Video A reports a 20% hook rate and a 70% hold rate. Video B reports a 50% hook rate and a 20% hold rate. Which video has the bigger problem—and which part should you edit first?
Under one explicit measurement convention, A is losing viewers before the three-second mark, while B is losing them after it. But that conclusion is only useful after you verify the formulas. In the worked example below, A actually produces more 15-second views than B, despite its much weaker hook rate.
That is the central lesson: hook rate and hold rate can separate early loss from later loss, but only when you write down the numerator, denominator, threshold, and counting rules. They identify a measured stage to inspect. They do not reveal an exact drop-off second, prove why people left, or tell you that an edit will increase sales.
Hook rate and hold rate answer different questions
“Hook rate” and “hold rate” are often custom labels rather than universal platform metrics. A practical convention is:
- Hook rate: early qualified video views divided by impressions.
- Conditional hold rate: later qualified video views divided by early qualified video views.
- Late-view rate: later qualified video views divided by impressions.
For a 30-second video using three and 15 seconds as the two milestones:
| Metric | Formula | Question it answers |
|---|---|---|
| Hook rate | 3-second views / impressions |
How often did delivery become an early qualified view? |
| Conditional hold rate | 15-second views / 3-second views |
Of the early viewers, how many reached the later milestone? |
| Late-view rate | 15-second views / impressions |
How much later viewing did the delivered impressions produce overall? |
The third rate matters because a conditional percentage can improve while the number of later viewers falls. When the two view counts are compatible, nested duration events from the same reporting scope, the relationship is:
late-view rate = hook rate × conditional hold rate
The labels alone are not enough. Motion’s April 2026 glossary defines hold rate as ThruPlays divided by three-second video plays. Foreplay’s creative-funnel guide calls ThruPlays divided by impressions “Watch Rate / Hold Rate.” Both calculations can be useful, but they answer different questions. Comparing the percentages as though they were the same metric is a denominator error.
Motion’s broader Metrics Cheat Sheet makes the distinction visible in another way: it discusses ThruPlay and a separate 15-second/3-second retention measure. The safest operating rule is simple: put the formula next to the label every time you report it.
Build a metric fingerprint before comparing videos
Before you diagnose the creative, record the measurement contract. At minimum, preserve:
| Field | What to record |
|---|---|
| Platform and placement | For example, platform plus Feed, Reels, Shorts, or in-stream—not one blended “social” row |
| Reporting period and export date | The dates being compared and when the report was pulled |
| Video length | The original asset length and, when different, the served length |
| Early event | Exact field name, threshold, and whether interaction can qualify it |
| Later event | Exact field name, threshold, completion rule, and whether it is nested inside the early event |
| Numerator and denominator | Raw counts, not only a rounded percentage |
| Counting scope | Impression sessions, player starts, people, paid views, replays, or another unit |
| Delivery context | Objective, audience, optimization, and placement mix |
This is not paperwork for its own sake. Current platform documentation shows why the fields matter.
TikTok’s Video play metrics, last updated in November 2025, says a two-second view is counted within an impression session and excludes replays. If someone swipes away and back, TikTok describes that as a new impression session and a new view. Its six-second view can qualify through six seconds of playback, full playback of a shorter video, or an engagement within the first six seconds. Average play-time metrics include replay time, while its quartile view counts exclude replays. A “six-second view” therefore is not always a pure six-second survival event.
TikTok’s August 2026 focused-view objective documentation adds objective-specific routes, including short-video completion, interaction, and an accelerated-view condition. Do not silently apply those rules to a different field or reporting period.
Google also defines a paid view differently by format. In About YouTube ads and view metrics, a skippable in-stream TrueView view can be triggered by 30 seconds, the end of a shorter video, or a click. An in-feed view can come from a thumbnail click or ten seconds of inline autoplay. A Shorts view can come from ten seconds, completion of a shorter video, or specified interactions. Google also says that swiping away and back to the exact same Shorts ad does not add another impression or view. These rules are not interchangeable with TikTok’s.
For Meta, use the current definition visible in your reporting surface or an accessible current official reference. The official Meta Python Business SDK field list includes fields for video plays, quartiles, play curves, and ThruPlay actions, but that code list does not define the current short-video, replay, or completion semantics. This article therefore does not turn a third-party formula into a universal Meta rule.
