Request Consultation →
What public records show about disclosed conflicts in draft submissions and article edit requests.
Disclosure makes a contribution visible. The public record still has to show what happened next.
We traced two distinct paths taken by contributors who identified a conflict of interest or paid relationship when they entered a Wikipedia process: submitting a new draft through Articles for Creation (AfC), and asking for a change to an existing article. For each selected case, we linked the initial public revision to later review, page-move or article-edit evidence.
The result is a case-linked account of 40 draft episodes and 26 eligible edit-request discussions entered between 1 January and 30 April 2026. These are separate, observable samples. Their counts describe these cases only; they are not Wikipedia-wide approval rates or a comparison of the two workflows’ effectiveness.
AfC acceptance concerns a new article draft. Edit-request mapping records whether requested text appeared in an existing article.
Verified AfC acceptances by the observation cutoff.
Two further drafts reached mainspace through other moves and remain a separate category.At least one requested change matched an article revision at some point.
At least one mapped change was still present.
The draft sample had been through review by cutoff, but most latest observed AfC states were declines. In the edit-request sample, an article change sometimes preceded a reviewer’s reply, came from the requester’s own account, or was later removed. A Done or Answered marker was therefore never treated as proof that the proposed text appeared in an article.
Read the denominators narrowly. The draft archive covered only part of a historical AfC frame, and the request archive used queue, template and heading clues. Selection was reproducible within those candidate sets, but neither set is a random sample of all disclosed conflicts on Wikipedia.
The four categories show the latest observed procedural state of each sampled draft at the cutoff.
Latest observed AfC classification on 23 September 2026. An ordinary mainspace move is kept separate from a verified AfC acceptance.
We followed one page-instance drafting episode from its first pending AfC submission. At the cutoff, six of the 40 had a move log explicitly describing publication of an accepted AfC submission. Two more were moved from Draft to mainspace, but their public move logs did not establish AfC acceptance. The other latest procedural states were 31 declines and one explicit rejection. A decline is not a rejection: an episode could be declined, revised and later accepted. Indeed, 35 of the 40 had at least one decline during follow-up, including some that later moved to mainspace.
At entry, 18 contributors had attributable evidence explicitly stating paid work relevant to the draft. The other 22 used a first-person COI declaration, usually the Article Wizard’s generic wording, without establishing whether the contributor was paid or specifying the relationship. We do not treat those 22 as unpaid. A generic declaration also does not, by itself, establish that a paid-disclosure requirement was fully met.
Article history establishes whether requested text appeared, who made the change, and whether any mapped text remained at the cutoff.
Coverage of the actionable original request at any point after entry. A match could have been made by the requester and could later be removed.
For each of 26 eligible request discussions, we compared the article before the request, subsequent article revisions and the version observed at cutoff. In 15 cases, all actionable parts of the original ask had a corresponding article change at some point; in six, some but not all did; in five, no new requested change was verified. These are observed text matches, not editorial approval. Content that was already in an article before a request was not counted as a newly implemented change.
Among the 21 requests with a mapped change, four were edited in the article by the same public account that made the request. Seventeen were edited by a different account string, which does not prove that the editor was independent or unpaid. The median interval to the first mapped change by a different account was 22.6 days among those 17 cases (middle half: 13.7–51.5 days). It is a conditional description, not an expected response time for a new request.
Two core changes were subsequently removed; 19 of the 26 requests still had at least one mapped change in article text at cutoff. Eight requesters had explicit entry-time paid evidence. The remaining compensation statuses were unclear, so this study does not compare paid and unpaid requests.
The marker shows the median; the bronze line shows the middle half of observed times. These are descriptions of the sample, not promised turnaround times.
All 40 draft episodes
Middle half: 0.2–71.4 daysSix accepted drafts only
Middle half: 28.3–111.0 days17 requests with a matching change
Middle half: 13.7–51.5 daysAcross all 40, the median interval from the sampled pending submission to first observed AfC review was 5.1 days (middle half: 0.2–71.4 days). Among the six accepted episodes only, the median time to AfC acceptance was 86.7 days. That conditional time is not an estimate of how long any new draft will take, and the unaccepted episodes are not assigned a time-to-acceptance of zero.
Each example connects a request or article revision to a specific coding distinction.
