The 70/20/10 pillar split, measured
Treat 70/20/10 as a starting prior. Tag videos by audience overlap, measure each pillar, and reallocate the mix when the numbers stop matching the split.
70/20/10 is the split almost everyone writes on a whiteboard: 70% proven core, 20% adjacent or trending, 10% experimental. It is a prior, not a measurement, and it is not a law of YouTube. Treat it as the allocation you start with because you have no data, then replace it with a split your own tags can defend.
The useful version of the idea does not count topics. It counts audience overlap. Paddy Galloway's public framing is the cleaner one: roughly 80% of videos should target people who already watch you, and 20% should be experiments aimed at people who do not. That is a statement about who the video is for. "Educational / behind-the-scenes / trends" is a statement about what it looks like, and those two are not the same column.
70/20/10 is a prior, not a quota
A quota says episode 17 must be experimental because the mix is light this month. A prior says: in the absence of evidence, most of the calendar should talk to people you already reach, a smaller slice should sit next door, and a thin slice should be allowed to fail in public.
Three reasons the quota version fails.
Topics are not pillars. "Tutorials" can be core on one channel and experimental on another. A pillar is a repeating promise to a repeating audience. If two tutorials pull disjoint commenters, they are two pillars sharing a costume.
Trend-chasing is not the 20%. Evergreen versus trending is a decay-curve question: some files earn most of their views in a week, some accrue for months. You can have an evergreen experiment and a trending core. Mixing the two taxonomies on one calendar is how the 20% becomes "whatever is going around."
Cadence is not the mix. Shipping more files does not create a 10% you can learn from. It creates more noise. Do not raise posting rate to hit a ratio. Cadence experiments belong after the measurement loop is visible, and content velocity without tags is just a faster way to not know which pillar worked.
Start at 70/20/10 if you have fewer than 12 tagged videos. Write it on the calendar as a guess. The rest of this page is how the guess gets replaced.
Define pillars by who they are for
Name at most four pillars. Five is how they collapse back into "content." For each one, write a single sentence of the form: this is for [person] who already believes [X] and is here to [job].
Then tag every video with one audience label, independent of the pillar name:
| Audience tag | Test | Typical share at the start |
|---|---|---|
| Core | A regular viewer would recognise this as "one of ours" in the first frame | 70% |
| Adjacent | Same person, different job-to-be-done | 20% |
| Experimental | Built for a person you do not currently reach | 10% |
Galloway's 80/20 maps onto this table as core-plus-some-adjacent versus experimental. 70/20/10 just splits the existing-audience bucket. Neither ratio is the point. The point is that experimental is defined by who is missing, not by "we used a new caption style."
A format is not a pillar. This or That can run inside a core pillar and inside an experimental one. If you treat the format as the pillar, you will read a dead payload as a dead shape, or the reverse. Hold format still inside a pillar when you can; when you change both in the same week, the tag cannot tell you which lever moved.
Audience fatigue still applies. A core pillar that has become interchangeable from video to video is a format problem sitting in the 70%. Reallocation will not fix sameness. A different skeleton will.
A tagging scheme you can actually keep
If tagging takes more than 30 seconds per video, it will stop. One sheet, one row per publish.
Columns:
- Publish datetime (with timezone)
- URL
- Pillar (one of your four names)
- Audience tag (core / adjacent / experimental)
- Format (the skeleton, not the topic)
- Length band
- Metric (the same one you would use to kill a series)
- Value at 24 hours
- Impressions at 24 hours, so you can see when a row is too small to vote
Rules that keep the sheet honest:
- One primary pillar. Secondary tags are how everything becomes "a bit of both" and the mix is always 100% of everything.
- Tag before publish. Tagging afterwards is how hits get relabelled experimental and misses get relabelled tests. The audience tag is a prediction, not a memoir.
- Length band is a column, not a vibe. Completion rate is not comparable across a 20-second punchline and a 55-second explainer. Compare inside the band.