A short-video threshold can run in the wrong direction
Suppose a six-second video has 600 views at 25% and 500 views at two seconds. The 25% milestone occurs at 1.5 seconds, so 600 / 500 = 120% is arithmetically possible. It is not evidence that retention increased after two seconds. The numerator represents an earlier event.
The quarter mark is later than two seconds only when the video is longer than eight seconds. It is later than three seconds only when the video is longer than 12 seconds. Translate every percentage milestone into seconds before calling a ratio “hold.”
Two viewing patterns with the same denominator
The following examples are original synthetic teaching data, not campaign results. Both videos are 30 seconds long, receive 10,000 impressions, and use the same first-pass, continuous-playback counting contract. Replays, seeking, interaction shortcuts, and accelerated completion are excluded. The lines in the chart connect measured milestones as a visual guide; they do not invent observations between them.
| Milestone | Pattern A count | A as % of impressions | Pattern B count | B as % of impressions |
|---|---|---|---|---|
| Playback starts | 8,000 | 80% | 8,000 | 80% |
| 3 seconds | 2,000 | 20% | 5,000 | 50% |
| 7.5 seconds | 1,700 | 17% | 2,500 | 25% |
| 15 seconds | 1,400 | 14% | 1,000 | 10% |
| 22.5 seconds | 1,000 | 10% | 600 | 6% |
| 30 seconds | 800 | 8% | 400 | 4% |
Pattern A: investigate the opening first
Pattern A’s hook rate is 2,000 / 10,000 = 20%. Its conditional hold rate is 1,400 / 2,000 = 70%, and its late-view rate is 1,400 / 10,000 = 14%.
The largest measured loss is between playback start and three seconds: 6,000 of 8,000 starts disappear, a 75% loss from that stage. From three to 15 seconds, the loss is 600 of 2,000 early views, or 30%.
That pattern justifies inspecting the opening before rewriting the whole video. Check whether the product, problem, or promise is legible without sound; whether setup delays the point; whether the first frame resembles the content that follows; and whether a placement or autoplay difference explains some of the start-to-three-second gap. “Delayed product clarity” is a candidate explanation only if the footage supports it.
Pattern B: investigate the bridge from hook to proof
Pattern B’s hook rate is 5,000 / 10,000 = 50%. Its conditional hold rate is 1,000 / 5,000 = 20%, and its late-view rate is 1,000 / 10,000 = 10%.
It loses 3,000 of 8,000 starts before three seconds, or 37.5%. It then loses 4,000 of 5,000 early viewers before 15 seconds, or 80% of the early cohort. The opening earns attention, but most of that audience does not reach the later milestone.
Inspect the connection between the opening and the body. Does the hook promise one outcome while the demonstration delivers another? Does the proof arrive after excess setup? Is the demonstration difficult to follow? Does a transition interrupt continuity? Those are editing hypotheses, not conclusions encoded in the percentage.
Notice the result that a one-metric ranking would miss: A has the lower hook rate and the higher conditional hold rate, yet it generates 1,400 15-second views. B generates only 1,000. The impression-based late-view rate exposes the difference.
A better hold rate can coexist with fewer later viewers
Consider two more 30-second scenarios with the same 10,000-impression denominator:
| Scenario | 3-second views | 15-second views | Hook rate | Conditional hold | Late-view rate |
|---|---|---|---|---|---|
| X | 4,000 | 1,600 | 40% | 40% | 16% |
| Y | 2,000 | 1,200 | 20% | 60% | 12% |
Y’s conditional hold is 20 percentage points higher, a 50% relative increase. But it produces 400 fewer 15-second views, a 25% decline. The arithmetic is consistent: 40% × 40% = 16%, while 20% × 60% = 12%.
This does not make conditional hold useless. It makes it a cohort question. Use it to understand what happened after the early threshold, then pair it with hook rate and late-view rate to understand the whole delivered funnel.
