BuildLink Analytics Corporation
LinkLab Research

What production standards must a short-form vertical video meet?

Researched 26 Oct 2026 · Re-verified and corrected 4 Nov 2026

Short answer: the first frame is the thumbnail and the first 2 seconds are the gate — no logo, greeting, or title card survives it. Build 28–45 seconds, one claim, a visual change on every new thought, captions inside the strict safe box, and a voice that rewards unmuting. And the anti-slop standard that governs all of it: specificity — real numbers, real documents, a named place. Evidence below is labelled by TYPE, not dressed up as confidence: PLATFORM (the platform's own document), SURVEY (self-reported responses), VENDOR DATA / VENDOR SURVEY (a company's own data or commissioned survey), PRACTITIONER (a creator/consultant dataset), SELF-REPORTED (an agency or case claim), TARGET (a practitioner goal — not platform-published), STANDARD (this pipeline's editorial rule).

2s
The hook gate. The proposition must be on screen and readable — the swipe decision happens here.
3.75s
Average TikTok watch time across 5.6M videos (Metricool 2025 — a mean, not a median viewer).
70–90%
Viewed-vs-swiped on the best-performing Shorts (practitioner study, 5,400 Shorts). Under 60% rarely performed well.
69%
Say they watch video sound-off in public (Verizon/Publicis survey, n=5,616). Design muted-first.
The honest part
No platform publishes a retention threshold, a ranking weight, or an organic safe-zone spec.

Anyone quoting you a single “you need X%” number is quoting a practitioner target, and this standard labels it that way. The safe-zone numbers for organic posts are measured by editor tools and disagree with each other by 100px or more — the two-tier box in §03 is a risk-managed synthesis, not a guarantee. Several widely repeated numbers in this niche are folklore; where they appear here, they are flagged as folklore, not laundered into facts. And there are no controlled studies of faceless formats in the home-services niches at all — §06 is mechanism fit plus named, self-reported cases, labelled as such.

§01 · Hook & retention
The proposition lands in the first 3 seconds
TikTok's own creative guidance: introduce the content proposition in the first 3 seconds, prioritise the hook in the first 6 — “90% of ad recall impact is captured within the first six seconds.” That is advertising guidance; it prices ad creative and must not be quoted as an organic benchmark.
Platform
The hook decision happens in ~2–3 seconds
Metricool's 5.6M-video study puts TikTok's average watch time at 3.75s and falling. It is a mean across all videos — it establishes when the decision happens, not what a median viewer does. (An earlier “median viewer gone by ~2s” version of this claim was withdrawn — see the corrections log.)
Vendor data
Skip rate is now a printed metric
Since August 2025, Instagram Insights shows skip rate — “the percentage of views from people who decided to skip your reel during those first 3 seconds” — and Meta's developer reference defines it the same way (reels_skip_rate). Keep 3-second skip rate under 30%; Instagram publishes no benchmark, so that figure is a target, not a platform number.
Platform
Benchmarks that define a good Short
Shorts viewed-vs-swiped: 70–90% on the best performers, under 60% rarely performed well (Galloway/Gileta — 5,400 Shorts, 33 channels, 3.3B views, 2023; direction reaffirmed since, exact averages are not 2026 guarantees). Length-banded view-through targets (Retensis, explicitly “not platform averages”): Shorts >60% / >50% / >40% for <15s / 15–30s / 30–60s.
Practitioner

Opening devices that work — ranked for a text-first pipeline

  • 1. Specific number + consequence: “A $14,000 roof. The invoice line that padded it: $2,300.”
  • 2. Bold contrarian / myth claim: the fiction stated plainly, to be broken.
  • 3. Document reveal: open on the estimate, clause, or inspection line — already zoomed, one line circled. Nobody else in the feed opens on paperwork.
  • 4. Direct question naming the audience: “Selling in [Town] this spring? Read clause 4 first.”
  • 5. Pattern interrupt by motion: a scale change, a redacted line un-redacting, a counter ticking — in the first 1–2s, not at 5s.
  • Combine two (number + curiosity gap is the best-documented pairing). The hook statement is fully readable on screen by second 2, and the spoken first line says the same thing.

