I Read the First Five Words of 25 Shorts. Most Did Not Name the Subject

The first-five-words pilot results

What counts as naming the subject in five words

An explicit-subject opening had to name the person, object, topic, or question the viewer would use to identify the premise. 'This phone battery failed' qualifies because the object and event are present. 'You won't believe what happened' does not, even though it signals surprise.

A situational-action opening could establish enough activity to orient the viewer without naming the central object. 'I walked into the meeting' gives a role and action, but the actual conflict may still be missing. The category is descriptive, not automatically weaker than an explicit noun.

Generic or incomplete language included greetings, discourse markers, pronouns without a visible referent, suspense shells, and fragments whose meaning arrived after token five. A visual-first Short with no speech can still be strong; the category says nothing about the images unless they were studied separately.

The coding difference in plain language

These are recreated examples, not quotations from sampled creators.

Five words are not a stable multilingual measuring stick

Languages package meaning differently. A pronoun can carry information that English expresses with a noun. Compounds, clitics, code-switching, honorifics, and tokenization rules change what counts as one word. A five-token window is convenient for a pilot, but convenience is not measurement equivalence.

The sampled captions included Hindi fragments and other multilingual signals. Automatic captions omitted or distorted some words. YouTube's own automatic-caption documentation warns that pronunciation, accents, dialects, background noise, overlapping speakers, and multiple languages can produce incorrect or missing text. Creators are told to review and correct it.

The full study should therefore begin with one validated language cohort and use human-verified first spoken clauses. Separate language replications can follow with fluent coders and documented tokenization. Combining them first and apologizing in the limitations later would produce a cleaner chart and a worse result.

The caption timestamp was not a speech-onset measurement

Nineteen of the 25 recovered tracks began within 250 milliseconds, and the median first cue timestamp was 160 milliseconds. That looks like a speaking-speed result until the measurement is examined. The timestamp belongs to an automatically aligned caption cue, not a laboratory annotation of the first audible phoneme.

Automatic captions can shift the start of a cue, group several words, or delay availability while processing. A long initial silence can even prevent automatic captions from generating normally. The correct label is 'first recovered caption-cue time,' not 'time the creator began speaking.'

If speech onset matters, trained reviewers should inspect the waveform under a frozen acoustic rule. If semantic onset matters, they should mark the point where the clause becomes specific. Those two timestamps can differ, and neither should be inferred from ASR metadata alone.

Raw Shorts views would be the wrong outcome

A view-count-ordered sample mixes channel scale, upload age, topic demand, geography, repeat plays, and discovery surface. Since 31 March 2025, YouTube counts a Shorts view when the video starts or replays, with no minimum watch time. The earlier measure remains as engaged views in Analytics, according to YouTube's current Shorts view explanation.

Comparing five-word categories against raw views would therefore reward older uploads, larger audiences, repeat starts, and popular topics alongside any wording effect. Even a matched public cohort would remain observational. The stronger outcome is creator-authorized early audience response inside niche, duration, upload-age, and channel-scale bands.

The useful claim is not 'nouns get views.' It is, at most, whether a verified opening-language category is associated with an early audience measure after the study accounts for obvious clustering and confounders. The causal version would need an experiment that changes wording while preserving the rest of the video.

The full study's opening-language codebook

The pilot's three buckets should expand before performance analysis.

A first-five-words check you can use before posting

• Write the first spoken clause exactly as a cold viewer hears it; do not summarize what you meant.
• Circle the first concrete subject, object, action, conflict, or question.
• Remove greetings and discourse markers temporarily to see whether the premise becomes faster without changing your voice.
• Check the visual track separately; a deliberate visual-first opening may not need speech to carry the subject.
• Review automatic captions against the audio, especially for names, dialect, code-switching, and overlapping speakers.
• Test two versions with target viewers if the opening choice matters. Do not infer success from word category alone.

The validated first-five-words study I would run next

Start with 300 English-language clips and human-verify the complete first spoken clause. Use two trained coders and preregister the six mutually exclusive categories above. Audit transcript verification against the audio and stop publication if accuracy falls below the registered 95% threshold.

To test association, use creator-authorized analytics or tightly matched public cohorts within niche, upload-age band, channel-size band, duration, and discovery source. Treat channel as a random effect when repeated channels are allowed. Publish the distribution of wording categories before any outcome model.

A filler leaderboard can be a secondary descriptive feature after recognition accuracy is established. It should never become a shame list or a claim that one phrase kills a Short. The primary research question is whether specific opening information predicts a defined response inside a valid comparison—not which creator used 'so' most often.

Claims this pilot cannot support

• High-performing Shorts usually name the subject in the first five words.
• Generic openings cause viewers to swipe.
• Creators should always speak within 160 milliseconds.
• Five English-style tokens measure equivalent information across languages.
• No-speech openings are weaker than spoken openings.
• The first five words predict exact views, retention, or recommendation.

Research and accessibility sources

YouTube's caption creation guide explains that caption files contain text and timing, with some formats also carrying position and style. That makes captions useful research material, but it does not make automatic tracks a verified transcript.

Research with Deaf and hard-of-hearing viewers documents the difference between caption availability and caption quality. The YouTube creator captioning study found that manually added caption signals can affect confidence and trust. This pilot did not audit accessibility; the source strengthens the case for human verification and context-rich labels.

No sampled transcript excerpt is reproduced on this page. Aggregate categories preserve the research point without turning creators' speech into a searchable quotation archive. The internal source frame remains subject to the documented platform-data refresh or deletion deadline.