- Format is copied from the bible, not invented per row. If you cannot name the skeleton, you cannot tell a pillar miss from a format miss.
Twelve videos is the first time the sheet is allowed to change the calendar. Before that, you are filling the prior. A 90-day brand calendar is the right container for the guess. An AI content calendar is fine as the publishing surface. Neither replaces the tags.
The reallocation rule
Every 12 videos, or every four weeks, whichever is slower, compute a three-video rolling median of your metric for each audience tag, inside each length band you actually use. Then apply this, in writing, before you look:
Promote an experimental skeleton into adjacent when its rolling median has sat at or above the core median for two consecutive windows. Demote an adjacent skeleton into experimental (or out) when its rolling median has sat below 80% of the core median for two consecutive windows. Do not move more than one skeleton per window. Never let experimental hit 0% of the calendar.
The 80% is a starting threshold, not a published benchmark. Set it from your own spread once you have a few windows. What the sentence is doing:
- Two consecutive windows. One lucky file is not a pillar. Feed variance on short-form will fake a promotion if you let it.
- Against the core median, not against a screenshot from another channel. Experimental that "did well" at 12,000 views is not a promotion if core sits at 40,000 on the same metric. Views are the wrong metric anyway; use Viewed vs. Swiped Away, completion, or sends per reach.
- One skeleton per window. Moving three things at once recreates the original confusion at a new ratio.
- Experimental never hits zero. A calendar with no experiments has no way to grow the core. The 10% is the price of having a next 70%.
Worked numbers, as arithmetic, not as a benchmark. You published 12 videos: 8 core, 3 adjacent, 1 experimental. That is 67/25/8, close enough to the prior. The experimental file's 24-hour completion is above core. You do nothing. One file cannot promote a skeleton. You run that skeleton twice more. If the three-video median still clears core, it becomes adjacent in the next 12, and one adjacent slot returns to core so the existing-audience share does not collapse.
If nothing experimental has cleared the bar in three windows, change the skeleton, not the percentage. Raising experimental from 10% to 30% because you are bored is how the core starves. The 70% is the channel. Protect it on purpose.
Reallocation is also how you stop a dead adjacent format from occupying the slot a working one would use. That is the same instinct as milking a format while it delivers, applied to the mix rather than to a single run: move the hours to the tag the median still defends.
Do not use the 12-video window to raise cadence. If the loop is visible and you can staff another core episode at the locked quality, that is a staffing question. It is not a mix question.
FAQ
What if my experimental videos outperform core immediately?
Then you may have mis-tagged them. If regulars are watching and sending, the audience was core and you labelled it experimental because the topic felt new to you. Re-read the test in the table: would a regular recognise it as "one of ours" in the first frame? If yes, it was never experimental. If no, and the median holds for two windows, promote it. A single over-performer is still one row.
Is 70/20/10 different on Shorts, TikTok and Reels?
The ratio is not platform-specific. The metric is. Shorts: Viewed vs. Swiped Away. TikTok: average watch or completion, length-banded. Reels: sends per reach — Mosseri's published Reels predictions are reshare, watch-through, like, and a tap through to the audio page. Keep one sheet per surface if you publish natively on more than one, because a Reels send and a Shorts swipe-away are not one number. The audience tags can be the same.
How does this sit next to an evergreen / trending mix?
Orthogonally. Tag both. A video can be core-evergreen or experimental-trending. The evergreen versus trending split tells you which decay curve you bought. The audience tag tells you who you bought it for. Collapsing them into one 60/25/15 or one 70/20/10 is how both measurements die.
Can I automate the mix so I stop thinking about it?
You can automate publishing. You cannot automate the audience tag without lying, because the tag is a prediction about who the video is for, made before the numbers exist. Schedule the slot, batch the production, keep a human on the twelve-video review. A calendar tool that posts the wrong 70% at a perfect cadence is still the wrong 70%.