Price and bold-claim hooks sell more per view on TikTok Shop, while borrowed openers lag

1. Introduction

Short-form commerce video is judged in its opening moments. If a viewer scrolls past in the first two seconds, nothing later in the video can convert them. This paper asks a narrow question: among top-selling TikTok Shop videos, which hook types and which hook sources are associated with more sales per view? Hook sources include footage filmed for the video, borrowed clips, stock footage and AI-generated clips.

The question matters because hook choices are cheap to change and easy to test. They are also increasingly contested, as borrowed clips and AI-generated footage become easier to drop into an opener. Creators and brands need to know whether these shortcuts sell as well as filmed material.

Prior work focuses mostly on attention rather than sales. Online attention is long-tailed, and early popularity is a strong predictor of later popularity [1]. Multimodal benchmarks show that micro-video popularity can be predicted from content signals [2]. The current state of the art finds that creator-level features carry the most predictive weight [3]. Less is known about whether content choices in the first seconds relate to conversion, which is the gap this paper addresses.

2. Data

The sample comprises 13,309 top-selling TikTok Shop videos across 29 product categories, collected between 2 September and 1 October 2026. Of these, 3,340 were watched and analysed in detail. All results below refer to this watched subset.

For each watched video we recorded the following: - the hook type, meaning what the first two seconds do (for example a price or deal, a bold claim, a question or a problem call-out); - the hook source, meaning where the opening footage comes from; - the overall video format; - views and sales.

Shares of sales are expressed as a percentage of total sales across watched videos.

The window is short, at 30 days, and the sample consists of top sellers rather than all videos. Some labels are rare. AI-generated hooks (20 videos), text cards (12) and stock footage (10) are examples, and estimates for these groups are flagged as lower confidence.

Daily rankings of top-selling TikTok Shop videos, products and categories in the US were collected from several independent sources and merged into one record per video. Where sources overlap, their sales estimates are cross-checked against each other, and videos whose estimates disagree by more than half are flagged. An AI analyst watched each video and recorded its opening, format, angle, production style, use of AI and hook source, without seeing how the video sold. Sales per view is a video group's revenue per view divided by the market's, shrunk toward the average for small groups, with 90% ranges. This paper covers 2026-09-02 to 2026-10-01: 13,309 videos tracked, 3,340 of them watched and analysed. Predictions come from a logistic model retrained daily and tested on recent weeks it had not seen.

3. Methods

Our main measure is a sales-per-view index. For each group, sales divided by views is compared with the market rate across watched videos, so that 1 equals the market average and 1.5 means 50% more sales per view. Small groups produce noisy ratios. We therefore shrink each group's estimate toward the market mean using empirical-Bayes methods [4], so that groups with few videos are pulled closer to 1. We report 90% intervals for every index and a confidence grade that reflects group size and interval width.

Separately, we fit a predictive model of whether a video is a top seller. It uses hook, source, format, category and interaction features and was trained on 2,667 rows. The model is evaluated on a time-split test, training on earlier videos and testing on later ones, to avoid look-ahead. Performance is summarised by the area under the ROC curve (AUC), a ranking measure where 0.5 is chance [5]. The AUC range is a bootstrap 90% interval [6].

The model was not restricted to analyses that never saw sales figures (cleanOnly is false). Some leakage of outcome information into the features therefore cannot be ruled out. Model odds ratios are descriptive associations, not effects.

4. Results: What the hook says

Hooks that lead with an offer sold most efficiently. Price-or-deal hooks reached 1.53× market sales per view (90% range 1.37–1.71, 193 videos). Bold claims reached 1.28× (1.20–1.37, 478 videos) and questions 1.27× (1.13–1.43, 184 videos). Bold claims are also widely used, carrying 16% of watched-video sales.

The most common hooks sat near the market average. Relatable POV openers carried the largest share of sales, at 20%, and sold at 1.05× (0.99–1.11). Problem call-outs carried 14% of sales and sold at 1.00× (0.93–1.08). Social proof, result-first and story openers had intervals that comfortably include 1.

Two product-forward openers sat below the market. Going straight into the demo sold at 0.89× (0.80–0.98, 242 videos). Unboxing reveals sold at 0.88× (0.76–1.02, 122 videos), with a range that only just reaches the market average. List or countdown hooks (20 videos) are too few to read.

