Wikipedia Article Approval Rates: What 1,000+ AfC Submissions Reveal

We provide a high-level summary of how Wikipedia's Articles for creation (AfC) workflow operates, why it's so important, and what our analysis of 1,009 submissions uncovered. 

Wikipedia terminology can be difficult to parse, especially as there are numerous closely related terms with subtle differences in meaning. We offer definitions for key terms and provide examples of how they're deployed by editors.

We summarize why there is such a large demand for new Wikipedia articles and how vendors attempt to satisfy that market need. Then we explain our own role in that ecosystem and what we wanted to know about the AfC process.

This section is split into two: (1) Wikipedia’s methodology for reviewing AfC submissions, and (2) Our methodology for compiling a dataset of submissions and assessing relevant datapoints.

How we compiled and assessed AfC submissions

To ensure we captured only viable submissions, we pulled our entries from the "1 day ago" category of the AfC pending submissions by age category page. We determined that if a submission survived the first 24 hours, it had typically passed the quick-fail assessment and was therefore deemed sufficiently viable to warrant a standard content review. 

We pulled "1 day ago" submissions from two date ranges in 2025—September 22 to September 28 and October 25 to October 31—for a total of 1,009 submissions. Pulling in two datasets separated by a month ensured that we captured a range of submissions (since some submissions come in groups linked to the same core subject) and a range of reviewing editors (as some reviewing editors are especially active over short periods of time). We believe these 1,009 submissions constitute a representative, methodologically sound sample. 

We assessed and coded the drafts using the following  eight categories: 

1. Submission date

   • The specific date when the draft was submitted.

2. Submission outcome 

   • This assessment captures the action taken on the specific submission,not the outcome of the draft to date. Some of the declined submissions we assessed were later improved, resubmitted, and approved. And, conversely, some of the approved submissions were later deleted from the mainspace. But we only considered the specific action (approval, declination, deletion, or other) that occurred to the time-dated submission under review. 

     Approved meant that a reviewing editor approved the submission by moving it to the mainspace. 

     Declined meant that a reviewing editor used a draft decline template to indicate the submission wasn't ready for Wikipedia. 

     Deleted meant that the entire submission draft was deleted from Wikipedia, usually because of copyright violations or "unambiguous advertising or promotion." In these instances, reviewing editors decided that Wikipedia should no longer host the content, even in draft form.

     We marked drafts as Other when it was not possible to determine the action taken on the submission or the action taken was not applicable—e.g., the AfC submission template was removed from the draft before a review could proceed.

3. Reason for decline (if applicable)

   • If a draft was declined, we noted the main decline reason provided by the reviewing editor (e.g. "This draft's references do not show that the subject qualifies for a Wikipedia article"). In some instances, a secondary reason was also provided, but we only captured the first reason.

4. Reviewing editor

   • The reviewing editor who approved or declined the submission. 

5. Approval or rejection date

   • The date on which the approval or rejection was made. 

6. Resubmission status

   • We noted (yes or no) if a draft had previously been submitted and rejected. We only looked at previous AfC submissions. In some cases the draft had previously been published and then reverted to draft status. We did not count that as a previous submission, since it wasn't done through AfC. 

7. Draft category

   • There are two main ways to categorize: 

     Sorting by primary characteristic (taxonomy / classification)

     This method involves placing an item into a single, predetermined, and often hierarchical group based on its most dominant or defining trait.

     Tagging all relevant attributes (faceted classification / tagging)

     This approach involves assigning multiple, descriptive labels (tags) to an item, acknowledging that an item can belong to several categories at once.

     We went with the primary characteristic approach because that's generally (although not always) how Wikipedia handles articles. This Project page shows how AfC submissions are sorted into different categories by a bot.

     We classified drafts into five main subjects (Arts & Culture, Business, Person, Organization, and Other). Our subject categories do not entirely align with Wikipedia's categories, most notably because we opted to split businesses and organizations. 

     Even though businesses and organizations fall under the same notability guidelines (WP:NCORP), in our experience editors assess for-profit businesses differently from other organizations, especially as "organization" is a broad catch-all for everything from religious groups to armed forces to environmental non-profits. 

     We also created subcategories for each of the five main categories. We discuss these subcategories in greater detail in the Findings section. 

     Wikipedia is a general-interest encyclopedia that includes entries about nearly every possible subject. Trying to accurately sort draft submissions using consistently applied criteria proved difficult, and some of the subcategory parameters are fuzzy by necessity.

     We know there’s significant interest in how Wikipedia handles drafts related to businesses and persons, and those categories (and associated subcategories) were thankfully easier to define and categorize. 

8. Evidence of LLM usage

   • This was a yes/no assessment of whether editors had left comments indicating concerns about LLM/AI usage. 

     While LLM tools are not strictly banned for all Wikipedia tasks, editors are forbidden from using them to generate new, unverified articles from scratch. Given this restriction, we wanted to know if reviewing editor feedback, either in the form of a decline notification or a specific draft comment, indicated suspicion of LLM usage (or AI usage more generally).

     For this category, we did consider feedback on previous submissions of the same draft, but we ignored feedback that occurred after the specific submission was approved or declined.

We summarize our findings across data analysis categories: Overall approval rate, Reason for decline, Approval rates by category, Reviewing editors, Time to review, Resubmission status, and LLM usage.

We provide a deep dive into two core areas that stood out in the data: the widespread use of LLM tools, a new problem facing Wikipedia, and the subjective nature of notability, a contentious subject since the encyclopedia's launch in 2001.

Articles for creation serves as a human check on AI-generated hallucinations and prose, and on low-quality, poorly sourced content more generally. But like all human decision-making, there is an interpretive element to the AfC assessments—especially regarding notability—that leads to unpredictable approval decisions.