of tradeoffs that worked well under some conditions and stopped working when those conditions changed.
The wider survey of discovery mechanisms is at the wider history.
The editorial premise: a human decides
The premise was simple enough to state in one sentence. Before a site appears in the collection, a person looks at it and decides whether it belongs.
That single rule produces every property the model had, good and bad.
Because a person decided, inclusion carried information. A listing meant that somebody had formed a view, which is different from a listing that means a crawler found a page.
Because a person decided, the collection could apply criteria a machine could not compute — whether a site was substantive, whether it was what it claimed to be, whether it added anything to the category it was asking to join.
Because a person decided, there was somebody to appeal to. A rejected submitter had a decision to argue with, made by a person, rather than an outcome produced by a system with no interior.
And because a person decided, the throughput of the entire collection was bounded by human attention, which is the source of every failure below.
How submission and review actually worked
The mechanics below are not reconstructed from memory. The model's own documentation is still published, because the Open Directory Project's successor is still running. Curlie's about page, retrieved on 6 September 2026, states that it "strives to be the largest human-edited directory of the Web", that it "is run by volunteer editors", and that "We started as the Open Directory Project (ODP), later became DMOZ, and in 2017, we launched Curlie to continue the 100% free directory." We could not open archived captures of the original Open Directory documentation, so the account below is taken from that successor's live submission instructions and editorial guidelines on that date. Read it as documentation of the model rather than as a record of any particular year.
A site owner chooses a category and submits their site to it. The instructions say to search the directory first to make sure the site is not already listed, then to "Submit the site to the single, most relevant, category", and note that "a category without the icon is not accepting suggestions". The submission joins what the editorial guidelines call a "category's unreviewed queue", and an editor responsible for that part of the tree reviews it: "An editor will review your suggestion to determine whether to include it in the Directory."
Three outcomes are documented. Acceptance. Relocation: the guidelines on subcategories instruct editors that links in general or catch-all categories "should be placed at the top of the most appropriate topical category or moved to an appropriate topical subcategory". And refusal, which the instructions state plainly — "We don't accept all sites, so please don't take it personally should a site not be accepted" — and which the site selection criteria restate as "no site is guaranteed a listing in Curlie, and we depend on editors to use their own discretion".
The listing itself belongs to the collection rather than the submitter. The editing style guidelines say that "directory annotations (i.e. titles and descriptions) should objectively describe a site's subject and/or contents", that a title "should identify the site, not describe it", and give as their intent advising editors "to avoid flagrant and egregious marketing hype, sales pitches, subjective commentary, excessive and unnecessary repetition, and other superfluous information". A submitter who is "dissatisfied with how the site is described or titled" is told to flag the listing, not to rewrite it.
Three features of that process are worth pulling out, because they are the ones later listing sites dropped.
The description belonged to the collection, not the submitter. That single decision is what stops a directory reading like a wall of advertisements, and it is the most commonly abandoned part of the model.
Category placement was the editor's call. Submitters routinely proposed a category more prominent than the one their site belonged in, and the editor's correction was where most of the taxonomy's quality came from.
Rejection was a normal outcome. A queue with no rejections is an intake form.
What it got right
Four things, and each of them is still correct.
Inclusion as a signal. When entry requires a decision, being in the collection tells a reader something. The moment entry is automatic, the collection can only tell you a thing exists.
Descriptions in a consistent voice. A reader can compare entries when the entries are written by the same hand to the same standard. They cannot compare marketing copy.
A visible structure. A browsable tree lets someone explore a subject they cannot yet name, which is the exact function that query-based tools cannot perform.
A named place to complain. Human decisions can be reviewed by humans. This is a real property and its absence in automated systems is felt constantly by anyone who has tried to correct one.
Failure 1: the queue
The submission queue is the failure most directly relevant to anybody doing this today, so it gets the most space.
