Lista de experimentos em Produto e Tecnologia
Esta lista de experimentos tem como objetivo servir de visão geral centralizada dos experimentos realizados no departamento de Produto e Tecnologia.
Cada experimento inclui uma breve descrição do problema que estamos tentando resolver, o tipo de experimento, a população selecionada e outros aspectos importantes do trabalho. Esta página está em construção e será atualizada regularmente.
Sobre esta lista
Os experimentos estão listados de acordo com seu status:
- Upcoming: The experiment is being designed and configured for data collection. We're setting up what we want to test, how we'll measure it, and getting everything ready. It is not active.
- Live: The experiment has been deployed and is actively collecting user interaction data with specified traffic on defined wikis. The experiment is active for a pre-defined duration.
- Analysis: The experiment owner is reviewing results and community thoughts.
- Decision: The owner has reached a decision, to be documented in the Outcome field and typically falls into one of three categories:
- "Ship it" – Recommend to roll out the winning version
- "Don't ship it" – Results show the feature didn't perform as expected; don't implement
- "We need to learn more" – Results are inconclusive; needs another test, qualitative feedback or further iteration
Experimentos
Próximos
| Experimento | Equipe | Prazo previsto | Tipo de experimento | Público | Problema a resolver / por que estamos fazendo isso | Compartilhe seus comentários |
|---|---|---|---|---|---|---|
| 'Review Changes' module in Home | Moderator Tools | TBD | A/B test | TBD | 'Review Changes' is a personalised module for Home which aims to introduce newer editors to the concept of reviewing the edits that other users are making. The feed of edits to review is personalised to each user, displaying edits to pages they've edited or watched, and to pages related to the ones they edit the most. We see this as step 1 on a journey the user could take towards becoming an experienced patroller.
We will observe whether the presence of 'Review Changes' leads to higher retention or other behaviour changes. |
|
| Article guidance:
Surface a "generic guidance" option for editors whose topic does not match an available outline |
Language and Product Localization | Start Date: 2026-TBD
End Date: 2026-TBD |
A/B test | Simple English Wikipedia and/or Turkish Wikipedia | A significant share of users (39% of total starting users, or 45% of users who reach the topic matching step) exit the Article Guidance flow at that point. This is likely a combination of unclear UI and Wikidata matching not surfacing the right result for the editor's topic. We believe offering a "generic guidance" path, one that does not depend on a Wikidata match or an outline existing, would move more editors through the workflow and increase the chances of more article creation attempts. | |
| Surface the availability of edit suggestions on articles to logged-in readers via controlled experiment | Editing | Start Date: 2026-TBD
End Date: 2026-TBD |
A/B test | TBD; registered users with ≤100 cumulative edits | As a logged-in Wikipedia reader who has edited before, does not do so regularly, and is arriving on-wiki seeking to learn about something/someone (broadly defined), I would value knowing when there are clear, actionable, and personally relevant improvements available within the article I am reading so that I can easily decide which (if any) I'd like to act on in this moment without needing to seek them out myself.
Hypothesis: If we surface the availability of edit suggestions on articles to logged-in readers through a controlled experiment, there will be a ≥# increase in the proportion of distinct registered users with ≤100 cumulative edits who publish at least one constructive edit (mobile). |
|
| Add semantic routing to existing search on pilot wikis | Apps (Android) | end Oct 2026 | Pilot Wikis (French, Arabic, and Japanese) | Logged out readers (Android) | Offer an optional semantic-search path on apps to demonstrate 1.5% increase in exploratory searchers for logged out app readers | |
| Hybrid (Semantic & Lexical) Search on Web Test | Reader Growth (Web) | 30 Oct 2026 | A/B Test (Semantic-Lexical variants) | Logged out readers (Web) | Test hybrid search on the web, so that we can increase search feature retention for logged out readers | |
| Retest Image Carousel | Reader Growth | 20 Nov 2026 | A/B testing | Logged out Readers (Web) | Our retest on the image carousel to measure the impact of incorporating community-requested features (e.g., captions and a jump to article section link that give context to the image thumbnails) on logged out reader retention |
Em andamento
| Experiment | Team | Anticipated Timing | Experiment Type | Audience | Problem to solve / why we are doing this | Share your feedback |
|---|---|---|---|---|---|---|
| Test methods for email confirmation enforcement for new accounts, when the user chooses to provide an email address during account creation | PSI | October 2026 | A/B Test | New users registering for web accounts, who choose to provide an email address. | Today, users can create an account with an attached email address, without confirming that they control that address. This means that many email addresses attached to new accounts remain unconfirmed, and so we cannot treat them as reliable for account recovery, security notifications, reminders, or other important communications. The current model also leaves room for preventable misuse involving invalid, third-party, or repeatedly reused email addresses. On English Wikipedia, 52.63% of new accounts that provided an email address confirmed it.
