How Does Wikipedia Decide to Delete an Article?

Three routes out, and the one everybody talks about is not the biggest

Most writing about Wikipedia and companies stops at whether an article can be created. That is the wrong end of the process. Creation is cheap; survival is the hard part, and the machinery that removes articles is both better documented and more decisive than anything on the way in. This guide is about that machinery.

There are three ordinary routes by which an English Wikipedia article stops existing. Speedy deletion lets a single administrator delete a page immediately under a numbered criterion, on the reasoning that some pages have no practical chance of surviving a discussion. Only administrators can execute a deletion; any editor may tag a page requesting one. Proposed deletion, universally called PROD, is described by its own policy page as “a way to suggest an article or file for uncontroversial deletion”: the page is tagged for at least seven days, and if nobody objects an uninvolved administrator either deletes it or removes the tag. Articles for Deletion, or AfD, is the public discussion — normally at least seven days, then closed by an editor who assesses what the discussion established.

We wanted to know which of these does the work, so we pulled the article-namespace deletion log for a full week through the MediaWiki API and classified every entry by the reason its deleting administrator recorded.

Two things in that chart are worth pausing on. The first is that PROD, the quiet seven-day route with no discussion, deletes more articles than AfD does. The second is that A7 and G11 — the two criteria a promotional company article is most likely to hit — account for only 60 article-namespace deletions in a week between them. That is not because promotional material is rare. It is because most of it never reaches the article namespace.

The same week's draft-namespace log records 3,173 deletions. Of those, 1,429 (45.0%) cite G13, abandoned drafts and Articles for Creation submissions, and 1,132 (35.7%) cite G5, creations in violation of bans or blocks. Another 121 cite G11. The draft namespace is where undisclosed paid editing is caught and where most company drafts quietly expire, which is the enforcement context behind whether you can pay someone to edit Wikipedia.

Articles for Deletion, in practice

AfD is the route that matters when an article already exists and somebody wants it gone. The mechanics are plainly documented and worth knowing exactly.

Who nominates. Any autoconfirmed editor. The page is explicit that an account is required: “Unless you log in, you cannot nominate an article on your own.” Anyone else has to ask someone to do it for them. There is no requirement that a nominator be an administrator, an experienced editor or a disinterested one.

What they are supposed to do first. The checklist known as WP:BEFORE asks a nominator to read the relevant policies, check the article's history, confirm it is not a candidate for a speedier route, and search for sources — specifically “a normal Google search, a Google Books search, and a Google News search”. It then states the rule that decides a great many discussions: “If you find that adequate sources do appear to exist, the fact that they are not yet present in the article is not a proper basis for a deletion nomination.” Notability is a property of the available sources, not of the current text. Alternatives come first too: if an article “can be fixed through normal editing… it should only be proposed for merging or redirecting as it is not a candidate for deletion”. Notifying the creator is described as courteous rather than mandatory, with a caution against selectively notifying editors likely to agree with you.

How long. Articles are “normally discussed for at least seven days, after which a decision may be reached”, and may close earlier where consensus is already clear. If a discussion is too thin to close, it is relisted onto a later day's log. The deletion process page caps this: discussions “should not be relisted more than twice”, and a third relisting requires the relisting editor to explain why.

How thin is thin. Thinner than the word “discussion” suggests. Counting bolded responses in the wikitext of our 544 discussions gives a median of two and a mean of 2.5, and 181 of them — 33.3% — had one or none. A hundred and nineteen (21.9%) carried a relisting notice. This is the single most misunderstood feature of the process: the typical AfD is not a debate among dozens of editors but a nomination, one or two replies, and a close.

Who assesses consensus, and what that means

The first thing every Wikipedia page on the subject says is that this is not a vote. The Guide to deletion: “Wikipedia is not a democracy and majority voting is not the determining factor in whether a nomination succeeds or not.” The AfD page itself: “While AfD may look like a voting process, it does not operate like one”, and “Justification and evidence for a response carries far more weight than the response itself.” The deletion process page is the most operational: “consensus is not calculated by number of votes”, closers weigh “strength of argument and cited recorded consensus”, and “arguments that contradict policy, are based on unsubstantiated personal opinion rather than fact, or are logically fallacious are frequently discounted”. Closers “must use their best judgement, attempting to be as impartial as is possible for a fallible human, to determine when rough consensus has been reached”.