The structure of the problem: submissions arrive at a rate set by everybody outside the collection, and are processed at a rate set by the people inside it. Those two rates have no relationship to each other. Nothing about receiving more submissions creates more reviewing capacity.
Once the arrival rate exceeds the processing rate for a sustained period, four things follow, and they follow mechanically.
The backlog grows without bound. Not "gets long" — grows, continuously, because the deficit compounds every cycle.
Waiting time becomes unpredictable. A submitter cannot tell whether they are waiting or have been forgotten. The documentation says as much without giving a number. Curlie's submission instructions, retrieved on 6 September 2026, tell a submitter only that "An editor will review your suggestion to determine whether to include it in the Directory" and that "Depending on factors such as the volume of suggestions to the particular category, response time can vary dramatically." Note what that concedes: the wait is a function of how much arrives in one category, which is exactly the mechanism above. Specific figures for typical waiting times circulate widely and we found none in any documentation we could open, so this piece carries none.
Pressure appears to bypass the queue. Where a queue is long and inclusion has value, some parties will look for a faster route through it. Any expedited path, whatever its intention, changes what a listing means, because a reader can no longer tell whether an entry was judged or accelerated. This is the mechanism, not an allegation about anybody.
The queue becomes the collection's actual editorial policy. Whatever gets reviewed is what gets in, so the order and depth of the queue quietly determine the shape of the collection, regardless of what the stated criteria say.
The only structural defences are unattractive: close submissions, cap intake per period, or narrow scope so far that arrivals stay below capacity. All three limit the collection's growth, which is why they are rarely chosen.
Failure 2: the category tree
A hierarchical taxonomy is excellent when the domain is small and stable, and degrades in specific ways as it grows.
Things belong in several places. The more a field matures, the more entries sit legitimately across two or three branches. Cross-listing multiplies the maintenance obligation; single placement makes the entry unfindable to anyone browsing the other branch.
The tree records the world at the moment it was designed. New categories of activity appear and have nowhere to sit. Retrofitting them means either an awkward placement or a restructure, and a restructure invalidates every existing path.
Depth trades against discoverability. A shallow tree has enormous unusable categories. A deep tree has precise categories nobody can find without already knowing the vocabulary.
The taxonomy encodes a worldview. Where a subject sits in a hierarchy is an editorial claim about what it is related to, and reasonable people will disagree. That is unavoidable in any taxonomy and is worth stating openly rather than presenting the tree as neutral.
Failure 3: the editors, and who they were
The model depended on people who knew a subject well enough to judge it and were willing to do the work for reasons other than money.
That produces a structural weakness with no villains in it. Editorial attention distributes unevenly across a large tree: some branches attract people who care deeply and are maintained carefully, and others attract nobody and become decorative. A reader browsing has no way to tell which kind of branch they are in.
The role also has no redundancy and no succession. When the person maintaining a branch stops, the branch does not fail loudly. It simply stops being updated, and it keeps looking exactly the same as it did when it was current.
And the position carries influence: deciding what enters a widely-used collection is a form of power, held by people who did not apply for it in any formal sense. We are not going to characterise individual editors of any historical collection, which would be both unfair and unsourceable. The structural point stands on its own — any system that depends on unpaid judgement inherits the distribution of who is willing and able to give it.
Failure 4: scale versus judgement
The last failure is the one that ends the model, and it is arithmetic.
Judgement costs time per item. The number of items grew. The number of hours in a volunteer's week did not.
So the collection faces a choice it cannot avoid. Keep the judgement and cover a shrinking fraction of the field. Or drop the judgement and keep coverage, at which point the listing no longer means anything was decided. There is no third option available by working harder, and none available through better software, because the expensive part is reading the thing rather than fetching it.
The economics underneath this — why intake gets funded and accuracy does not — are in the economics that killed them. The broader business models of the current listing category are at how directories work.
Which of these we have not solved
The point of this piece is not to explain why we are different. Three of the four failures apply to us in full.