What kills retention in the first seconds

  • Logo animations, brand intros, “Hey guys / welcome back,” greetings of any kind.
  • Black frames, fade-ins, slow zooms; a title card that precedes the claim instead of being the claim.
  • Throat-clearing (“So today I want to talk about…”) — the proposition arrives after the swipe decision.
  • A first frame that doesn't work as a still thumbnail.
  • Text too small or low-contrast to read at phone size in under a second; hook text under platform UI (§03).
  • Clickbait the body doesn't pay off quickly — it wins the stay and loses the mid-video cliff.
§02 · Pacing & structure
Default length: 28–45 seconds
One claim, 3–5 beats, one payoff. Stretch to 45–60s only when the payoff is cumulative (a line-by-line breakdown, a multi-step process). Under 20s only for a single-stat hit. The evidence conflicts — Galloway's held-duration lesson, TikTok's 21–34s ad band, a faceless-length analysis showing sub-60s medians ~57% higher than 61–120s (EchoEtch — a single independent analysis) — and is reported, not averaged away. Never pad: length only helped when duration was held.
Standard
Change the visual state on every new thought
New card, highlight move, document pan/zoom, number tick. In practice every 2–5s. Hard ceilings: no static frame longer than 6s; no beat longer than ~3s in the first 10s. An evolving document/chart build counts as a change; a frozen frame is a flatline. Honesty flag: no controlled study establishes “change every 2–3 seconds” — this is an editorial standard, not a research finding.
Standard
Text per frame: 3–7 words per card
Kinetic headline cards carry 1 idea and hold ≥1.0s (≥0.8s absolute minimum for a 2–3 word card). Mini-hook or open loop every ~7–10s; the peak fact at ~60–75% through, not at the end where few remain.
Standard
Word budgets — the arithmetic, shown
At a natural 150–165 wpm voiceover, raw capacity is 75–83 words (30s), 113–124 (45s), 150–165 (60s). Less ~10% headroom for the hook pause and the payoff hold, the working budgets are 68–74 / 101–111 / 135–149 words. Check any script with the tool below. Text-only variants (no voiceover): cap the total at ≤70 words at 30s — script to caption pace, ≈145 wpm (Jensema 1998, “Viewer reaction to different television captioning speeds,” American Annals of the Deaf 143(4):318–324, n=578). If it doesn't fit, cut words — never shrink the font or speed the cards.
Standard
Captions: phrase-level, not karaoke
No controlled evidence was found that word-by-word highlighting beats well-timed phrase captions on retention — every “karaoke wins” claim located traces to a caption-tool vendor. The accessibility research cuts the other way: phrase/sentence chunking produced fewer fixations and less effortful processing than word-for-word text (Rajendran, Duchowski, Orero, Martínez & Romero-Fresco 2013, Perspectives: Studies in Translatology 21(1):5–21, DOI 10.1080/0907676X.2012.722651), and subtitles that duplicate the audio can raise cognitive load for ADHD learners (Lewis & Brown 2012 — a redundancy finding, cited for exactly that). Standard: phrase captions of 3–7 words, clause-bounded; word highlight only as a subtle in-phrase accent, hook line at most. Bouncing, rainbow, emoji karaoke is a slop tell (§07).
Standard
Endings loop, they don't close
Make the last line re-open the first — end on the question the opening answered, or zoomed on the document detail the opening showed wide. Average view duration over 100% of runtime is the loop signal. No “thanks for watching,” no end card, no logo outro: all three break the loop and waste the last 3 seconds.
Standard
Word-budget checker

Enter a runtime and a script word count. The verdict classifies by word count against the budget range for that runtime (150–165 wpm, less ~10% headroom). Pace in wpm is shown for information only — it does not decide the verdict.

Budgets — 30s: 68–74 · 45s: 101–111 · 60s: 135–149 words. Text-only variant: cap at caption pace (≈145 wpm, Jensema 1998) — ≤70 words at 30s.

§03 · Safe zones & text design — 1080×1920

No platform publishes a fixed safe zone for organic posts. Ad templates are published and are stricter — so this standard runs two tiers. Pixel values below are the keep-out zones: keep essential text and logos out of them.