5. Results: Where the opening footage comes from

Footage filmed for the video dominates. It accounts for 1,979 watched videos and 89% of watched-video sales, and sells at 1.01× (0.98–1.05), effectively the market rate. Product close-up openers sold at 1.15× (1.00–1.32, 147 videos), a modest edge whose lower bound sits at the market average.

Borrowed clips from TV, film, news or viral sources sold at 0.76× (0.60–0.97, 41 videos). The residual 'other' source category sold at 0.65× (0.54–0.79, 37 videos). Stock footage pointed in the same direction at 0.78×, but it rests on only 10 videos and its range (0.58–1.04) includes the market average.

AI-generated hook clips showed a point estimate of 1.37×. This rests on just 20 videos, with a range of 1.00–1.87 that reaches down to the market average. We do not treat it as evidence that AI openers outperform. In the predictive model, 'no visible AI' carries an odds ratio of 1.71 for being a top seller, which points the other way. Text cards (12 videos, 1.03×) are similarly inconclusive.

6. Results: Formats and what the model picks up

Several formats sold above the market rate: - unboxing, 1.37× (1.20–1.57); - voice-over B-roll, 1.29× (1.18–1.41); - review or testimonial, 1.27× (1.17–1.37); - talking head, 1.18× (1.09–1.28).

Demo or tutorial is by far the most common format, with 1,183 videos and 34% of watched-video sales, yet it sold at 0.92× (0.88–0.96). Skits sold at 0.86× (0.77–0.97). An unboxing format sells well, while an unboxing-reveal hook does not. This suggests the opener and the overall structure play different roles.

The predictive model reached an AUC of 0.641 (bootstrap 90% range 0.573–0.696) on time-split testing. This is a real but modest ranking signal, and it was not trained on clean analyses only. Its strongest positive associations were: - filmed live action (odds ratio 2.54); - a visual pattern-break combined with a review or testimonial (1.92); - no visible AI (1.71); - before and after combined with filmed live action (1.70).

Its negative associations were: - product in the first second (0.62); - product close-up combined with a bold claim (0.60); - education (0.59).

Being boosted with ads also carried an odds ratio of 1.59. This is a reminder that paid distribution is entangled with the outcome.

7. Discussion

For creators and brands, the pattern is consistent. Openers that give the viewer a reason to care, such as a price, a claim or a question, are associated with more sales per view. Openers that simply start showing the product are associated with fewer. This fits the model's lower odds for having the product in the first second, and the below-market rate for straight-into-the-demo hooks. Filmed footage is the norm and sells at market rate. Borrowed clips lag, which suggests that attention captured with unrelated material may not carry through to purchase.

These results extend prior work, which mostly predicts attention [1, 2]. They show that content choices in the opening seconds also relate to conversion efficiency. The model's modest AUC is consistent with the finding that creator features, which we do not centre here, carry most predictive weight [3].

Alternative explanations deserve weight. Price-or-deal hooks may cluster in categories or price points that convert well regardless of the opener. Creators who use borrowed clips may differ in audience or experience. Ad boosting is associated with top-seller status and may be correlated with hook choice. The sample also includes only top sellers, so the comparisons are within winners rather than across all videos.

8. Limitations

Most videos in the sample are top-ranked, so findings mostly separate strong sellers from good ones rather than from all videos; typical and weak videos are being added. Results are associations, not causes. Paid promotion, creator audience size and product price are not fully controlled for. Revenue and views are estimates, not figures reported by TikTok. The evidence covers a single 30-day window (2 September to 1 October 2026) and only top-selling videos, so results describe differences among winners rather than across all TikTok Shop content. Several groups are small and medium or low confidence: AI-generated hooks (20 videos), text cards (12), stock footage (10), list or countdown hooks (20) and day-in-the-life format (8). The predictive model was not restricted to analyses that never saw sales (cleanOnly is false), so some outcome leakage is possible. All figures are observational associations and may reflect category mix, price point, creator differences or ad boosting rather than the hook itself.

9. Conclusion

In one month of top-selling TikTok Shop videos, hooks that lead with a price, a bold claim or a question sold at roughly 1.3–1.5× the market's sales per view. Going straight into the demo was associated with about 0.9×. Footage filmed for the video carries 89% of sales at market rate, while borrowed clips sit at about 0.76×.

Evidence on AI-generated and stock hooks is too thin to support conclusions. The predictive model offers only a modest signal (AUC 0.64, range 0.57–0.70). Repeating the analysis over longer windows, with models trained only on analyses blind to sales, is the natural next step.