The queue is unsolved. We accept submissions and review them by hand. If submissions exceed our reviewing capacity, our backlog will grow exactly as described above. Our only defences are the unattractive ones: a narrow scope, a cap on how many new entries we publish in a period, and a refusal to offer any faster route for any reason, including payment. We publish the cap so that a submitter can see the constraint rather than guess at it. That makes the queue honest. It does not make it short.
The reviewer problem is unsolved. Our entries name the person who reviewed them, which gives a reader someone to hold responsible, and does nothing about the underlying dependency on a small number of people's attention and taste. If those people stop, this collection stops being maintained the same week.
Scale versus judgement is unsolved and unsolvable. We chose judgement, which means the register will cover a rounding error of the businesses that would qualify. We would rather say that plainly than describe ourselves as covering a category.
The taxonomy problem is one we have only deferred. Our categories are few and our three-entry gate keeps thin hubs from publishing, which avoids the worst of the tree problem by keeping the tree tiny. It does not solve it. The first time an entry legitimately belongs in two categories, we will meet the same tradeoff everyone else met.
The one thing we do that the old model did not is decay: entries carry a check date, get re-checked on a published cadence, and become tombstones rather than silently persisting. That addresses rot. It addresses none of the four failures above. What our bar does differently sets out the inclusion rules in full, and the rest of this cluster is at the independent web.
In Plain English
The original directory idea was that a person looks at a site before it gets listed. That one rule is where all the value came from: a listing meant somebody had decided, descriptions were written in the collection's own voice rather than the submitter's, and a rejected applicant had a human to argue with.
It broke in four ways. Submissions arrived faster than anyone could review them, and the backlog set the editorial policy in practice. The category tree stopped fitting a world that kept producing new kinds of thing. Editorial attention was distributed unevenly and had no succession plan, so branches went quiet without looking different. And judgement costs time per item while the number of items kept growing, which forces a choice between meaning and coverage.
Three of those four are ours too. We have not solved the queue, we depend on a small number of people, and we chose judgement over coverage knowingly. The tree problem we have only postponed by keeping the tree small.
Sources
Each entry below is recorded in the source ledger under the identifier shown.
- S-020 · Curlie, about — About Curlie, Curlie. https://curlie.org/docs/en/about.html Retrieved 2026-09-06. Supports: that Curlie describes itself as a human-edited directory run by volunteer editors, and its own statement that it began as the Open Directory Project, became DMOZ, and launched as Curlie in 2017.
- S-024 · Curlie, submission instructions — Suggest a Site: Instructions, Curlie. https://curlie.org/docs/en/add.html Retrieved 2026-09-06. Supports: searching first, submitting to the single most relevant category, categories that are not accepting suggestions, that an editor reviews the suggestion, that response time varies with the volume of suggestions to a category, that not all sites are accepted, and that a submitter dissatisfied with the title or description flags the listing rather than rewriting it.
- S-022 · Curlie, site selection criteria — Site Selection Criteria, Curlie Editorial Guidelines. https://curlie.org/docs/en/guidelines/include.html Retrieved 2026-09-06. Supports: that no site is guaranteed a listing and that editors use their own discretion, and the existence of a per-category unreviewed queue.
- S-021 · Curlie, editing style — Editing Style Guidelines, Curlie Editorial Guidelines. https://curlie.org/docs/en/guidelines/describing.html Retrieved 2026-09-06. Supports: that titles and descriptions are written by editors to describe a site objectively, that a title should identify rather than describe a site, and the stated intent of avoiding marketing hype and sales pitches.
- S-023 · Curlie, subcategories — Subcategories, Curlie Editorial Guidelines. https://curlie.org/docs/en/guidelines/subcategories.html Retrieved 2026-09-06. Supports: that editors move links out of general or catch-all categories into the most appropriate topical category or subcategory.
Read our entry bar: what gets in, what is excluded automatically, and the cap on how fast we add anything.