Odinstat

How Odinstat Works: Methodology

Last updated: 2026-09-05

The Core Principle

Every number in Odinstat is computed from your own channel's data with standard, well-understood statistics. Nothing is estimated from "channels like yours", no industry benchmarks are passed off as your data, and no number is ever invented to fill a gap. When there isn't enough data to say something honestly, Odinstat says "not enough data" — that is a feature, not a failure.

Why your data only? YouTube's recommendation system rewards channels for what they uniquely do well. Copying the average of other channels optimises you toward being average. Your own history — what worked for your audience — is the signal that matters, so it is the only signal we use.

Where the Data Comes From

Two official read-only YouTube sources:

  • YouTube Analytics API — daily views, watch time, likes, comments, shares, subscriber changes, audience retention curves, traffic sources, demographics, and geography.
  • YouTube Reporting API — thumbnail impressions and click-through rate per video per day.

Known limits we surface instead of hiding:

  • YouTube reports metrics with a 2-3 day delay; recent days are not missing, just not reported yet. The app labels data with its actual "through" date.
  • Impression and CTR history only reaches back about 30-60 days from when a channel connects. Older videos therefore have no impression data — we show "—" rather than a made-up 0%.
  • A video's CTR is only shown once it has at least 100 tracked impressions; below that the number is statistical noise.
  • Views and likes are lifetime totals. When we compare newer videos with older ones on those totals, videos under 14 days old are left out and the comparison says so — an older video has simply had more time.

The Questions We Lead With

The analyses we put first are the ones that judge one of your videos against your own baseline, because those are the ones a creator can act on:

  • Outliers — which videos sit far outside your usual range, good or bad. Uses the median and the spread around it, so one viral video does not redefine "normal".
  • Thumbnails after reach — click rate compared with what is typical for how widely each video was shown, so a video shown to a million strangers is not judged like one shown to your subscribers.
  • Retention — your real drop-off curves, read for where people leave and whether the opening or the packaging lost them.
  • Retention for length — each video against what is typical for its length on your channel, so a 3-minute video and a 25-minute video are not compared as if length did not matter.
  • Traffic sources — where views come from and which sources work hardest per video.
  • Subscriber conversion — subscribers gained per view, so a small video that converts well can beat a big one that does not.
  • Shelf life — how many days each video took to fall to half its peak daily views, and whether that is a launch burst or genuine staying power.

Other analyses (trend shifts, publish timing, category comparisons, format comparisons, view projections, and more) remain available under "More analyses" and in Settings. They use the same data and the same honesty rules; they are further down because at most channel sizes they more often have to answer "too little data to tell".

The Statistics We Use

Each analysis names its method in the app (every insight card has an expandable evidence section showing the exact numbers and steps). The main tools, in plain terms:

  • Outlier detection — distance from your median in units of your typical spread (a modified Z-score), so extreme videos do not distort what counts as normal.
  • Reach- and length-adjusted comparisons — a simple fitted line (CTR against reach, retention against length) and each video's distance above or below it.
  • Group comparisons (Shorts vs long-form, recent vs earlier) — Welch's t-test and bootstrap resampling, which do not assume your videos behave like a textbook bell curve. These are observational: they describe what happened, not an experiment you ran.
  • Relationships (does CTR track views?) — Spearman rank correlation, which is robust to outliers.
  • Trend shifts — change point detection to find when your channel's trajectory actually changed, rather than eyeballing a chart.
  • Projections — exponential smoothing with uncertainty ranges, never a single fake-precise number.
  • Multiple comparisons — where one analysis runs several tests together (the correlation analysis, the improving-over-time report, and the Deep Dive question batteries), we apply a Holm-Bonferroni correction so we do not report flukes as findings. Single-test analyses report their own p-value and threshold.

Averages are weighted by what they measure: CTR is weighted by impressions, retention by views — a 5-view video never counts as much as a 50,000-view one. Videos with no tracked value for a metric are left out of that metric's average rather than counted as zero.

Where AI Fits (and Where It Does Not)

The Channel Assistant and retention interpretations use a large language model (Google Gemini) — but only as a writer, never as a calculator. The pipeline is:

  • Step 1 — we compute the statistics above from your data.
  • Step 2 — video titles are replaced with anonymous placeholders.
  • Step 3 — the model receives the pre-computed facts and writes a plain-language explanation.
  • Step 4 — titles are restored in the answer you see.

The model is explicitly instructed to describe only what the data shows and to say so when data is missing. AI answers can still contain mistakes — which is why every underlying number is also available directly in the app, so you can check any claim yourself.

Honesty Rules

Rules the product enforces everywhere:

  • No fabricated values: missing data renders as "—" or "not enough data", never as zero.
  • Minimum sample sizes: analyses refuse to run on too few videos rather than produce unstable conclusions, and a test that is not significant on a small sample says "too little data to tell", not "no effect".
  • Observational language: comparisons describe what happened on your channel. They are not experiments, and the copy never tells you to "post on Tuesday" from a correlation.
  • Every insight shows its evidence: the numbers, the method, and the raw data behind it are one tap away.
  • Caveats are part of the result: when a metric is a proxy or a date range is limited, the card says so.