The experiment tests methods for email confirmation enforcement for new accounts, when the user chooses to provide an email address during account creation. It does not require users to provide an email address, but requires confirmation if they do provide an email address. The experiment consists of two A/B tests: delayed enforcement and upfront enforcement. Delayed enforcement is now live, with upfront enforcement planned to follow. The purpose of the experiment is to determine which approach (upfront enforcement, or delayed enforcement) is the most effective at driving email confirmation, and is the least likely to create less negative impact on retaining active readers and editors. The results will provide directional evidence on which enforcement approach is more appropriate and where to focus further engineering investment. |
|
| Early onboarding: | Contributor Growth | Start Date: 2026-09-29
End Date: |
A/B test | Arabic, English, French, and Spanish Wikipedias | The current Welcome Survey introduces friction at a critical moment in the user journey (immediately after account creation) without offering any visible benefit in return. Key issues:
|
|
| Streamline the toolbelt on Minerva (Minimal Minerva) | Reader Growth | 31 August | A/B Test | Logged out Readers | If we streamline the toolbelt on Minerva, we will see practically significant improvement to logged-out reader retention. | |
| Listening to Wikipedia: Android text-to-speech experiment | Android Apps | October 2026 (30 days) | A/B test | Selected users of the Wikipedia Android app who have English set as their feed language and follow at least one topic. Eligible users are randomly assigned to a test or comparison group, so not everyone who meets these conditions will see the feature. | This experiment tests a short-form listening experience in the Android app’s “For you” feed. Selected users can listen to pre-generated audio of article introductions related to topics they follow.
The purpose is to learn whether readers find listening useful, return to use it again, and would like additional ways to listen to Wikipedia. The results will help the team decide whether further investment in listening experiences, such as more article sections, full articles, or broader language support, should be explored. This is an early experiment and does not mean that the model or voice being tested has been selected for a permanent feature. |
|
| Article recommendations to new account holders | Reader Experiences | 29 Sept 2026 | A/B Test (Treatment-Control variants) | Logged in Readers (Web) | We are hoping to enrich the reader experience and thus increase Internal referral rate by 1% for new account holders who identify as Readers by providing article recommendations Project page on MediaWiki | |
| Cross-wiki Code Collaboration | MediaWiki Content Platform Team | October 2026 | Pilot wikis (Wikipedia: Hindi, Punjabi, Odia, Telugu, Malayalam) | Users on Wikipedias that use the same kind of modules (such as infoboxes, citations, dates, language tools, etc.). | This experiment deploys the Produnto extension to understand if users will adopt this approach to sharing Lua modules. |
Análise
| Experiment | Team | Experiment Timeline | Experiment Type | Audience | Problem to solve / why we are doing this | Share your feedback |
|---|---|---|---|---|---|---|
| Reader → Contributor Attribution Extended Experiment #1
Build on the successful baseline experiment of reader-to-contributor registration rates across traffic sources (e.g. search, social media, AI assistants, and other referral channels). |
Various teams involved in Key Results IR1 and IR2 | Early June – Mid July 2026 | Invisible (no discernable UI change) | Sampled anonymous readers across Wikipedia projects, sampling the top 50 largest Wikipedias | Wikimedia would like to understand how different types of readers become contributors. This experiment instruments a privacy-preserving pipeline to estimate reader-to-contributor conversion rates across traffic segments such as referrer source and device class. Building on the successful Reader → Contributor Attribution Baseline Experiment, this extended experiment expands data collection to improve estimate precision, evaluate additional traffic segments, and provide deeper insight into how readers become contributors. Findings will inform future reader-to-contributor research and product decisions related to contributor growth. |
``` |
Decisão
| Experiment | Team | Experiment Timeline | Experiment Type | Audience | Problem to solve / why we are doing this | Share your feedback | Outcome |
|---|---|---|---|---|---|---|---|
| Guide Temporary Account users toward full account creation | Contributor Growth | Start Date: 2026-06-30
End Date: 2026-07-31 |
A/B test | 50% of temp account holders on French, Arabic, Spanish, and English Wikipedias | Temporary Accounts lower the barrier to starting an edit, but they also limit our ability to support and retain contributors over time. Temporary Account users cannot receive email, and experienced editors have fewer ways to teach, mentor, or follow up with them. In addition, we do not consistently communicate the benefits of creating a full account, nor do we guide Temporary Account users toward registration at moments when motivation is likely to be highest.