Platform zoneTopBottomLeftRightSource / status
Meta Reels/Stories — ads (official)2696726565Meta Ads Guide percentages (14% / 35% / 6%)
TikTok — ads (official template, scaled)240660120120 + right railTikTok Ads Manager in-feed specs — template files, read at scale
YouTube Shorts — ads (Google template)28867248192Google Ads Help, Demand Gen specs (safe area 840×960 from y=288)
Reels — organic (measured)~220~430~50~140Creatora 2026 guide, measured
TikTok — organic (measured)~130~484~60~140Creatora 2026 guide, measured
Shorts — organic (measured)~380~380~60~120Creatora 2026 guide, measured
Strict box — prices, CTAs, anything promotable
x 120–888, y 288–1248 (768×960) — the intersection of the three ad templates. If a frame might ever be boosted, it is built to this box.
Standard
Organic working box — headlines & captions
x 65–940, y 250–1440, and no caption baseline below y=1440. The right edge stops at 940 because below y≈800 the right ~140px is action-rail territory. Shorts exception: the measured organic Shorts top keep-out (~380) is larger than this box's top (250) and larger than Google's ad template (288) — on Shorts specifically, keep essential text below y≈380 until a better measurement exists. The box does not silently satisfy the worst case on every platform.
Standard
Caption placement & logo
Captions sit centre-to-lower-middle, baseline around y 1150–1350 — above the description block on all three platforms, never bottom-anchored like TV subtitles. Logo top-left inside the safe box (≈x120, y300) or inside the headline lockup — lower-left collides with Shorts' subscribe block. Hook text must also survive the Reels 3:4 grid / 4:5 feed centre crop.
Standard
Type sizes
Headlines/hook text 72–110px bold sans; captions 54–72px; nothing essential under 48px; secondary labels ≥42px or deleted. Sans-serif only, bold weight for captions; line height 1.1–1.25 headlines, 1.2–1.4 captions. No thin weights, serifs, or decorative faces at caption sizes.
Standard
Caption line length — fit is physical
Applied limit: ≤28 characters per line at 54px, max 2 lines — ≈24 at 64px, ≈21 at 72px. The governing rule: a bold-sans character runs ≈0.56 × the font size, and the line must fit the 875px-wide working box (875 ÷ (0.56 × 54) ≈ 28). Reading speed ≤17 characters per second — deliberately stricter than the TV standard below, because Jensema's caption-comfort data puts difficulty from ≈170 wpm, and 17 CPS ≈ 170 wpm at average word length. Broadcast reference (16:9 TV — not the applied limit): Netflix's Timed Text Style Guide allows 42 characters per line at up to 20 CPS, written for a television screen; at vertical caption sizes a 42-character line runs ≈1,250–1,750px wide and physically does not fit the frame.
Standard
Contrast
WCAG AA is the floor — 4.5:1 for normal text, 3:1 for large; design to 7:1. White or brand-light text on a solid/semi-opaque dark box, or dark text on a solid light card — never bare text over a busy background. Stroke/shadow in addition to the box, not instead of it. Meaning never by colour alone. And the rule every source repeats: preview on a real phone, per platform, before a template locks.
Standard

The two boxes and the price pill, drawn at true 1080×1920 coordinates — the pill sits inside the strict box (y 1180–1244, text baseline y 1225), in the caption-baseline band. An earlier version of this diagram let the pill cross the strict box's bottom edge; corrected in the Nov 4 pass (see the corrections log).

§04 · Audio — muted-first, sound-rewarding
The muted evidence
69% of surveyed viewers watch video sound-off in public, 25% even in private; 80% say they're more likely to watch to completion when captions are available (Verizon Media + Publicis Media, Apr 2019, n=5,616 US adults — self-reported behaviour from two advertising companies: direction, not measurement). Facebook's internal tests (Feb 2016, no method published) found captioned ads averaged +12% view time. The famous “85% of video is muted” figure is 2016 muted-autoplay feed data — not a 2025 short-form fact, and it is not used as one here.
Survey
The sound-on counterweight
88% of TikTok users call sound essential and 73% say they stop and look at ads with audio (Kantar × TikTok — a survey commissioned by the platform, treated as directional). In the dedicated Shorts/Reels surfaces, sound-on is the norm once a viewer is in the surface. The same video is muted-first in a feed and sound-on in a swipe surface — both tracks must work.
Vendor survey
The dual-track standard
(1) Burned-in text carries 100% of the meaning — a muted viewer loses nothing factual. (2) Voiceover — human, or high-quality TTS with human editing of pauses, emphasis, and pronunciations — at 150–165 wpm, not reading every graphic label aloud. (3) A music bed under everything, licensed for business use on each platform; trending consumer audio is often unlicensed for business accounts. (4) Sound effects sparingly, as punctuation on reveals and beat changes. (5) Mix: voice clearly on top, music ~15–20dB under, master ≈ −14 LUFS (production convention, not a research finding). Flat, monotone synthetic voice is a retention and slop problem in itself (§07).
Standard
§05 · Per-platform renders

One master idea and script, four renders — different safe-box check, different title/caption conventions, clean watermark-free export each time. The identical file with the identical caption on all four reads as mass-produced on every surface.