Run three data checks before ordering an edit
1. Separate zero, missing, and tiny denominators
| Reported counts | Hook rate | Conditional hold | What you can say |
|---|---|---|---|
I=0, V3=0, V15=0 |
N/A | N/A | There was no delivery; no creative diagnosis is available. |
I=1,000, V3=0, V15=0 |
0% | N/A | No early-view denominator exists. Hold is not 0%. |
I=1,000, V3=2, V15=1 |
0.2% | 50% | The observed rate is 50%, but one additional early view moves it to 66.67% if it also reaches the later milestone, or 33.33% if it does not. |
I=1,000, V3=2, V15=null |
0.2% | N/A | A missing later field is not a behavioral zero. |
A 50% hold rate based on one of two early views should not carry the same decision weight as 1,000 of 2,000. There is no universal minimum sample in this article because repeated exposure, delivery allocation, and user dependence affect uncertainty. The operational point is narrower: inspect the raw denominator before treating a percentage as stable enough to brief an edit.
2. Break out placement before trusting the total
Placement mix can reverse the aggregate story. In this synthetic example, every placement’s conditional hold improves by ten percentage points, but the combined rate falls because delivery shifts toward the lower-hold placement.
| Period | Placement | Impressions | 3-second views | 15-second views | Conditional hold |
|---|---|---|---|---|---|
| 1 | Feed | 9,000 | 2,700 | 1,350 | 50% |
| 1 | Reels | 1,000 | 100 | 20 | 20% |
| 1 | Total | 10,000 | 2,800 | 1,370 | 48.93% |
| 2 | Feed | 1,000 | 300 | 180 | 60% |
| 2 | Reels | 9,000 | 900 | 270 | 30% |
| 2 | Total | 10,000 | 1,200 | 450 | 37.50% |
The total is weighted by early-view counts, not by impressions and not by a simple average of the placement percentages. Before you roll back a creative, compare like-for-like placement rows and inspect how the denominator composition changed.
3. Audit the export grain
Use raw counts at the same ad, placement, date, objective, and video-length grain. Remove subtotal or TOTAL rows before summing. Treat null as missing, not zero. Do not divide a paid-only early count by an organic-inclusive later count. Do not pool quarter milestones from six- and 30-second videos as though both represent the same elapsed time.
If the export contains only a rounded “hold rate” and not its numerator and denominator, you cannot reconstruct or validate the exact calculation. Request the raw fields instead of manufacturing counts from a displayed percentage.
Milestones identify a stage, not an exact drop-off second
Pattern B loses 2,500 events between three and 7.5 seconds and another 1,500 between 7.5 and 15 seconds. Those milestone counts do not reveal whether the first loss happened at 3.1 seconds, 7.4 seconds, or gradually across the interval. Drawing a smooth curve through the points does not create analytics that were never collected.
Average watch time has the same limitation. Imagine 100 starts on a 30-second video:
- In distribution X, every start watches 16 seconds: 1,600 total seconds, 16-second average, 0% completion.
- In distribution Y, 50 starts watch all 30 seconds and 50 watch two seconds: also 1,600 total seconds and a 16-second average, but 50% completion.
“Average watch time is 16 seconds” is not proof that the problem is at second 16.
Open a real retention or key-frame report when you need finer timing. YouTube’s audience-retention documentation says dips can mean viewers stopped or skipped, while spikes can reflect watching, rewatching, sharing, or a section that was unclear enough to replay. It also warns that a segment’s absolute views can exceed the video’s overall view count when portions are watched multiple times. TikTok’s Video Insights documentation offers second-by-second Key Frame Analysis, but labels those values as estimates that can differ from summary metrics because of attribution, privacy, app-version, and technical conditions.
A curve narrows the time window. The footage, delivery context, and a controlled comparison are still needed to investigate the cause.