All of which places an unusual amount of weight on one person. So we counted who that person was.

Not every closer must be an administrator: non-administrators routinely close discussions whose outcome does not require the delete button, and an administrator performs the deletion itself. The practical reading for anyone whose article is at AfD is that the audience is small, experienced and reading for policy argument rather than enthusiasm. The AfD page's own advice is blunt about which of those two it rewards: “the closer is most likely to read your actual comment and assess it against policy”.

One outcome deserves separate mention because it is where a tenth of everything lands. A soft delete is what happens when a discussion attracts essentially no participation: the article is deleted, but as though it had been PRODded, and the deletion process page states it can be restored “for any reason on request”. Fifty of our 543 closes were soft deletes. If nobody defends an article, it goes — but reversibly.

Where a real discussion produces no rough consensus, the default is the other way. “The result of the discussion is to maintain the ‘status quo’, generally with the content being kept.” Five point seven per cent of our sample closed that way.

What happens to company articles specifically

AfD discussions are voluntarily sorted into topic lists by WikiProject Deletion sorting, and those notices are recorded in the discussion's own wikitext. That gave us a way to isolate the subset this guide's readers care about. Of the 544 discussions, 70 had been sorted into the Companies, Organizations, Products or Business lists.

A company article that reaches AfD is roughly two and a half times less likely to be kept than an article on any other subject nominated in the same week, and half again as likely to be deleted. That is the number to carry out of this guide. It is consistent with what we found in what sources Wikipedia company articles actually cite: the surviving corporate articles rest on a narrow band of sources that most companies do not have.

A worked example makes the shape concrete. Altimesh, a software company, was nominated on 1 September 2026 with a one-sentence rationale: “Delete for not passing WP:GNG. With a WP:BEFORE I could not find any proof of GNG or WP:NCORP.” It was closed as delete on 9 September. Appsbar, nominated the same week on the grounds “Cannot find any independent in-depth coverage. Website appears to be dead”, was closed on 12 September as a redirect to the article about its founder. Neither discussion ran to more than a handful of comments. That is what the process looks like from the inside: a short evidentiary claim about sources, briefly tested, and acted on.

The guidelines that get cited

We counted which policy shortcuts appeared in the text of the 70 company-sorted discussions.

The general notability guideline sets the baseline: “A topic is presumed to be suitable for a stand-alone article or list when it has received significant coverage in reliable sources that are independent of the subject.” Each italicised term is defined. “Significant coverage” means coverage that “addresses the topic directly and in detail, so that no original research is needed to extract the content” — which rules out a mention in a list, a quotation in someone else's story, and a funding-round paragraph. “Independent of the subject” “excludes works produced by the article's subject or someone affiliated with it”. And “presumed” is doing quiet work: it “means that significant coverage creates an assumption, not a guarantee, that a subject merits its own article”. Clearing GNG does not settle the question; it opens it.

NCORP is the guideline for organisations and companies, and it is deliberately stricter. It says so: it follows the general guideline “with a stronger emphasis on quality of the sources to prevent gaming of the rules by marketing and public relations professionals”. Four of its provisions decide most company discussions.

On independence, NCORP does not stop at the subject and its staff. “The author must be unrelated to the company, organization, or product. Related persons include organization's personnel, owners, investors, (sub)contractors, vendors, distributors, suppliers, other business partners and associates, customers, competitors, sponsors and sponsorees.” Read that list against a typical media plan. A customer case study fails. A partner's announcement fails. A profile arranged by an agency the company pays fails on the material rather than the byline.

On depth, “deep or significant coverage provides an overview, description, commentary, survey, study, discussion, analysis, or evaluation of the product, company, or organization”. Every word there describes something the writer did to the subject, not something the subject said.

On audience, there is a floor: “Attention solely from local media (e.g., the weekly newspaper for a small town), or media of limited interest and circulation, is not an indication of notability. At least one regional, statewide, provincial, national, or international source is necessary.”