YouTube Shorts

  • Signals: viewed-vs-swiped and held duration; search is a second, durable discovery path.
  • Length here: 30–50s, held.
  • Title: ≤100 characters, the search phrase a homeowner actually types front-loaded in the first ~40–60 visible characters. Hashtags in the description (3–5) — YouTube ignores all hashtags if more than 60 are used.
  • Trap: watermarked re-uploads are suppressed; the inauthentic-content policy is §07.

Instagram Reels

  • Signals: watch time, likes per reach, sends per reach (Mosseri, 22 Jan 2025). Sends weigh more for unconnected reach. (“Send = 3–5× a like” is folklore — no published weights.)
  • Length here: 25–45s. Design for the send — a checklist, a price range, a fact worth sending to one specific person.
  • Caption: first line = a second hook (it truncates under the video). Hashtags: maximum 5 hashtags per post or Reel — a platform cap announced 18 Dec 2025 and rolling out (the old limit was 30; hashtags do not meaningfully drive reach).
  • Cover: hook-safe in the centre 3:4 crop (feed shows 4:5, grid 3:4).

TikTok

  • Signals: completion + average watch time weigh heaviest; finishing a longer video is a strong interest signal.
  • Length here: 21–40s. Native, fast, sound-on, question-led; comments and questions drive distribution.
  • Caption: question or claim + local keywords; hashtags are a weaker signal than watch behaviour and the spoken/on-screen words. Pin a comment that asks the follow-up question.
  • Trap: business accounts can't use most trending consumer music; overly polished, TV-commercial creative gets skipped — TikTok's own guidance favours native.

LinkedIn video

  • Signals: dwell time, completion, comment quality; relevance and expertise over raw reach. Feed autoplay is silent — captions are the primary channel.
  • Length: guidance range 45–90s (60–90 cited for consideration/demo content); this pipeline's render standard is 45–75s. Under 30s for a single tip.
  • Post copy does the title's work: first 2 lines before “see more” = hook + claim; source any stat in the copy; 3–5 hashtags at the end; link in the copy or first comment.
  • Not a trap: external links are not the reach killer folklore claims — in Metricool's 2025 study of 577,180 posts from 47,735 company Pages, posts with links drew +4.9% impressions and +13.6% interactions. Native video still beats a video link.

Cadence: no reputable study establishes a universal best time or frequency for these niches. Consistency (e.g. 3–5/week per platform) and iterating on your own retention graphs beat any published “best time” list — treat such lists as unsourced.

§06 · Faceless formats that win — home services, construction, real estate

Ranked by combined support — platform evidence, named cases, and mechanism fit. These formats supply the specificity faceless content otherwise lacks; that is their whole advantage.

