Keeping a Reels experiment log that decides something

What the log is for

A Reels log is not a diary and not a content calendar. It has one job: compare two Reels fairly and decide what to do next. A field that never changes a decision is weight you carry for nothing.

The trap is memory. You remember the Reel that went far past your usual numbers and forget the eight ordinary ones it should be measured against. Without a written baseline you compare a Reel with an impression of your account, and that impression moves with the week and your mood.

Seven fields, and no more

One entry per Reel, written when you post or right after. Fill the first six in before you post and leave them alone afterwards. That is your own rule, not a lock in a tool, and it is what makes the entry comparable later.

• Posted at: date, local time and timezone, so readings line up.
• Working title: short, recognisable, no adjectives.
• Variable: exactly one, from a closed list such as hook, length, topic, format or posting time. A closed list stops you inventing a category every week.
• Variant: what is different, in plain words. A question instead of a statement. Forty seconds instead of twenty.
• Hypothesis: one sentence, see below.
• Link: the post itself, so you can find it again in four weeks.
• Numbers: later, each with its age, plus the time you read it in your own note.

What to leave out

Three things look useful and are not. First, the follower count as a control: it moves on its own and explains almost nothing about one Reel. Second, likes as the headline number: they respond to something other than views and make old and new entries hard to line up. Third, a verdict written in the same minute as the post.

A free note field is fine. Put observations there, not conclusions. Opens on the pour, no intro, filmed at seven in the evening is an observation. This one is a winner is a conclusion you cannot check later.

The hypothesis template

One sentence with four parts: the observation, the expected direction, the metric and the age. Because of what you noticed, you expect a variant to produce more or fewer of one metric at one age than your usual Reel.

Example: because my last three Reels opened on a statement, I expect a question hook to produce more views at 24 hours than my usual Reel.

Set the threshold when you write it, not when you read the result. Example: it counts as confirmed if 24-hour views land 20 percent above my median for that age. A bar fixed in advance separates a test from a story.

If the sentence needs an and, you have two tests. Split them.

Every number needs an age

Record views at fixed ages: about 24 hours, about 3 days, about 7 days. Note in your own log when you actually read the number. A reading at 31 hours is not a 24-hour value, and treating it as one spoils the comparison for every later entry.

Use the same metric for the decision every time and let the others run as context. Missing stays missing: if you were on a train at the 24-hour mark, leave the field empty. An empty field is honest, a guessed number is not.

Cumulative numbers only. If your source shows the total since posting, record the total, not the difference to yesterday.

The twenty minute weekly review

Pick a weekday and a time and keep both. The review is short because the log did the work during the week. Sort the entries into three groups: still open, ready to judge, already decided.

Read the oldest undecided entry first. Judge only entries whose last measurement age has arrived. Write three things into it: the decision, the number behind it, and one line about next time. Then choose next week's test from an open question in the log.

Do not add entries during the review. Writing and judging in one sitting is how a log turns into a journal.

Share, repeat or drop

Compare with your own median at the same age, not with another account and not with a different age. A seven day total against a 24-hour number flatters whichever Reel is older.

You need a band, because raw numbers wobble. One workable rule: 20 percent or more above your median means share, 20 percent or more below means drop, in between means repeat. Treat the band as a convention, not a law of the platform.

Two guards keep it honest. With fewer than three earlier comparable tests you have no median worth trusting, so mark the entry unproven and keep collecting. And do not retire a variant after one bad Reel. One result is noise, the same direction twice is a reason to stop.

Share does not mean post more of everything. It means the pattern goes into your next three Reels unchanged, so you find out whether it holds.

From entries to a rulebook

Once a month, read the decided entries by variable instead of by date. Count how often each variant beat your median and on how many entries the count rests. Two matching entries are a hint, not a rule. With three to seven earlier tests the confidence stays low, and only from eight does it become medium.

Keep the rules short and dated, and delete the ones that stop working. The log is the raw material, the rulebook what you use for the next script.

Where a tool earns its place

All of this works in a spreadsheet with seven fixed columns. What it does not do is keep the age of every number, compute the median of earlier tests at the same age, and refuse a verdict when the sample is thin. That last step is the one most people skip.

marketingapp.io keeps the log in that shape. A Reel is one test with one variable from a fixed list, a variant, a hypothesis and a posting date; the posting time is stored as one moment, without a timezone column. Views are recorded at about 24 hours, 3 days and 7 days, a hand-entered value counts for the fixed age you pick, and a missing value stays missing. The comparison uses the median of your earlier tests at the same age and channel, and until three comparable tests exist it says there is too little data instead of guessing. Share at 20 percent or more above, drop at 20 percent or more below, repeat in between.

You can also let an agent write the entries. Connected over MCP, it reads the state of the workspace, logs a test or a measurement you confirmed, and saves a research find. It does not publish a Reel, does not read private Insights, and the product never asks for your Instagram login. The judgement stays with you.

Questions