And on trade press, a warning that surprises a lot of B2B companies: “Trade publications must be used with great care. While feature stories from leading trade magazines may be used where independence is clear, there is a presumption against the use of coverage in trade magazines to establish notability.” That presumption runs against the grain of most B2B PR strategy, and it sits oddly beside the reach advantage we documented in whether trade publications outperform national outlets. Both things are true: a trade feature can be the most commercially useful placement a company gets and still not help it survive an AfD.

Excluded outright as dependent coverage: press releases, advertising material, paid or sponsored articles, self-published material, patents, and “other works in which the company, corporation, organization, or group talks about itself”. That is the mechanism behind our finding in whether media coverage gets a company a Wikipedia article, and it is why the source-by-source verdicts collected in Wikipedia's perennial sources list matter more than the volume of clips.

Speedy deletion and PROD

Two of the numbered speedy criteria are aimed squarely at company pages. A7 covers “any article about a real person, individual animal, commercial or non-commercial organization… that does not indicate why its subject is important or significant”. Note the standard: an indication of importance, which is lower than notability and is assessed from the text rather than from the sources. G11 covers “pages that are exclusively promotional and would need to be fundamentally rewritten to serve as encyclopedia articles, rather than advertisements”. The word “exclusively” is the limit: a promotional tone is a cleanup problem, whereas a page with nothing in it but promotion is a deletion problem.

G4 is the one that catches people who try again. It covers “sufficiently identical copies, having any title, of a page deleted via its most recent deletion discussion”. Recreating the deleted text at a slightly different title does not evade it; the criterion says “any title” explicitly. G12 handles copyright infringement, and G5 deletes creations made in violation of a ban or block, which is the criterion used against sockpuppet accounts.

PROD sits between the speedy criteria and AfD, and its rules are short. It “must only be used if no opposition to the deletion is expected”. Any editor — including the article's creator — “may object to the deletion by simply removing the tag; this action permanently cancels the proposed deletion via PROD”. And it is single-use: “PROD is one-shot only: It must not be used for pages PRODed before or previously discussed at AfD or FfD.”

That combination is worth understanding if your article is tagged. Removing a PROD tag is permitted, takes one edit, and cannot be undone through PROD; the nominator's remaining option is AfD, which is public. A company that removes a PROD tag without also fixing the sourcing problem has usually bought itself a discussion it is, on our numbers, likely to lose.

After deletion: history, salting, and coming back

Deletion is not erasure. The deletion policy states that it “removes the current version and all previous versions from public view”, and administrators retain access to the deleted page and its full edit history unless the content has been oversighted, a stronger form of removal used for material such as defamation or personal information. Readers can ask an administrator for a copy, or request a temporary undeletion for review, though “these requests are likely to be denied if the content has been deleted on legal grounds”.

The practical consequence is one that clients rarely anticipate. If a paid or undisclosed editor wrote a company's article and it was deleted, the diffs, usernames and edit summaries are all still there for administrators to read, and they are read when the same company appears again. Deletion closes the article, not the file.

Coming back has three doors. Where the deletion was a soft delete or a PROD, Requests for undeletion restores it on request. Where there was a discussion, deletion review is the appeal, and its scope is narrow: it is for cases where someone believes the closer read the consensus incorrectly, where a speedy deletion fell outside its criteria, where there was a substantial procedural error, or where “significant new information has come to light since a deletion that would justify undeleting”. It is explicitly not for arguing “technicalities (such as a deletion discussion being closed ten minutes early)”, not for a “disagreement with the deletion discussion's outcome that does not involve the closer's judgment”, and not “to repeat arguments already made in the deletion discussion”. It runs at least seven days.

The third door is a new article built on new sources, which is almost always the right one. The exception is when the title has been salted. Creation protection prevents new pages from being created at a specified title, and the protection policy names the case: “This type of protection is useful for pages that have been deleted but repeatedly recreated.” The term comes from salting the earth. A salted title can be unlocked, but the route is procedural rather than technical — ask the protecting administrator, request a reduction in protection level, or go through deletion review, and in each case present a draft that makes the case on its merits. Seven per cent of the nominations in our week were second or later nominations of the same article, so repeat attempts are common enough that this machinery gets used.

Can the subject say anything?

Yes, and this is where companies most often make things worse.