1 · Price transparency — line-by-line cost breakdown
The strongest niche signal. A remodeling firm's line-by-line “how much does it cost to build a basement” is described as its breakout video; its “Four Renovations That Add Zero Value” video is reported at 6.6M views, no ad spend, with an ad version generating 767 leads at $18/lead (LeadPhoenix — a single-source agency claim with no independent audit; read at the source page on Nov 4, 2026, and printed exactly as the page states it). The estimate is the visual — rendered document, progressive line highlighting. Secondary: home-service TikTok campaigns at $22–$34 cost per lead, with the honest caveat that TikTok leads close at ~10–15% vs ~30–40% for active-search leads.
Self-reported
2 · Myth-busting / the two-case story
Agent Shannon Gillette's Reel contrasting two sellers — priced-high-to-negotiate vs priced-to-generate-demand — held up by NOWBAM as a reusable framework for agents. No view count was published; the format is the evidence.
Practitioner
3 · Document walkthrough
Contract clause, permit requirement, inspection finding, warranty exclusion. Nobody without the document can copy it — that un-copyability is both the hook and the anti-slop proof. Redact personal data; zoom to the one line that matters; translate it into a dollar consequence.
Standard
4 · Process explainer
“What a home inspection actually checks, room by room.” One step = one card gives a natural beat structure, and checklists are highly sendable/saveable — the Reels signal that matters.
Standard
5 · Before/after — as data
Days on market, sale-to-list ratio, repair cost before vs after — numbers on the same branded chart. With no people footage, the data version is the honest version of the format, and numbers are more credible than a filtered photo.
Standard
6 · Local market update in numbers
“[Town] this month: 3 numbers that matter,” with a named source and date on screen. Hyper-locality is targeting, proof of expertise, and series-ability in one — and the format most likely to be sent to a specific neighbour or client.
Standard
7 · “What we found” — the failure reveal
The disaster under the floor, the panel that shouldn't have passed. Contractor-marketing sources claim finding content out-engages polished finished-work posts several-fold — an agency claim, presented as a hypothesis to test, not a benchmark. The mechanism is plausible: recognition (“that might be my house”).
Self-reported
§07 · Failure modes — how a short reads as slop
The scale of it
Of the first 500 Shorts shown to a fresh account, 104 (21%) were AI-generated slop and 33% “brainrot” (Kapwing, Oct 2025 — a single-account feed snapshot, not a platform share). Merriam-Webster named “slop” its 2025 Word of the Year: viewers now have a word for the pattern and skip on recognition.
Vendor data
The platform line — YouTube's inauthentic-content policy
Partner Program policy (renamed from “repetitious content,” 15 July 2025; clarified 16 July 2026): content must not be “mass-produced, generic, repetitive, or manipulative,” and the July 2026 clarification names three non-monetisable categories — (1) Generic or Repetitive Content (looks made with a template; interchangeable across uploads), (2) Unsatisfying or Off-putting Content (emotionally manipulative formulas; shock for its own sake), and (3) AI Personas Related to Sensitive Topics (AI personas presenting as human experts on health, legal, financial, or political topics — YouTube's examples name an AI “doctor” and AI podcast hosts offering financial guidance). In YouTube Help's own examples: “AI-generated content made with generic or unoriginal templates giving the impression of mass production without adding the creator's original, authentic insights or perspective.” A channel with too much of any category “will not be allowed to monetize” (YouTube Help's wording; TechCrunch reported the consequence for channels dedicated to this content as removal from the Partner Program, regardless of whether the content is AI-generated). Enforcement is channel-wide, judged on the overall pattern.
Platform
Why this pipeline sits close to the line
Faceless real-estate pricing and contract content neighbours categories (1) and (3) directly: template-identical pricing videos are category-1 risk; an AI persona presented as a pricing or legal authority is category-3 risk. AI narration and AI-assisted production are not themselves violations — sameness and fake expertise are. A programmatic pipeline is disqualified by being interchangeable, not by being faceless: every video needs a distinct, specific, human-originated claim and evidence. That is an editorial requirement, not a rendering setting — and it is why the anti-slop gate below carries the persona question.
Standard

The tells — read in under 3 seconds

  • Visual: default editor fonts and gradient backgrounds; identical layout, colours, and cadence in every video; slow pans over stills (the slideshow tell); walls of small text; text clipped by UI or the grid crop; a single wrong digit in a price breakdown destroys the format's whole claim to authority.
  • Structural: logo/greeting intro; claim delayed past 3s; one idea stretched over 60s; “wait for it” with a generic payoff; end card; every video the same length to the second.
  • Audio: monotone robotic TTS, mispronounced names and numbers; trending meme music under serious advice; complete silence under static text.
  • Content: vague claims with no number, place, date, or source; advice that could be about any business in any town; fake urgency; a CTA before any value was delivered.

