We ranked all 423 NOS uploads from the last 90 days two ways: by how much the audience engaged with them (like rate), and by how many people the algorithm showed them to (views). On a healthy channel, those two rankings roughly agree. On yours, they're inverted.
Your best content isn't underperforming. It's undiscovered.
| Video | Like rate | vs channel avg | Views |
|---|---|---|---|
| Op zoek naar vermiste kinderen door bendegeweld | 5.71% | 3.3× | 7,565 |
| Tekeningen Jip en Janneke hangen binnenkort naast de Nachtwacht | 5.66% | 3.3× | 5,673 |
| Bioloog David Attenborough verkent al 100 jaar de wereld | 5.59% | 3.2× | 23,326 |
| Hulp bieden in Zuid-Sudan met gevaar voor eigen leven | 5.48% | 3.1× | 9,111 |
| Te weinig hulpmiddelen komen Gaza in | 4.81% | 2.8× | 3,349 |
Missing children, aid workers in Zuid-Sudan, Gaza, national culture. This is NOS at its best — and the algorithm shows it to almost no one. Meanwhile your channel-average Short does 20,833 views.
The mechanism: YouTube decides how widely to distribute a video within its first hours, based on early signals — click-through, first-seconds retention, early engagement. Your humanitarian stories open like everything else you publish: establishing shot, formal framing, no sound design. They don't generate the early signals, so the algorithm never tests them with a wider audience — even though the people who do find them engage at 3× your average. The content is right. The packaging suppresses it. And that gap is only detectable when you rank an entire catalog both ways at once.
Sample: all 423 NOS uploads, 90 days ending Jun 28, 2026. "Most-engaged" = the 5 highest like-rate uploads; "most-viewed" = the top 10 by views. The 13× headline compares average views of the two groups (132,041 vs 9,805). Public like and view counts via YouTube Data API; like rates computed only where YouTube exposes public like counts.
Across both channels' top Shorts, every high performer travels one of two paths:
Emergency, disaster, national event. Wins regardless of production — no music, no engineered hook needed. This is how NOS wins.
| NOS heatwave Short | 140K |
| NOS severe weather Short | 133K |
| Music / sound design | None |
The catch: the news is only dramatic some of the time. Your Shorts average 20,833 views with high variance — big spikes, deep troughs.
Music + empathy hook + text question. Creates urgency from any story. RTL applies it to 100% of sampled Shorts — systematically.
| RTL sampled Shorts range | 34K–58K |
| Music / sound design | 100% |
| Peaks vs NOS | Lower |
The payoff: a much higher floor. RTL's Shorts average 39,495 views — nearly double yours — with far less variance, on the same news cycle.
Structural attributes from deep analysis of the top 5 Shorts per channel (18 attributes each). Channel averages from all Shorts in the 90-day window: NOS n=155, RTL n=149.
Exactly one of your five top Shorts uses music and an empathy-driven voice hook: the Defqon.1 cancellation — interviews with disappointed festivalgoers. It's structurally identical to RTL's default formula, and it did 68K views with zero topic urgency. No disaster, no emergency — a cancelled festival. When you accidentally use RTL's playbook, it works. You just never do it on purpose.
Within RTL's own long-form catalog — same team, same thumbnails, same production formula — Dutch-proximate and personally relatable stories average 32,511 views; far-away or abstract stories (Texas, Venezuela, tournament predictions) average 2,918. An 11× gap that production cannot close.
The hierarchy: proximity > production > topic. Pick close-to-home stories, package them with Path B tools — that's the whole playbook.
Sample: RTL's top 10 vs bottom 10 long-form videos by views, 90-day window — a within-channel control that holds production constant.
We ran each channel's #1 most-viewed video through the full Prism pipeline. Every scene gets a machine-scored tension state and intensity level. Plotted over time, the two videos have opposite signatures:
Flat. Every scene scored calm, intensity 2–3. Ceremony observed from a distance — nothing builds, nothing releases.