The Guide to deletion is permissive about participation as such: “The author of the article can make their case like everyone else.” It also offers a line that is easy to skip and worth reading twice: “Please do not take it personally if an article you've contributed to is nominated for deletion.”

The constraints come from the conflict-of-interest guideline. COI editors are “strongly discouraged from editing affected articles directly” and are asked to propose changes on talk pages instead. On disclosure the guideline is unambiguous: an editor with a conflict of interest “should always let other editors know about it, whenever and wherever you discuss the topic” — which covers a deletion discussion. On creation, “you should generally refrain from creating articles about yourself, or anyone you know, living or dead, unless through the Articles for Creation process”.

If money is involved, disclosure stops being a guideline and becomes a term of service. The Wikimedia Foundation Terms of Use require contributors to disclose “each and any employer, client, intended beneficiary and affiliation with respect to any contribution for which you receive, or expect to receive, compensation”, by at least one of three routes: a statement on the user page, a statement on the talk page accompanying the paid contributions, or a statement in the edit summary. Anyone advertising editing services off-wiki must also “disclose all Wikipedia accounts you have used or will use for this service in the public posting on the third-party service”. We treat that requirement at length in whether you can pay someone to edit Wikipedia.

The other failure mode is recruitment. The Guide to deletion names both forms: “A sockpuppet is an account created by a vandal or bad-faith contributor in an attempt to bias the deletion process. A close variation is the ‘meatpuppet’, people recruited from outside Wikipedia to try to alter the result of a discussion.” Because closers discount arguments not grounded in policy and are alert to accounts with no other history, a flurry of first-time accounts saying an article should be kept is worse than silence: it tells the closer what is going on, and it is exactly the pattern G5 exists to punish afterwards.

What actually helps, on the evidence of the discussions we read, is narrow and unglamorous. Post on the discussion page, disclose the connection in the first sentence, and do one thing: name specific sources, with publication, date and author, and say in a line each why they satisfy significant coverage independent of the subject. That is the argument the closer is reading for. Everything else — the company's size, its funding, its customer list, the unfairness of it all — is discounted by design.

For companies that reach this point and lose, the adjacent structured-data routes are a genuine fallback rather than a consolation, because they run on different rules: see whether Wikidata is easier to get into than Wikipedia and the broader reference-presence picture in Wikipedia and reference presence.

Methodology and limitations

All policy and guideline wording is quoted from the English Wikipedia pages listed in the sources below, each read on 26 September 2026. These are the operative documents: on Wikipedia the policy page is the primary source, and we have quoted rather than paraphrased wherever a definition does the work. The counts are ours. For the AfD sample we retrieved the templates transcluded on each of the seven daily logs for 1–7 September 2026 through the MediaWiki action=parse API, filtered to Articles for Deletion subpages, and fetched the wikitext of all 544 in batches through prop=revisions on 26 September 2026. Outcomes were parsed from the signed closing line (“The result was…”); 543 of 544 had closed and one had not, and seven closes used wording our parser could not map to a standard outcome and are reported as “other”. Topic subsets come from the WikiProject Deletion sorting notices recorded in each discussion. For the deletion-log figures we queried list=logevents for delete/delete actions in namespace 0 and namespace 118 over 8–15 September 2026 and classified the free-text reason by regular expression. Scripts were written for this report and the queries are reproducible from the API parameters stated in each chart source line. Six limitations matter. First, this is one week. AfD volume and the mix of subjects vary seasonally and with the activity of individual prolific nominators, and a single week cannot establish a trend. Second, deletion sorting is voluntary and incomplete, so the 70 company-sorted discussions are a sample of company AfDs and not a census; some company articles were certainly nominated that week without being sorted. Third, counting a policy shortcut is counting vocabulary, not persuasion: we cannot tell from the text which argument the closer actually relied on. Fourth, our count of responses per discussion matches bolded list items and will miss unusually formatted comments, so treat it as approximate. Fifth, 257 of the 1,917 article-namespace deletions carried a free-text reason our classifier could not map to a criterion, and housekeeping criteria such as G8, R2 and G6 inflate the total relative to deletions that reflect a judgement about a subject. Sixth, we found no peer-reviewed study of AfD outcomes for companies that we could read in full, and we are not going to cite one second-hand; the figures here are our own measurement, presented as such, and readers should reproduce them rather than trust them.