What polished looks like

  • Specificity on screen: a real figure, a real document, a named locality, a dated source — visible, not just spoken.
  • Motion with meaning: highlights, zooms, and builds follow the argument. Decoration moves randomly; polish moves semantically.
  • Variation within a system: same type, colour, and voice identity — but a price breakdown does not look like a myth-bust.
  • Human fingerprints: a point of view, a recommendation, a “here's what I'd do.” Programmatic rendering, human judgment.
  • Craft basics done invisibly: proofread text, correct captions, consistent loudness, a clean loop. Slop is recognised by the accumulation of small carelessness.
The anti-slop gate — any “yes” = do not publish
  • ☐ Could this exact video be re-skinned for a different business or town by changing one word?
  • ☐ Is any frame a still image panning slowly with text over it — no build, highlight, or argument motion?
  • ☐ Are the template, length, cadence, and structure identical to the last 5 videos?
  • ☐ Is any claim unsourced, undated, or not locally specific? Is any number unverified?
  • ☐ Would a viewer in the trade call the voice, music, or wording generic?
  • ☐ Does an AI persona present as a human expert on a financial, legal, medical, or political topic? (YouTube's July 2026 category 3 — non-monetisable. AI narration of our own sourced material is fine; a synthetic “expert” persona is not.)
Production standards checklist — apply directly

Script

  • ☐ One claim per video. Hook line written first, containing a number, a named audience/place, or a myth to break — ideally two of three.
  • ☐ Hook fully on screen by 1.0s, fully readable by 2.0s, proposition stated by 3.0s — no logo, greeting, or title card before it.
  • ☐ Word budget at 150–165 wpm less ~10% headroom: 68–74 (30s) / 101–111 (45s) / 135–149 (60s). Text-only: ≤70 words at 30s (caption pace ≈145 wpm).
  • ☐ Mini-hook/open loop every 7–10s; peak fact at 60–75%; payoff delivered, then the loop line — no outro, no end card.
  • ☐ Every factual claim carries an on-screen number + source + date, or is deleted. Rendered text proofread character-by-character — especially prices.

Structure & pacing

  • ☐ Default runtime 28–45s (LinkedIn render 45–75s); 45–60s only for cumulative payoffs.
  • ☐ Visual state change on every new thought (typically every 2–5s); no static frame >6s; no beat >3s in the first 10s.
  • ☐ End frame/line flows into frame 1. Judge against the §01 length-banded targets and Shorts viewed-vs-swiped ≥70%; average view duration >100% of runtime is the loop signal.

Text & captions

  • ☐ Headline cards: 3–7 words, 1 idea, hold ≥1.0s.
  • ☐ Captions: ≤2 lines, ≤28 characters/line at 54px (scale down at larger sizes; the line must physically fit the 875px box), ≤17 characters/sec, clause-bounded breaks. Phrase-level by default; word highlight only as a subtle in-phrase accent.
  • ☐ Bold sans-serif. Headlines 72–110px, captions 54–72px, nothing essential <48px. Line height 1.1–1.25 / 1.2–1.4.
  • ☐ Contrast ≥4.5:1, design to 7:1 — text on a solid/semi-opaque box, plus stroke/shadow. Meaning never by colour alone.

Safe zones (1080×1920)

  • ☐ Strict box (prices, CTAs, anything promotable): x 120–888, y 288–1248.
  • ☐ Organic working box: x 65–940, y 250–1440; on Shorts, essential text below y≈380 (the measured Shorts top keep-out exceeds the box top); caption baseline y 1150–1350, never below y=1440.
  • ☐ Hook text survives the Reels 3:4 grid / 4:5 feed centre crop.
  • ☐ Logo top-left inside the box (never lower-left). Real-device preview on each platform before a template locks; re-check after UI updates.

Audio

  • ☐ Fully comprehensible muted (burned-in text = 100% of meaning) and rewarding sound-on.
  • ☐ Voice human or human-edited TTS: correct pronunciations, natural pauses, 150–165 wpm. Flat monotone TTS = reject.
  • ☐ Licensed business-use music bed under all renders; SFX only as punctuation. Voice ~15–20dB above music; master ≈ −14 LUFS.

