How to read Shorts analytics

An 18 August 2026 Similarweb snapshot recorded 161,345 monthly visits to storyshort.ai, with 50.1 percent arriving from organic search and a bounce rate of 36.8 percent. Those three numbers come from one dashboard, yet each answers a different question, and reading them as a single verdict is how most channel decisions go wrong. This page is a method for separating them.

The short answer

Read Shorts analytics by answering two questions in order: how many people were shown the video, and of those shown, how many stayed. Impressions and swipe-away rate answer the first; the retention curve answers the second. Compare every number against your own median across recent uploads, never against another channel, and never against a single day, because Shorts distribution arrives late and unevenly. Record what you changed alongside each measurement, or the numbers have no meaning.

The steps

1. Separate the two questions the numbers answer

Every metric on the Shorts analytics screen belongs to one of two questions. The first question is reach: how many people were shown the video at all. Views, impressions and feed placement belong here. The second question is hold: of the people who were shown it, how many kept watching past the opening seconds. Average view duration and the retention curve belong here. A video can fail at either stage for opposite reasons, and applying the fix for one failure to the other makes the next upload worse. Before opening any report, write down which question the number answers. If you cannot place a metric into one of the two questions, ignore it until you can.

Output: a two-column list of your metrics, labeled reach or hold.

2. Read the retention curve's first drop before anything else

The retention curve shows the share of viewers still watching at each moment. On a Short, the shape that matters most is the first drop, in the opening seconds. That drop is the note the hook left: steep means the opening promised something the viewer did not believe, shallow means the opening earned the next ten seconds. Everything after the midpoint describes whether the body held, which matters less, because most viewers who leave do so early. Read the curve from the left edge inward, note where the steepest fall ends, then check whether that point matches the moment your script changes subject. When the two points coincide, you know what to rewrite.

Output: a timestamp marking where the first drop ends, matched against the script's first change of subject.

3. Compare a video only to your own median

A median over your own recent uploads is the only fair baseline. Another channel's numbers carry a different subscriber base, a different niche saturation and a different posting history, so no comparison to them survives contact with reality. Take your last ten Shorts, order their average view durations, take the middle value, and treat that as zero. Each new video is above, below or near your median, nothing more. The same applies to views: the bank of attention YouTube gives a Short varies between uploads for reasons outside the video itself, which is why view count alone cannot rank your work.

Output: a single median figure per metric, computed over your own last ten uploads.

4. Do not read a single day

Shorts distribution is delayed and uneven. A video published today may receive most of its feed placements over the following days, in bursts, sometimes weeks apart after a resurface. Reading the numbers on day one measures the delay, not the video. Set a fixed review window, such as seven days after publish, and read once inside it. Between reviews, do not open the dashboard. Checking daily produces decisions based on noise, and noise always looks like signal when it moves in the direction you hoped.

Output: a calendar entry naming the review date for each upload, and no dashboard sessions before it.

5. Write down what you changed, next to the numbers

Measurement without a record of what changed is not measurement. If you altered the hook, the length, the caption style or the posting time, write the change down beside the metrics it might affect, dated. One variable per upload is the rule: two simultaneous changes make both results unreadable. After several cycles, this log becomes the only part of your analytics that actually predicts anything, because it tells you which of your own interventions moved which of your own medians. Tools that produce the videos can help here by keeping inputs consistent. ViewMade, for example, renders each Short from sourced archive footage with a media credit file listing every clip's origin, so the visual input stays documented across uploads and the log tracks script and structure rather than guessing at footage provenance.

Output: a dated changelog column sitting next to the metric column, one change per row.

What this will not fix

Analytics tells you what happened, not why. The retention curve shows where viewers left; it does not show whether they left because the hook was weak, the topic was wrong for the audience the algorithm picked, or the previous video set an expectation this one broke. Only a controlled next attempt answers why: change one thing, keep everything else identical, publish, wait out the review window, compare to median. If the reason were readable directly off the dashboard, every channel reading the same screen would fix the same problem the same way, and retention data would stop being useful. It remains useful precisely because the cause requires an experiment the dashboard cannot run for you. Budget for that experiment before drawing conclusions from any single curve.

Where to go next

Once you can read a single Short honestly, /guides/how-to-measure-whether-a-format-is-working extends the same median-based comparison across several uploads sharing one format, which is the level at which format decisions should be made. To see whether the problem sits upstream of any individual video, /guides/how-to-audit-your-own-channel walks the whole channel against its own history before you touch the next script.

<!-- faq -->

Frequently asked questions

How long should I wait before judging a Short's performance?

Seven days is a defensible default window, applied identically to every upload. Shorts receive feed placements in delayed, irregular bursts, so a video that looks dead on day two can resurface later. What matters is not picking the perfect window but using the same one every time, so your comparisons stay valid across uploads. Write the review date into your changelog when you publish.

Which single metric matters most for Shorts?

None, taken alone. Swipe-away behavior in the first seconds correlates most closely with how far the video distributes, but reach depends partly on factors outside the video. Read the first drop on the retention curve together with impressions: strong impressions with a steep early drop points at the hook, weak impressions with a shallow drop suggests the packaging or topic did not earn the test at all.

Should I delete a Short that performed badly?

Deleting removes the data point from your median without removing whatever caused the result. Keep it, label it in your changelog with what was different about it, and let it lower the bar honestly. A median built only from survivors stops being a baseline and becomes a flattering fiction. The exception is content that violates platform policy, which is a moderation decision, not an analytics one.

How many videos do I need before my median means anything?

Ten is the smallest sample that produces a usable middle value, and even then treat it as provisional. Below ten, one outlier drags the median enough to flip a verdict. Until you reach ten uploads under roughly similar conditions, spend the time recording changes and reviewing curves rather than ranking videos against each other. The changelog you build now is worth more than an early verdict.

Do I need paid tools to do this?

No. Every method on this page runs on the native analytics screen, a spreadsheet and a changelog. Paid tools add automation around production and scheduling, not better answers to the two questions of reach and hold. If a tool claims to tell you why a video failed, it is running the same single-variable guesswork you would run yourself, just priced monthly.