The Editorial Filter: How We Curate the Daily Drop
Every morning at 13:00 UTC — that's eight Eastern, five Pacific — our system runs through several dozen Baby Trump videos surfaced by the YouTube Data API overnight and surfaces the ones we think are worth your attention. The result is deceptively simple: a single "Top Drop" on the homepage, a feed of fifty-odd secondary picks underneath. The pipeline behind it isn't simple at all. This is an explanation of how we decide what makes the daily, why it matters that something has to be filtered out, and what we'd do differently if we were starting over.
Why We Filter at All
The naive version of this site would just embed everything tagged "baby trump" on YouTube. We tried that for two days in our first week of internal testing and the feed was unusable. The raw query returned an enormous amount of off-genre content: actual baby videos with "trump" in unrelated context, news clips about the president's grandchildren, AI-generated content tagged "baby trump" that was actually about a different politician, foreign-language videos with auto-translated titles that picked up the keyword, and a long tail of low-effort uploads that were clearly AI-spam channels trying to ride the trending tag.
The unfiltered feed had a roughly five-to-one ratio of off-genre to on-genre results. A casual visitor would scroll for thirty seconds, conclude that the site was mostly noise, and not come back. The filtering exists not as an editorial flex but as table stakes: without it, the site doesn't function as a discovery tool.
The Filter Stack
The pipeline runs in four stages. Each one is cheap to compute and the cumulative effect is to take roughly two hundred candidate videos per day down to the ~50 that make the feed and the single one that becomes the Top Drop.
Stage 1: keyword regex. Every candidate's title and description has to actually contain the phrase "baby trump" or a close variant ("baby donald trump", "baby trump & baby Biden", etc.). This single filter removes roughly 30% of the raw results — mostly tangentially-related videos that surfaced for one of the less-specific query phrases. It's a blunt instrument but a fast one.
Stage 2: channel-history check. If a candidate video is the only Baby Trump video the uploading channel has ever posted, it's deprioritized. This catches both the algorithm-chasing one-off uploads and the AI-spam channels that cycle through trending tags. Channels with a sustained body of Baby Trump videos (BNN obviously, but also the smaller creators who've made the genre a core part of their output) get a corresponding boost.
Stage 3: engagement signal. We pull view count, like ratio, and comment volume from the YouTube API for every candidate. A video that has cleared the keyword and channel filters but has a like-to-view ratio below a floor (currently about 2%) gets demoted. The like ratio is a surprisingly good proxy for "did the video actually land or did it just collect impressions" — the AI-spam videos and the off-genre videos tend to have low like ratios even when they accumulate views from algorithmic recommendation.
Stage 4: recency weighting. All else equal, a fresh video is preferred over a stale one, but the weighting is gentle. Top performers from earlier in the season stay in the feed for weeks if their engagement holds; the Top Drop rotation specifically pulls from the top ten ranked candidates by a date-deterministic selector, which means the same featured video shows up across all visitors on a given day but a different one shows up tomorrow.
What We Don't Filter For
A few things we deliberately do not gate on, even when readers occasionally suggest we should:
Political leaning of the joke. Our filter is content-shape neutral. A video that mocks Baby Trump and a video that mocks his opponents both clear the same engagement and channel-history floors. The genre is broadly satirical of political theater itself, and the strongest entries tend to be the ones that deflate everyone in the room. Filtering for partisan alignment would defeat the point of having an editorial filter at all.
Production polish. As discussed in our companion piece on what makes the form work, polish is not a strong predictor of comedic quality and is sometimes a negative one. We don't penalize rougher animation or simpler backgrounds; the engagement-signal stage tends to handle this correctly without us having to make a subjective call.
Video length. Within a wide band — thirty seconds to twelve minutes — we don't weight by runtime. The data shows that shorter videos win more often on average, but the occasional long debate-compilation video legitimately earns its placement and we don't want a rule that would cap it out.
Editorial Overrides
The algorithm doesn't run unsupervised. A small number of editorial overrides exist and they all run in the direction of removing things rather than promoting them. Specifically:
A video that's been flagged by a viewer as containing content that crosses into targeted harassment, dehumanizing language, or violence imagery is reviewed and typically pulled from the feed. We don't host the underlying videos — they're embedded from YouTube and YouTube's own moderation runs in parallel — but our editorial position is that we won't surface that material on our front page even when YouTube hasn't removed it.
A video that's clearly mislabeling its creator (re-uploads of another channel's content with the credit scraped off, AI-translated re-voiced versions of someone else's video, etc.) gets dropped. The genre has a creator ecosystem and the site should reward people for their own work, not the people downstream re-uploading it.
A video flagged by its original creator as one they'd like removed from our curation is removed. We've had this happen twice; in both cases the request was honored within twenty-four hours.
The Top Drop
The single featured video on the homepage is selected deterministically by date from the top ten ranked candidates. This means every visitor on a given day sees the same Top Drop, and the selection rotates predictably day by day. We do this rather than randomizing because the deterministic selection lets us add light editorial commentary in the daily email (which features the same Top Drop) without a mismatch between what the email and the site are showing.
The daily email goes out at 13:00 UTC to anyone signed up, includes the day's featured video, a high-resolution thumbnail, the channel name, the description, and direct links both to YouTube and to the on-site detail. It's the single publication moment of the day; everything else on the site is a discovery surface around it.
What We'd Do Differently
If we were rebuilding the curation today, two changes would be on the list. First, the channel-history filter is currently a binary boost rather than a continuous weighting, which means a brand-new channel that posts a legitimately great video has to overcome a meaningful initial penalty. A logarithmic weighting on channel track record would handle this more gracefully without losing the spam-filtering benefit.
Second, the engagement signal weights all sources equally, which over-rewards videos that pick up algorithmic traffic relative to videos that find smaller but more-engaged audiences. A view-velocity rather than absolute-view metric would correct for this, but requires more historical data than we currently store. We're accumulating it; expect this to ship later in the year.
Until then, the filter does the job it was designed to do: take the noisy raw feed of "baby trump" search results and surface the entries actually worth your time. If you ever spot something we missed, we want to know — submissions to hello@baby-trump.com get reviewed within a few days, and channels we like tend to get added to the search query set.