Per-platform & measure

  • ☐ Clean, watermark-free export per platform — never cross-post a downloaded TikTok/Reel.
  • ☐ Shorts: title ≤100 chars, search phrase in the first ~40–60; 3–5 hashtags in the description (never >60 total); link the related long-form.
  • ☐ Reels: design for the send; first caption line = second hook; hook-safe cover; max 5 hashtags — platform cap announced Dec 2025, rolling out.
  • ☐ TikTok: proposition spoken + on screen in the first 3s; native, not TV-polished; pin a question comment.
  • ☐ LinkedIn: native upload; first 2 post lines = the hook; source your stat in the copy.
  • ☐ Measure your own: Shorts viewed-vs-swiped ≥70% (investigate <60%); Reels 3s skip rate <30%; TikTok average watch ÷ length >50%. Find the exact second of the retention cliff and fix that beat next time.
Could not verify — open questions
  • 1. No platform publishes a retention threshold, a ranking weight, or an organic safe-zone spec. The popular “send = 3–5× a like,” “85–92% watch muted,” and “change every 2–3 seconds” claims were not substantiated by any primary or controlled source found.
  • 2. Karaoke vs phrase captions: no controlled comparison exists — only vendor claims in favour, and accessibility research raising cognitive-load concerns. Test it on the pipeline's own retention graphs.
  • 3. Faceless niche case data is thin and self-reported (one agency's 6.6M-view / 767-lead claim — read at the source page Nov 4, 2026 and printed as the page states it; the viral agent Reel's agent is named in its source article, but no view count was published). No independent audit exists. Format rankings here are mechanism-based hypotheses to validate against your own analytics.
  • 4. LinkedIn sound-off percentage and best posting times for these niches: direction is consistent, exact numbers are weakly sourced and were not adopted as standards.
  • 5. Safe-zone pixel values disagree between reputable measured sources by 100px+ and change with app updates — the two-tier box in §03 is a risk-managed synthesis, not a guarantee.
Corrections — Nov 4, 2026 (three passes)

Published Oct 26, 2026. An external critique triggered a full re-verification on Nov 4, 2026 — every flagged item re-checked against the document itself, independent sources, or primary sources, never against this research's own notes. A second pass the same evening re-reviewed the page's checker and the YouTube policy; a third pass fixed micro-nits. Where the critique was wrong, that is recorded too. The log stays public — that is the deal.

Pass 1 — confirmed and fixed

  • Instagram hashtags were outdated. “3–5 max is current practice” missed the platform cap of 5 per post or Reel (announced 18 Dec 2025, down from 30). §05 and the checklist now state the cap.
  • Word budgets didn't match their own rate. The printed budgets matched a 140–160 wpm basis, not the stated 150–165 — neither the raw band figures nor the band less 10% headroom. Recomputed and printed with their arithmetic (§02, checklist): 68–74 / 101–111 / 135–149. The old page checker had the same defect baked into its code.
  • The caption line limit couldn't fit the frame. Netflix's 42 characters is a 16:9 TV standard; at vertical caption sizes the line physically overflows the 875px box. Applied limit is now ≤28 characters at 54px, scaled by size, with the fit rule stated.
  • “Median viewer gone by ~2s” withdrawn. Metricool's 3.75s is a mean across all videos; it supports a first-2–3-seconds hook gate, not a median claim.
  • Evidence labels were inflated. A single confidence grade has been replaced throughout by evidence-type labels: Verizon/Publicis = SURVEY, Kantar × TikTok = VENDOR SURVEY, Metricool = VENDOR DATA, Galloway/Gileta = PRACTITIONER, TikTok's 90% ad-recall figure = PLATFORM (ad context, not organic).
  • LinkedIn lengths contradicted each other (45–90 / 60–90 / 45–75 in three places). Now distinguished: guidance range 45–90s; the pipeline's render standard is 45–75s.
  • A checklist figure (“AVD >85%”) had no source and matched nothing in the research. Withdrawn; the checklist now points at the §01 length-banded targets.
  • ≤17 CPS is now explained: Netflix allows 20 CPS (≈200 wpm); Jensema's caption-comfort data puts difficulty from ≈170 wpm — so this standard caps at ≈170 wpm ≈ 17 CPS.
  • Organic-box gaps are stated, not hidden: the box's right edge is set by the action rails (written that way now), its top does not cover the measured Shorts keep-out (the Shorts exception is explicit in §03 and the checklist), and this page's own safe-zone diagram let its price pill cross the strict box's bottom edge — moved up 30px, into the box it teaches.
  • Secondary sourcing replaced at the load-bearing points: Meta safe zones → the Meta Ads Guide itself; TikTok template → TikTok Ads Manager specs; YouTube template → Google Ads Help; policy → YouTube Help; accessibility → the Lewis & Brown (2012) and Rajendran et al. (2013) primaries; skip rate → Instagram's Aug 2025 announcement and Meta's developer definition; Verizon → its May 2019 origin post, named.
  • Metricool's LinkedIn claim scoped: 577,180 posts from 47,735 company Pages (not personal profiles); link posts +4.9% impressions / +13.6% interactions.

