@Isaac and @diego - we are planning to do a scrappy experiment in Q1 where we ask people for their interests and use that data to show recommendations in the Suggested Edits feed. This feels quite related to the DE1.3.5 recommendation algorithms research, so I figured you might have ideas, or at least be interested in what we are thinking about here. Feel free to chime in, follow along, or I'm happy to meet and chat through ideas if you want to discuss and learn more.

Excited to see this move forward! First question I guess is what makes a good recommendation. Given the focus on retention, I'd say three factors to pay attention to.

• Motivated to do: these are newcomers so editing is hard/confusing. They need some reason to make an attempt. This is going to be easier if they are familiar with the topic and even more so the closer they feel to being a relative expert of sorts on it. For example, folks know a lot about the country they live in, but so do many other people. Much better to find the most niche article that they care about (like their town/neighborhood), as they will have relatively more expertise to contribute and that can boost confidence. Things like trending/most-viewed can help motivate folks but these are extrinsic factors and it's going to be much more effective to get a newcomer to a topic that they intrinsically believe is important.
• Capable of completing: these are Suggested Edits so there's less of a burden on the newcomer to spot an issue themselves, but the smaller the article, the less context they need to work through to understand a potential edit. This also gets to the familiarity piece -- the more familiar you are with the topic, the easier to assess whether the suggested edit is a good one.
• Convince them to stick around: this is a product of how motivated they were to do the edit in the first place, how impactful the edit feels (there is a reason why add-an-image significantly boosts retention whereas add-a-link seems to have almost no effect despite boosting activation), and what happens after the edit. Regarding the last point, articles that are already higher quality or have more editors contributing to them are also probably places where a newcomer is more likely to trigger a revert. Lower-quality articles with fewer edits tend to be safer places to start.

Translating the above into recommendations:

• I'd say that we almost certainly want to avoid the higher-popularity/higher-importance/higher-quality articles -- i.e. not use trending / most-viewed / vital / featured. These might receive more click-throughs (newcomers will be more likely to recognize these and think they might want to edit them), but they're also going to be more likely to end with disappointment (don't end up publishing the edit or get reverted when they do) and cause consternation within the experienced editor community. That's because these articles will be larger, have more editors already watching them, and often the content is just inherently more complex and difficult to summarize too. Taking this a step further, for pretty much any strategy taken, I'd suggest boosting lower-quality content as an easier place to start. For example, with geosearch, I'd consider not just showing the nearest articles but perhaps also including a few more distant articles if they're lower quality.
• I would use random or very high-level topic-based interests only as a fallback as opposed to primary mechanism for surfacing candidates. Some editors presumably like the randomness but for most, these tend to lead to confusion as they're pretty unlikely to overlap with the newcomer's actual expertise -- e.g., I'm into sports but what do I know about this random football player?
• Geo-based is good! More likely to be relevant/motivating and hopefully lots of lower-quality articles that could be found. To get around the location-privacy stuff, perhaps a way to search for an area the person knows about?
• Generally mixing strategies is a good idea so long as there are explanations with each suggestion so it's not confusing. There really is no perfect strategy when we lack context so trying a couple things is a nice way to hope something sticks. To this point, making it easier to scan multiple candidates quickly might be worth trying (as opposed to cards that need to be individually swiped to explore as we do with the current Newcomer Tasks module)?
• Presumably you'll need something like the Search box in the prototype. I think some UX iterations would be great to help make this less open-ended and more clear what people should put in -- e.g., is this a good place for semantic search? How do we best prompt people to find something of interest to them? Should we pre-populate some ideas if people are stuck -- e.g., "try places you've lived" etc.
• To me, the goal is to get the newcomer to a single article where they feel comfortable editing. This isn't about building a full interest profile - they can either repeat the process for their next edit or eventually build an edit history we can use. In that sense, I'd rather have a deeper process that aims to get to a single good recommendation than a shallower process that finds several okay candidates. For example, let's say we start with geosearch. From that list of results, we should encourage the newcomer to select whatever is most interesting interesting to them. Now we can use that article as the seed to generate the actual candidates -- e.g., morelike to get similar articles that also have a recommendation available (and ideally are lower quality). Or look at the links in the article they selected and use them as candidates (again having recs available and lower quality).
• One note: it's probably less important to get the choice of article right if you switch this into much more of an explicit learning experience -- i.e. heavily-guided experience with pre-quiz perhaps, something like add-a-link for the edit, and then post-publish information about how to view the article history and what to expect (and perhaps a nudge to their mentor to review their edit). In that case, I'd say just go with something low quality so the edit is less likely to be controversial but you can probably choose something random at that point. And only on success do you try to help the newcomer find a more relevant article.

those are great prompts, @Isaac!

to add more context, here's our current thinking on account setup (aka the welcome survey).

if someone starts the account creation process from an article, we use that article as a seed article to populate the interests screen. beyond letting people search for articles themselves, we want to offer a few one-tap suggestions - articles related to the seed article (or additional articles that a person may add manually on this screen via the search). those interests are then used to populate the suggested edits once someone lands on the newcomer homepage. they won't necessarily be the exact articles the system suggests for editing.

now we're trying to figure out what to show when someone starts account creation from the Main Page or other pages that can't be used as a seed article.

we considered showing random articles, but that feels too random.

during account setup, the idea is to present a few articles that feel relatable, so new accounts can pick one or two as interests. those interests then inform the pool of suggested edits, increasing the likelihood that a newcomer will feel inclined to act on them.

