How Find ranks videos
Remaining views over 48 hours on TikTok and 7 days on YouTube, then brand fit.
Find forecasts how many more views a clip should collect while a planted comment can still be seen. The input is the search snapshot only: current views and age. There is no second fetch, and there is no 30-day total.
Cards are then grouped, scored for brand fit, and ordered. The same grouping is what you see on Search results and on get_status cards.
Groups
get_status puts each card in bucket. Prefer recommended.
bucket | Meaning |
|---|---|
recommended | Inside the plant window, at least 5,000 remaining views, and a brand-fit score high enough to lead with. |
low_views | Still inside the plant window, but remaining views are under 5,000. |
not_recommended | Same remaining-view floor, but the title and caption do not fit the product well enough. |
too_old | Older than 48 hours on TikTok, or older than 7 days on YouTube. A new thread is unlikely to be seen, even if the clip already has a lot of views. |
The 5,000 floor is the same at every age. It is applied to remaining views, not to views already on the clip. A two-minute video with views can clear it. A six-day Short with more current views can miss it.
Inside each group, cards are sorted by remaining views, then by brand-fit score. Brand is scored on every card that is still shown, including low remaining and too old.
Formula
Search gives current views V and age t in hours. Later views are treated as a multiplier of current views (Szabo and Huberman). Remaining over horizon H is the extra views expected between now and t+H:
remaining = V * (exp(-t / tau) - exp(-(t + H) / tau)) / (1 - exp(-t / tau))That is the closed form of exponential saturation:
V(t) = Vinf * (1 - exp(-t / tau))Vinf is implied by the views already seen. Age is floored at one minute so a publish time equal to now does not divide by zero. There is no cap on remaining. A young clip with views has a high views-per-hour rate, so the same formula predicts a large remainder.
| Parameter | TikTok | YouTube Shorts |
|---|---|---|
Horizon H (also the plant window) | 48 hours | 7 days (168 hours) |
Time constant tau | 18 hours | 72 hours |
tau is Rumora's stand-in for a median curve. It is not a table fitted on our corpus. Pinto and KTFN would estimate the multiplier from a daily or hourly history. Search does not give that history, so the exponential is the one-point substitute.
A missing publish date, or zero views, produces no remaining number. Those cards fail the floor and land in low_views unless they are already past the plant window.
Sources
The quantity (remaining views over a nearby horizon) and the idea that later views scale with current views come from these papers. The exponential tau values and the 5,000 floor are ours.
- Gabor Szabo and Bernardo A. Huberman, Predicting the popularity of online content (Communications of the ACM, 2010). Later popularity is a multiplier of current popularity on the log scale.
- Henrique Pinto, Jussara M. Almeida, and Marcos A. Goncalves, Using early view patterns to predict popularity of YouTube videos (WSDM 2013). The shape of early daily views separates videos that will keep growing from videos that already peaked. Find sees one snapshot, not that shape.
- Shisong Tang et al., Knowledge-based Temporal Fusion Network for Interpretable Online Video Popularity Prediction (The Web Conference, 2022). On Douyin (TikTok's sister For You stack) they forecast the next 72 hours from the previous 24. That is why TikTok's plant window is two days, not a month.
- Will It Go Viral? Grounding Micro-Video Popularity Prediction on the Open Web (WebShorts / Shorts-Cast, 2026). YouTube Shorts popularity is tracked through day 7. That is why YouTube's horizon is a week.
Click-by-click search is in Find videos. Agent polling is in Find then comment.
Last updated on