Checked — the critique was wrong, or the document stands

  • Galloway/Gileta view figures: confirmed against a contemporaneous (Apr 2023) recap. The critique's 1.7M figure is the same study cut by video length; this research's 4.1M vs 1.8M is the cut by average view duration. Both cuts are now printed so they can't be conflated.
  • LeadPhoenix “6.6M views”: the critique couldn't fetch the page and suspected the URL tag. The page was fetched in this pass: its headline and body attribute the 6.6M views to the “Four Renovations” video and the 767 leads to its ad version, exactly as printed here. It remains labelled SELF-REPORTED.
  • Jensema 145 wpm is a real, citable study (1998, American Annals of the Deaf 143(4):318–324, n=578) — under-cited before, not unsourced; the full citation is now carried.
  • Lewis & Brown (2012) is a real paper — now cited for what it actually studied (see Pass 2).
  • Shannon Gillette's name is present throughout the cited NOWBAM article about her Reel. Not unsourced.
  • The EchoEtch length figure checks out (8,125 vs 3,457 median views/day, sub-60s vs 61–120s) and is now cited and labelled a single independent analysis.
  • Facebook's +12% captioned-ad figure is real — Facebook's own internal tests, announced Feb 2016 with its caption tool; no method published, muted-autoplay era, labelled as such.

Pass 2 — the checker's re-review (same evening)

  • Calculator verdict bug — introduced by Pass 1, and owned. Pass 1 corrected the checker’s budget constants but left its verdicts classifying by raw wpm, so a 72-word 30s script (in budget) was told “below pace” and an 80-word script (over budget) was told “in budget.” Verdicts now classify by word count against the budget range; wpm is display-only. The checker on this page is built that way: 30s/72 words = in budget · 30s/80 = over budget · 30s/65 (130 wpm) = below-pace verdict · 30s/50 (100 wpm) = under/air verdict.
  • YouTube policy updated to the July 2026 clarification (Pass 1 missed it entirely): the three named categories — Generic or Repetitive Content; Unsatisfying or Off-putting Content; AI Personas Related to Sensitive Topics — are in §07, with the AI-persona question added to the anti-slop gate.
  • The withdrawn policy “quote” — the checker corrected himself, and he is right. The sentence does exist in YouTube Help's examples, near-verbatim. The original sin was printing a slightly different wording as a quote; Pass 1 withdrew it outright; the official wording is now restored verbatim in §07 (“…without adding the creator's original, authentic insights or perspective”).
  • Lewis & Brown rescoped: it studied subtitles and ADHD learners (a redundancy finding) and no longer carries the fragmentation claim; Rajendran et al. (2013) carries segmentation, with its full citation and DOI.
  • Primary links completed (Meta Ads Guide page itself; the Verizon origin post; Facebook's +12% given a sources entry via its Feb 10, 2016 announcement as reported by MarTech; the skip-rate entry given a URL), and the Instagram cap wording softened — announced 18 Dec 2025 and rolling out gradually; the operative rule (max 5) is unchanged.
  • The Shorts top-keep-out exception, stated four times on the old page, now appears twice — in the §03 rules and in the checklist.
  • EchoEtch was flagged by the checker as unchecked by him; it was read in Pass 1 and stays labelled a single independent analysis.

Pass 3 — micro-nits (late, same day)

  • The TechCrunch sources entry now carries its URL.
  • The skip-rate source of record is Social Media Today's own article (“Instagram Adds Retention Metrics for Reels,” 24 Aug 2025), with Meta for Developers' reels_skip_rate definition as the primary for the metric itself. (That article's URL does not surface in search results; it is cited by title and date rather than by a guessed link.)
  • “Removed from the Partner Program” was replaced with YouTube Help's own wording (“will not be allowed to monetize”); the removal phrasing is now attributed to TechCrunch's reporting, where it came from.
Sources

All first read 25–26 Oct 2026 via web index; the load-bearing ones were re-checked against primary or independent sources on Nov 4, 2026. Labels are evidence types, as used throughout this document.

← All LinkLab research