In the Michael Jones request, the requester proposed a speaking section and then inserted it directly about 43 seconds later. A reviewer later declined the request. The article text matched; the sequence was not editor-mediated approval.
The Living Films request led to an added infobox listing. A later editor removed that listing. Its “ever matched” and “present at cutoff” values differ.
A paid representative offered branded and brand-neutral FDA wording for a dry-eye article. An editor added a brand-neutral version with a secondary citation. A separate later branded sentence was removed; the neutral core remained. We did not code that removal as a full reversal.
The Purandar Mishra discussion asked editors to consider a 1959 birth date. The pre-request article already stated that year. Later formatting changes and a request marker did not create a new factual addition in the opening ask.
Some reviewers explicitly raised sourcing or promotional-language concerns, including in a company rewrite discussion. Those comments show how a particular request was reviewed; they do not amount to a systematic Stonebridge quality rating of all sources or proof that disclosure caused the outcome.
The frames, selection rules and coding boundaries determine what these counts can support.
We reconstructed candidate leads from public AfC sorting revisions and from historical COI edit-request queue/Talk-page records. The A archive included 10,600 queried draft page-instance histories, collected in page-ID order, but was incomplete. Explicit disclosure phrases plus historical queue linkage yielded 852 candidate A leads. The B search covered 1,397 resolved Talk-page histories and yielded 179 strong heading/template leads. Those are discovery counts, not counts of all eligible English Wikipedia contributions.
Before outcome coding, we used a fixed SHA-256 ordering seed to draw 40 leads for each workflow. All 40 A leads met the episode rules. Of the 40 B leads, 14 were excluded after checking actual request inception: nine were old discussions still visible in 2026; the others failed a historical pair, marker, substantive-request or entry-window requirement, or disclosed the relationship only after the request began. The 26 remaining B discussions form the request denominator. One excluded case illustrates the timing rule: the request preceded the relevant employment disclosure by about a day.
For A, acceptance required a move log explicitly linked to an accepted AfC submission; an ordinary Draft-to-mainspace move was a separate outcome. We tracked first review, later review and procedural status through 23 September 2026 at 14:16:40 UTC. For B, each new, actionable part of the original request needed a corresponding article revision. We recorded requester versus different-account edits and later removal. A Talk-page response, template closure or request permission alone was insufficient. The public data tables link each included case to its entry revision and outcome evidence.
Interpretation limit: These are reproducibly selected cases from incomplete, observable candidate frames. They cannot establish a Wikipedia-wide likelihood of acceptance, the effect of disclosing a conflict, whether an unidentified contributor was paid, or whether every accepted article stayed published. Three requests concern the same CafePress article. We therefore report counts and conditional medians, without population confidence intervals or causal comparisons.
We wrote a source-independence and tone rubric, but did not complete a source-by-source assessment for every sampled case. This report does not present source-quality correlations or call any draft AI-authored from its prose. Classification was model-assisted and checked again by the same AI assistant against the stored public evidence; automated checks verified IDs, dates, denominators and data exports. There was no independent human validation. Deleted or suppressed material and undisclosed relationships are outside the observable record.
The Wikimedia Terms of Use require disclosure of relevant paid contributions. English Wikipedia’s COI guidance strongly discourages direct editing where a conflict exists and describes disclosure and request routes; its paid-contribution guidance adds detail. AfC and edit requests are different editorial workflows. Our case categories describe public statements and events; they are not compliance judgments.
Other researchers and practitioners already examine nearby questions. Lumino Digital’s 2026 AfC report studies a larger general-submission set from 2025, with a different collection and unit; its reported percentages cannot be used as a control group here. The COI Edit Request Tracker follows request queues, and a Wikimedia-hosted research project addresses paid disclosure and COI. This report’s contribution is the linked, bounded case record across the two separate processes, not a claim to be the first or largest study.
The public files connect every included case to its source revision and coded outcome.
The two tables contain the sampled case IDs, entry revisions, observed classifications, cutoff flags and public evidence links. The data dictionary defines the fields and limits; chart data gives the counts behind both figures. The tables minimize account details; source revisions remain on Wikipedia under Wikimedia’s own terms and licenses.
Research and publication by Stonebridge Creative Technologies. This report is independent of, and is not endorsed by, the Wikimedia Foundation or English Wikipedia. Case classifications reflect public evidence available through the observation cutoff and may need revision if the public record changes or a documented coding error is found.