Turn the diagnosis into one editing hypothesis
Use the metrics to decide what to inspect, then verify the candidate issue in the actual creative. The worksheet below is filled rather than left as a generic template. Download the filled funnel worksheet or use the video funnel calculator.
| Measured pattern | Verify in the footage | One-variable edit hypothesis | What must stay comparable | What would weaken the hypothesis |
|---|---|---|---|---|
| Pattern A: low hook, stronger conditional hold | Product or promise is genuinely unclear in the opening | Move existing product/demo information into the first three seconds; keep the later story intact | Length, offer, audience, placement, objective, metric contract | Opening was already clear; start rate or placement explains the gap; hook rises while 15-second views per impression fall |
| Pattern B: strong hook, weak later continuation | Hook promise and 3–15s payoff do not match | Keep the opening; reorder proof so it answers the hook sooner | Opening, length, offer, placement, audience | Early click-out is desirable; delivery mix changed; late-view rate does not improve |
| Localized decline during a demonstration | Viewers cannot tell what action or result the demo proves | Simplify the demonstration sequence or add only the missing context | Hook and surrounding shots | Rewatch spike reflects interest rather than confusion; decline persists away from the demo |
| Stable opening followed by setup-heavy decline | Several seconds pass before new proof or value appears | Compress setup and bring the strongest proof earlier | Core claim and total test scope | Setup is necessary for comprehension; shorter cut changes the threshold definition and is not analyzed separately |
| Dip adjacent to a transition | The transition obscures the product, text, or causal sequence | Replace that transition only | Adjacent shots, audio, copy, duration | No real timestamp-level dip exists; the issue appears across multiple unrelated moments |
Define the test outcome before delivery. For an opening change, examine hook rate and later views per impression, not hook alone. For a body change, examine conditional hold and later views per impression while confirming the opening and delivery conditions stayed comparable. Keep the business outcome—purchase, qualified visit, install, lead, or another chosen result—beside the attention metrics. A re-edit is a hypothesis until it is actually compared.
Benchmarks are context, not an edit command
A benchmark can help you spot an unusual result only when its formula and population resemble yours. Billo’s H1 2026 hook-rate report reported a 25.44% average across 88,329 Meta sales-objective video ads with more than 1,000 impressions, using three-second plays divided by impressions. The public method identifies the period and inclusion rule, but not the country mix, video-length mix, placement mix, number of advertisers, or averaging weight.
That makes it a documented vendor-sample reference, not a universal target. Your own like-for-like history may be more useful than a cross-industry average. A benchmark also cannot determine whether a low hook is caused by the opening, the audience, the placement, the objective, or a change in delivery.
Most importantly, watch time is not proof of sales. A high hook can attract curiosity without producing qualified action. A click can also shorten video viewing because someone leaves for the landing page; Motion’s metric guide explicitly notes this as one possible reason a high-CTR asset may show lower ThruPlay. Choose the business winner with the business outcome, then use viewing metrics to understand what the creative appears to be doing along the way.
The practical diagnostic sequence
When a video looks weak, do not begin with “make the hook punchier.” Begin with a measurement contract:
- Write the hook and hold formulas with their raw counts.
- Confirm the later event is actually later, compatible, and nested inside the early event.
- Preserve platform, placement, reporting dates, video length, objective, and counting rules.
- Calculate hook rate, conditional hold, and late-view rate together.
- Check zeros, nulls, tiny denominators, subtotal rows, and placement mix.
- Use milestones to choose a stage; use a real curve to narrow the time window.
- Inspect the footage and form one falsifiable editing hypothesis.
- Evaluate the edit with both attention metrics and the chosen business outcome.
Hook rate asks whether delivery became an early view. Conditional hold asks whether those early viewers continued. Late-view rate reconnects the later milestone to all delivered impressions. Together, with explicit definitions, they tell you where to look next—without pretending the ratio has already found the cause.
Sources
- “What is Hold rate?” — Motion, updated April 17, 2026
- “Metrics cheat sheet” — Motion Help Center, April 15, 2026
- “How to See Ad Performance in Meta Ads Manager” — Foreplay
- “Video play metrics” — TikTok for Business, updated November 2025
- “About the video views objective and focused view optimization” — TikTok for Business, updated August 2026
- “About video insights in TikTok Ads Manager” — TikTok for Business, updated February 2025
- “About YouTube ads and view metrics” — Google Ads Help, accessed September 8, 2026
- “Measure key moments for audience retention” — YouTube Help, accessed September 8, 2026
- Meta Python Business SDK,
adsinsights.pyfield list — Meta - “Video Ad Hook Rate Benchmarks H1 2026” — Billo, August 6, 2026