Early data from the Temporary Accounts banner suggests there are opportunities to better align messaging, timing, and presentation with user intent and readiness. The current experience does not clearly explain why creating a full account is valuable, nor does it prompt users to register at moments that feel relevant and motivating. As a result, many Temporary Account users continue editing without converting, limiting long term engagement, community connection, and retention. |
Ship it!
Experiment results show that the treatment group is 87% more likely to create an account than the control. | |
| Guided article creation
Help new editors easily understand and apply the existing practices and policies, and navigate the task of starting a new article with ease. |
Language and Product Localization | May-June 2026 | A/B test | 50% of users on Simple English, French, Turkish, Wikipedias | Currently, editors need to make a significant effort to create new content from scratch. As a result, reviewers have to process many low quality contributions and revert/fix them. Finally, readers see content creation as a mystery, which makes it harder for them to decide to become contributors in the future. | Ship it!
We observed a statistically significant +17.2% relative improvement in 30-day article survival rate (p=0.033, CI +1.1% to +25.0%), covering mobile and desktop combined across all editor experience levels. This confirms the feature has real positive impact and justifies continued investment. Reaching statistical significance specifically for junior editors on mobile requires either a longer window or a larger wiki set. | |
| Suggestion Mode MVP | Editing | May-June 2026 | A/B test | 50% of users on English, French, German, Japanese, Spanish, Russian, Polish, Dutch, Hebrew, Arabic, Czech, Swedish, Korean, Chinese, Vietnamese Wikipedias | Hypothesis: If we present junior contributors who enter the mobile VisualEditor with immediately actionable edit suggestions, then the proportion of edit sessions that result in someone publishing a constructive edit will increase by >=10%. | Ship it!
When newcomers see a suggestion, the impact is large, positive, statistically significant, and most pronounced on mobile: constructive activation (+110.00%), edit completion (+22.06%), constructive edit rate (+9.33%), edit abandonment w/o changes (-23.78%), 2nd-week retention (+24.67%). | |
| Incident Reporting System enwiki trial | Product Safety and Integrity | May-June 2026 | Progressive rollout | Logged-in users on English Wikipedia | The Incident Reporting System (IRS) is being introduced to help contributors more easily find the right place to seek help when facing harassment or other issues. We are doing this because deploying IRS to the English Wikipedia requires special care, given its size and existing processes. The purpose of this trial is to generate real-world data on how the system performs in this environment. Through a controlled, limited rollout, we aim to evaluate how the system is used and whether it operates without introducing significant additional burden on functionaries. This data will inform whether IRS is effective and suitable for broader use on English Wikipedia. |
Remain on English Wikipedia.
We completed the trial, and ended up showing the Report button to 100% of eligible users for the last two weeks of the trial. During the trial, the tool did not generate a high volume of reports overall, but we also did not seen signs or hear concerns of volunteer disruption. After public feedback from English Wikipedia functionaries and other volunteers, we have left the incident reporting system enabled on English Wikipedia going forward. | |
| Reference Previews Improvement
An experiment to test whether we can serve readers better when a footnote click in read mode shows the full bibliographic information rather than flying them to the reference list. |
Technical Wishes (WMDE) | June 17 to July 1, 2026 | A/B test | 10% of desktop for logged-in and logged-out readers/contributors on German, Swedish, Farsi, Hindi, Somali, and Western Frisian Wikipedia. | The assumption that we are testing is: Most readers who click on a footnote marker do NOT intend to navigate to the reference list. They are satisfied with viewing all the bibliographic information and details in the pop-up/in-line. | To be confirmed | |
| Scale the new Home Feed from Android to iOS | App Experiences (IOS) | 4 Sept | AB Test | Logged out Readers (IOS) App retention | Test new Home Feed against control group with access to Explore, as measured by an A/B test. | Decision pending, early results indicate: Daily app open rate is statistically in the treatment’s favor (99% chance to win; 95% CI +0.037% to +0.398%), but GrowthBook calls it a draw. No detrimental effect from the explore feed | |