Conclusion

Wikipedia's deletion process is more legible than its creation process and much less merciful. An article can be gone in an hour under a speedy criterion, in seven silent days under PROD, or in a week-long discussion that two people read. Company articles do worst of all: in the week we counted, two-thirds of company, organisation and product nominations ended in deletion and one in fourteen was kept. The reason is not hostility, it is NCORP, which excludes as dependent almost every source a communications team can arrange and presumes against the trade press that a B2B company usually has most of. The three things worth knowing if you are on the receiving end: the argument that works is a short list of specific sources tested against “significant coverage, independent of the subject”; the conflict of interest must be declared before anything else is said, and must be declared as a matter of the Terms of Use if anyone is being paid; and the history does not disappear when the article does, so anything done badly the first time is still on file the second. If the sources that would satisfy NCORP do not exist yet, no amount of process knowledge will substitute for them, and the honest answer is to wait.

Frequently Asked Questions

Sources

1. primary data Wikimedia Foundation. MediaWiki Action API, English Wikipedia. Our own queries: action=parse on the seven Articles for Deletion daily logs for 1–7 September 2026; prop=revisions for the wikitext of all 544 transcluded discussions; list=logevents with leaction=delete/delete for namespaces 0 and 118 over 8–15 September 2026. Retrieved 26 September 2026.
2. policy document English Wikipedia. Wikipedia:Articles for deletion. Nomination eligibility, the seven-day period, the WP:BEFORE checklist, notification of the creator, and the statement that AfD “does not operate like” a vote. Read 26 September 2026.
3. policy document English Wikipedia. Wikipedia:Deletion process. “Consensus is not calculated by number of votes”; weighting of arguments; rough consensus; the two-relist limit; soft deletion and restoration on request; the no-consensus default to the status quo. Read 26 September 2026.
4. policy document English Wikipedia. Wikipedia:Deletion policy. Removal of all versions from public view, administrator access to deleted history, oversight, and routes to undeletion. Read 26 September 2026.
5. guideline English Wikipedia. Wikipedia:Notability (organizations and companies) (WP:NCORP). Definitions of deep coverage and independence, the list of related persons, the audience floor, the presumption against trade magazines, the list of dependent sources, and the “gaming of the rules” rationale. Read 26 September 2026.
6. guideline English Wikipedia. Wikipedia:Notability (WP:GNG). The general notability sentence and the definitions of significant coverage, reliable, sources, independent of the subject, and presumed. Read 26 September 2026.
7. policy document English Wikipedia. Wikipedia:Criteria for speedy deletion. Wording of A7, G4, G5, G6, G7, G8, G11, G12, G13 and R2, and the principle that only administrators may delete. Read 26 September 2026.
8. policy document English Wikipedia. Wikipedia:Proposed deletion. The uncontroversial-deletion purpose, the seven-day period, any editor's right to remove the tag, and the one-shot rule. Read 26 September 2026.
9. policy document English Wikipedia. Wikipedia:Deletion review. Grounds for review, the explicit exclusions, the seven-day minimum, and the pointer to Requests for undeletion. Read 26 September 2026.
10. policy document English Wikipedia. Wikipedia:Protection policy. Creation protection (“salting”), when it is applied, and how to have a salted title unlocked. Read 26 September 2026.
11. guideline English Wikipedia. Wikipedia:Conflict of interest. The strong discouragement of direct editing, the duty to declare a conflict wherever the topic is discussed, and the Articles for Creation route. Read 26 September 2026.
12. policy document English Wikipedia. Wikipedia:Guide to deletion. “Majority voting is not the determining factor”; the author may make their case; definitions of sockpuppet and meatpuppet. Read 26 September 2026.
13. terms of service Wikimedia Foundation. Terms of Use, paid-contribution disclosure. Requirement to disclose employer, client, intended beneficiary and affiliation, the three permitted disclosure locations, and the off-wiki advertising requirement. Read 26 September 2026.
14. worked example English Wikipedia. Wikipedia:Articles for deletion/Altimesh and Wikipedia:Articles for deletion/Appsbar. Two company discussions from the sample week, quoted for their nomination rationales and closes. Read 26 September 2026.