An arc. Opens at maximum intensity (5/5, shock), sustains (4/5), then releases (2/5, the memorial). Classic dramatic structure — 2.8× the views.
No analyst drew these curves. The pipeline scores tension and intensity for every scene of every video it processes, automatically and identically each run. The full scene-level data behind both charts is below — click to verify.
Two fully-processed videos shown as an illustration of what the pipeline detects — not a statistical claim. At catalog scale, this scoring runs on every video.
Everything above traces back to structured extraction. Below is the complete pipeline output for the two videos in the previous section — every scene, every speaker, every on-screen text entry. Same system that processed all 1,930 videos. Click any claim's source and check it.
Every finding above is extracted directly from the video content — not generated, not interpreted, not inferred. Run the same video through the system tomorrow and you get identical output. No hallucination, no drift. The data is structured, stored, and queryable — not a one-time report that changes when you ask twice.
Three production-level changes, all within your current team's control, all derived from the findings above:
Your #1 video opens on vehicles arriving; RTL's opens on maximum-intensity human stakes. First-seconds retention is the strongest early signal the algorithm reads — and it's decided before your story even starts.
1 of your 5 top Shorts uses sound design. 5 of RTL's 5 do. Your one exception — Defqon.1 — did 68K without topic urgency. Move from 20% to 100% and Path B becomes yours too.
RTL's dominant hook is a text question (60% of sampled top Shorts); yours is visual action. Apply the question hook specifically to your high-engagement humanitarian stories — the buried treasure from Finding 01.
None of this requires new journalism. It requires packaging the journalism you already do so the algorithm gives it a chance.
Everything in this brief came from public data. With read-only YouTube Analytics access, we correlate retention curves with the scene-level extraction you saw above — the exact second viewers leave, matched to what was on screen at that moment. This is what that output looks like:
Representative example — retention percentages are illustrative until we have your analytics. The scene timeline on the left is real (see raw extraction above). The correlation engine is what access unlocks.
Read-only access, revocable any time. We correlate your retention curves, traffic sources, and audience data with the scene-level content analysis — and deliver a diagnosis tied to specific videos and specific seconds.
This runs on Prism — a video intelligence platform that processes content through a single multimodal AI call, extracting every dimension simultaneously. The analysis above isn't manual research. It's infrastructure: it runs continuously and works on any channel in any language.
| Channel | Subs | Videos | Avg Views | Momentum | Shorts Avg |
|---|---|---|---|---|---|
| NOS | 328K | 423 | 19,797 | -32% | 20,833 (n=155) |
| RTL Nieuws | 185K | 393 | 29,286 | +59% | 39,495 (n=149) |
| NU.nl | 107K | 392 | 37,858 | +4% | 40,689 (n=300) |
| De Telegraaf | 185K | 584 | 25,996 | -13% | 42,466 (n=48) |
| Hart van NL | 18K | 138 | 6,675 | -58% | 13,960 (n=26) |
| Attribute | NOS | RTL |
|---|---|---|
| Vertical-native | 100% | 100% |
| Self-contained | 100% | 100% |
| Professional quality | 100% | 100% |
| Text on screen | 100% | 100% |
| Branding | 100% | 100% |
| Hook timing | 0 sec | 0 sec |
| Cuts per 10 sec | 3.6 | 4.2 |
| Music / sound design | 20% | 100% |
| Primary hook type | visual_action (60%) | text_question (60%) |
| Avg duration | 62s | 58s |
| Attribute | Top 10 (19K–72K views) | Bottom 10 (1K–5K views) |
|---|---|---|
| Custom designed | 100% | 100% |
| Text overlay | 100% | 100% |
| Has faces | 90% | 80% |
| Avg text words | 7.3 | 6.0 |
| High saturation | 60% | 50% |
Data: YouTube Data API v3, 90-day rolling window. Thumbnail analysis: Gemini 2.5 Flash, 22 attributes per image (n=10 per channel cross-channel + n=20 within-channel control). Video structural analysis: Gemini 2.5 Flash, 18 attributes per video (n=5 per channel). Full datasets available on request.