Insights

Insights measures the posts ChannelForge publishes for you, from your own connected accounts, and tests what goes with more views and more engagement — inside each channel, with posts compared at the same age, and with a plain label that says how sure each answer is. Use it to decide what to make more of, and what to stop.

Every number on the page is a measurement or a test of measurements. Nothing is estimated, extrapolated or forecast, no model writes a number or a sentence, and nothing on the page promises views or earnings.

What it does#

What is collected, per platform#

Each post is read with the account that published it, inside your workspace — never with another workspace's connection. Every call only reads.

A number a platform does not report is never counted as zero. It is listed under Data gaps, with the reason and, where you can fix it, what to do (for example, reconnecting an account whose sign-in was refused).

How often#

Young posts decide a short's fate, so they are read most often:

Each workspace is measured once an hour, at a minute of its own, away from the hour and the half hour when posts usually go out. A collection that found something new is analysed, at most every 6 hours. A post the platform says no longer exists is marked gone and not asked for again.

How posts are compared fairly#

A post that is two days old cannot be compared with one that is two months old by their raw views. Insights uses three measures:

• Views for their age — a post's views against the typical post of the same age on the same platform in your workspace, shown as × typical (×2 is twice the usual for its age, ×0.5 half). How views grow with age is measured inside each post — from its first and last measurements once it is a day old, taken at ages at least a quarter apart — never by lining old posts up against new ones. Until 10 posts on a platform have been measured that way, posts are compared only with posts of a similar age.
• Settled views — views once a post is past 72 hours.
• Rates — the engagement rate (likes, comments, shares and saves per view) and, on Instagram once it is permitted, the share of the video watched. A rate counts only for posts with at least 100 views: on fewer, it is noise.

What is tested#

Comparing posts only with posts of their own channel matters: channels differ in reach and niche, and that is not the title's doing.

Measured alerts#

Some things need no test, only a measurement. They appear badged Measured:

• Not delivering (Facebook) — five or more settled Reels, nine in ten of them at five views or fewer.
• Stall (Instagram, and YouTube once permitted) — a channel's last three posts all at a quarter or less of its own usual views at 24 hours (the median of up to ten posts before, at least five, when that is 100 views or more). The alert says whether most of your channels on that platform fell at the same time.
• Starved (the same platforms) — a channel whose usual views at 24 hours, over ten or more posts, are under a twentieth of your other channels' there (two or more of them).
• Breakout — a post under a week old with at least three times your workspace's median settled views on that platform, and 50 or more. The stricter rule behind follow-ups is in Breakout follow-ups.

A channel is only ever compared with channels of your own workspace — never with another customer's. An alert that stops being true is gone from the next analysis.

Find it in the app#

• Insights in the sidebar, after Analytics, opens the workspace's overview.
• Each channel's name in the Channels × platforms table opens that channel's own Insights page; the channel page links there too.
• The Analytics page keeps the plain counters of the last 30 days; with Insights on they refresh hourly while a post is young.

Every member of the workspace can read Insights, viewers included. There is nothing to switch on.

Reading the overview#

Each finding has a short headline, its numbers, the sample it rests on and its q-value. How it was tested opens the full sentence code wrote from the numbers. A finding about the same subject on other outcomes is folded into one line ("Also: settled views ×1.66 (suggestive)"). Each finding has an evidence id (F001, F002 …) so it can be cited.

An example headline, as the page writes it:

Strong · Instagram · F014
A number in the title: ×1.42 the views at 72 hours of the same channel's other posts on Instagram
×1.42 (95 % CI ×1.18–×1.73) · n = 38 (15 and 23) · q = 0.031

Here 38 posts were compared, 15 with a number in the title and 23 without.

A channel's own page#

A channel's Insights page shows each platform (the channel's median against the typical post of the same age in your workspace, the test against your other channels, followers with a small chart), the measured alerts on that channel, its best and weakest posts for their age with a views-by-age line for each (hover or focus a point to read its measurement), the findings about the channel, the findings across your channels that apply to it, and its accounts' data gaps.

How sure each answer is#

Every answer is a statistical test, labelled in plain words:

What the statistics do, without the formulas:

• The permutation test asks how often an effect at least this big would appear by chance. It shuffles which posts had the feature (only inside each channel) 2,000 times; the share of shuffles that reach the real effect is the p-value.
• The 95 % interval is found by re-drawing the posts 2,000 times. A narrow interval means the effect is pinned down; one that crosses ×1 means it could be nothing.
• The q-value (Benjamini–Hochberg) corrects for asking many questions at once: ask forty and one or two will look real by luck. Labels rest on q, not on p.
• The same data always gives the same numbers. Each test is seeded from its own question, so its p-value and interval do not change because something else was added to the analysis; its q-value, which allows for every question asked, can.

A null result is a finding too. "Not detectable" on titles with a question means your audience, so far, does not care either way — and you can stop worrying about it.

Credits and plans#

Insights costs nothing to run: the measurements and the analysis hold and charge no credits, and they do not appear in Activity. The credits per 1,000 views column uses what your episodes were charged.

Insights is on the Creator plan and above. On the Trial the page explains what it is and links to the plans. The same analysis, Facebook's and Instagram's parts only, is also available through the API on plans that include it: see the API reference.

Limits and good to know#

• Your data stays yours. Each workspace is analysed alone. Nothing of one customer's numbers enters another's comparison.
• How long it is kept. Measurements are kept 400 days and the record of each collection 30 days; the last three analyses are kept. Posts of the last 180 days are analysed. A deleted workspace's measurements are deleted within days (see Privacy).
• Carousels are not measured; video posts are.
• Permissions still to come. Reach, plays and watch time on Facebook and Instagram need permissions Meta grants after its App Review of our app; YouTube's watch time and retention need a YouTube Analytics permission the connection does not ask for yet. Until then those numbers are data gaps. When a permission arrives, you may need to reconnect the account once.
• Meta's hourly allowance is respected: past three quarters of it, a collection stops reading Facebook and Instagram for that hour, so publishing keeps the rest.

Troubleshooting#

Related#

• Hook lab — openings tested on the same measurements
• Retention to script loop
• Breakout follow-ups
• The coach — reads these findings each week
• Monetization tracker
• Platform policies — what we do and do not do with YouTube's data