geo-based selection seems promising. again, these articles won't necessarily be what the system suggests for editing - they're just how a new account expresses interest.

that's also why trending articles feel lower-risk: they're only used to express interest, which then maps to the pool of suggested edits, still controlled by community configuration among other things.

thanks for the additional context @AAlhazwani-WMF !

that's also why trending articles feel lower-risk: they're only used to express interest, which then maps to the pool of suggested edits, still controlled by community configuration among other things.

yeah, this is a really good point so let me revise a little what I was saying. I had been thinking of them as recommendation candidates but you're right that if we apply the same trending-as-the-seed as opposed to trending-as-the-recommendation-list, I'm much more in favor of including it as an option. They should hopefully be somewhat diverse but mostly recognizable so not a bad indication of where someone's interests lie. In that sense, featured articles are also probably fine. I'd still be a little hesitant around Vital Articles as these tend to be much more historical and general, so presumably okay as "things I'm interested in" but not a great seed for actual recommendation candidates as the recommendations are going to tend to be more general/historical as well and those aren't great places to start for newcomers either. If we go this route, we'll have more overlap between newcomers (i.e. different people choosing the same seed articles) so we'll likely just want to introduce a bit of randomness into how we expand the seed to reduce overlap -- e.g., don't always take the same top-5 candidates from morelike but randomly select 5 from the top-20 to display.

Another idea as I look at the flow for the from-an-article scenario: I want to expand on my previous point about it feeling more important to find one good article recommendation than several okay options. What if instead of this being a start-up process that the editor is expected to do once, we treat this as the flow for resetting one's Suggested Edits feed that can always be easily re-entered? My sense is that folks' interests are pretty diverse and context-dependent. For example, looking at my own edit history, I have a background interest in journalism that I tend to edit on but I'll go on little sprints where e.g., I visited a new museum that I liked and was under-documented so I did some expansion work for a bit there. Or when I'm doing more gardening during the summer, I tend to be a bit more interested in editing articles about plants. All to say, I think folks often have background interests but then they go through little sprints where they focus on a different topic for a bit before returning back to their base interests. I've heard this from folks too around current events -- e.g., doing more football-related editing during World Cup. This is pretty anecdotal so also leaving myself a TODO to do some quick qualitative coding around this and come back with actual data. Getting back to design, if we pivoted to this model:

• I'd drop the idea of selecting multiple interests ("What are 3 of your interests?") and instead focus on just gathering a single interest ("Find an article you're interested in"). For the account-creation-from-seed-article flow, this would be pre-populated with that seed and it would essentially just be a confirmation screen for the user (or they could remove it and choose something different). In the no-seed flow, it would be a list of candidates as we're discussing. This should make this part of the flow a lot faster. Because we only want one article, we could then switch most of the screen real estate to previewing the suggested article candidates that would be generated once a seed is selected so folks could decide whether to proceed or keep iterating.
• On the Homepage, there'd be a prominent "Reset my feed" or something like that that would take the person back to find-a-seed screen and let them reset their interest.
• We could also build in a URL parameter for the Homepage that that would let someone set the Suggested Edits feed from any arbitrary seed article. This way edit-a-thons etc. could easily link into the feed while personalizing it to their event.
• If someone gets to the end of their feed of Suggested Edits, we could have a "Try another interest" button or something like that so they could kick off the process again.
• Maybe we still have an option too where folks can switch their feed to "personalized based on recent edits" or something like that so they eventually can have a general feed that's always available and a bit more diverse. But again it would always be easy to take on a new "persona" for a little. I particularly love this as it feels like a much more Wikipedia-based form of personalization than how most social media platforms do it (allow folks much more control over the personalization and the ability to easily "forget" their history and become a different user for a bit).

Thoughts?

Coming back to what the task is actually asking: Which articles should be presented to the user in the interest selector when no seed articles are available?

• Random articles
  • trivial, already implemented.
• No suggestions:
  • trivial, and probably the first we should A/B test
• Most viewed / trending articles via generator=mostviewed
  • simple, needs very minor post-processing to remove special pages
• Geolocation + trending per country
  • kinda doable. We have an api for that: https://wikimedia.org/api/rest_v1/metrics/pageviews/top-per-country/DE/all-access/2026/08/05 but it is not filtered by project. This should be doable manually in a good-enough way for the experiment. But if that is to scale then we need this for a per-project basis. E.g. the most-viewed pages from Germany in French Wikipedia.
• Geolocation-based suggestions
  • tricky. The geo-search does not allow large bounding boxes. Maybe we can do some magic for country-code->country name->wikipedia-article? But this might need more depth.
• Curated lists as a random pool:
  • simple if you can define the curated list. The scalability issue has already been noted.
• mixed: the combination of the above

A/B/C testing this should be relatively straight-forward, barring sufficient number of users ever get that far.

Overall, this is something we can do at any time after the new early onboarding has moved forward. It is not blocking us in any way.

I'm moving this to QA with the note that this is for Product (@KStoller-WMF, maybe @AAlhazwani-WMF ) to review and decide how to move forward, not QA-Lead.

In T433151#12196848, @KStoller-WMF wrote:

Thank you, Michael. As a first step this is really all I was looking for.

Geolocation-based suggestions
tricky. The geo-search does not allow large bounding boxes. Maybe we can do some magic for country-code->country name->wikipedia-article? But this might need more depth.

The search area is the same size as the search around the central spot: 20km x 20km. And even then, the results seem to center around its center as opposed to sampling randomly from all the articles available in the box.