| Introduce a new content type in Apps home | Reader Experiences Apps | 20 August- 5 concepts tested in the experiment, outcome results were supported | A/B test | Logged in Readers (Apps) | This test is part of Phase 2 of the revamp of the App's Explore feed, where we're testing out new additions to Home, with the goal of increasing daily returns to the app. This was also mentioned in a recent community wish | Proceed with further Testing: We aim to build & A/B test Addition of new articles by topic to the "For you" feed. Explore how the experience of seeing new articles emerge by topic can be more dynamic. Consider adding a trending map to the Community Feed showing Top Read by country, or a location filter for country to the Top Read module. | |
| Share Card
An experiment to test the usefulness of a share feature that gives readers options to create a visually compelling card about an article (or portion of an article) they are reading. They can share this card to a friend or post it on their social media. |
Reader Growth | Mid May – June 2026 | A/B test | 10% of mobile logged-out readers on Arabic, Chinese, French, and Vietnamese Wikipedias and 0.1% of mobile logged readers readers on English Wikipedia | Readers increasingly prefer to get their information from people they trust, which has fueled a rise in finding information through social media and social contexts. We want to see if people enjoy sharing information from Wikipedia. | Align on whether/how to scale the article-level Share button | |
| Encouraging users with unconfirmed emails to confirm them | Product Safety and Integrity | Mid May for 2 weeks | A/B/C test | Logged-in users with unconfirmed emails on English, French, German, Polish, Japanese, Hebrew, Wikidata, Commons, Mediawiki, Meta | Wikimedia projects do not currently require email confirmation in order to edit and take other actions as a registered account. As a result, many user accounts have associated emails that have never been confirmed. This is contributing right now to users being locked out of their account, with no safe way for them to regain access. This experiment will measure the effectiveness of persistent communication that encourages user accounts that have an unconfirmed email address to confirm it, or set up a new confirmed email address. | Ship it! – The results from this experiment are encouraging. We saw an 80% increase on average in email confirmation rates across the projects in the experiment. We saw negligible impact in email removal rates. Next, we will be rolling this banner out to all communities in June. You can read more about the experiment results here. | |
| Donor Appreciation Badge
Testing two variations of a "thank you" badge shown to donors after a recent donation. The badge is visible only to the donor on their own browser. One variation includes a simple thank-you message, while the other includes a lightweight animation when tapped. |
Reader Experience & Fundraising | June – July 2026 | A/B/C test | Logged-out donors who recently donated on mobile web browsers (1% on English Wikipedia, 10% on German Wikipedia) | We hope to deepen the relationship between donors and the Wikimedia movement through different approaches to appreciating and engaging supporters. Findings may inform future opportunities to invite donors to create accounts for deeper engagement with Wikipedia, including reading and editing. | Inconclusive, not scaling: We need to iterate on this further under different donor sentiment experiments in DE5.4 KR (We may run a future iteration that invites anonymous donors to created WP accounts in order to save their donor status and get access to more reader features).
| |
| Surface Reading Lists to logged-out readers on mobile web to increase account registration rate | Reader Experience | May-June 2026 | A/B test | 10% of logged-out readers on German, Spanish, Italian, Portuguese, Polish, Dutch, Turkish, Urdu Wikipedias | Account creation is the gateway to meaningful participation on Wikimedia projects. This experiment aims to encourage account registration at the moments where motivation is naturally highest, such as creating reading lists. | Further Optimisation is required before scaling. 93% users responded that the feature is useful to them, exceeding our goal of 50%. | |
| Explore Feed Refresh
Testing a refreshed Explore Feed with a more modern design, clearer structure, and lightweight personalization, with a goal of making it easier for readers to discover interesting content and come back to the app more often. |
Apps | Mid May | A/B test | New and casual Wikipedia app readers (especially logged-out users) | The Explore Feed is not engaging enough; most users don’t come back to it, don’t customize it, and don’t clearly understand its purpose. | We exceeded our 10% goal, and saw an 10.5% statistically significant increase in engagement with the feed.
| |
| 25th Birthday Reading Challenge
Encourage readers to build a daily habit by reading one article per day and completing a 25-day streak. |
Apps | May 11 – June 18, 2026 | Experiment | Wikipedia app users (logged-in and logged-out) | Many readers open Wikipedia when they need information, but fewer return regularly. This experiment explores whether a simple, goal-based reading challenge with progress tracking and rewards can encourage more frequent engagement and habit formation. | we exceeded our goal of 5% increase compared to that cohort’s past behavior before joining the challenge.
|