Bold-claim openings are associated with higher sales per view in Womenswear & Underwear, while format and production differences remain within noise

1. Introduction

Womenswear & Underwear sellers on TikTok Shop face a practical question: given a product and a short video, which creative choices are associated with turning views into sales? Raw sales figures cannot answer this, because attention on short-video platforms is long-tailed and early popularity strongly predicts later popularity [1]. A video that sells a lot may simply have been seen a lot. We therefore focus on sales per view, which separates how well a video converts from how far it travels.

Prior work on short-video popularity prediction has made substantial progress with multimodal models [2], and the current state of the art finds that creator-level features carry most of the predictive signal [3]. That literature predicts views and engagement rather than commerce, and it suggests that content features alone should be expected to explain only part of performance. This paper examines the creative features within a single category, openings (hooks), formats and production, and adds the category's seasonal pattern to frame the months ahead.

2. Data

The sample comprises 586 top-selling Womenswear & Underwear videos on TikTok Shop observed between 2 September and 1 October 2026, a window of about one month. Of these, 165 were analysed in detail; across all categories, the wider market comparison set contains 3,340 analysed videos.

For each analysed video we recorded the opening hook type, the overall format, the production style, and sales and view outcomes. Hook and format categories with few videos (six to eight) are retained but flagged as low confidence. The seasonal index is the category's monthly sales relative to its own average month (100 = a normal month).

The short window is an important constraint: results describe one month of top sellers and may not generalise to other periods, particularly ahead of the seasonal swings described below.

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: 586 videos tracked in Womenswear & Underwear, 165 of them watched and analysed; the market baseline uses all 3,340 watched videos. Predictions come from a logistic model retrained daily and tested on recent weeks it had not seen.

3. Methods

For each creative attribute we compute a sales-per-view index relative to the market, where 1 equals the market average. Because many groups contain few videos, raw ratios are unstable; we apply empirical-Bayes shrinkage that pulls small-group estimates toward the market mean in proportion to their uncertainty [4]. We report 90% intervals for each index and grade confidence as high, medium or low based on group size and interval width. Share of sales is each group's percentage of sales among analysed videos.

Separately, we fit a predictive model of whether a video is a top seller, using creative attributes and their pairwise interactions, trained on 2,667 rows from the wider market. The model is evaluated on a time-split hold-out (trained on earlier videos, tested on later ones) and summarised by the area under the ROC curve, a measure of how well it ranks sellers above non-sellers [5]. The AUC range is a bootstrap 90% interval [6]. The model was not restricted to analyses that never saw sales (cleanOnly = false), so some leakage of outcome information into the features cannot be ruled out.

4. Results: openings

Bold-claim openings are the clearest signal in the category. Across 37 videos they sell at 1.3× the market average (90% range 1.05–1.60, high confidence) and account for 26% of analysed sales, the only hook whose interval sits entirely above 1. Relatable POV openings are the most common hook (40 videos) and carry a similar 26% share, but sell at roughly the market average (1.03×, range 0.85–1.24).

Question openings show the highest point estimate (1.36×) but rest on six videos with a range of 0.88–2.08, so the evidence is weak. At the other end, result-first openings (0.77×, range 0.56–1.06, seven videos), trend or sound openings (0.83×, 0.61–1.13) and going straight into the demo (0.86×, 0.65–1.14) sit below average, though every one of these intervals includes 1.

5. Results: formats and production

Format differences are small relative to uncertainty. Reviews and testimonials are the largest format, with 49 videos and 31% of analysed sales, selling at 1.06× (range 0.89–1.26). Routine / GRWM (1.06×, 0.79–1.41) and talking head (1.03×, 0.79–1.33) are similar. Demo or tutorial videos take 22% of sales but sell at 0.92× (0.73–1.16). Before-and-after (0.87×) and slideshow (0.87×) formats sit lowest, both with intervals spanning 1.

Production is dominated by filmed live action, which accounts for 146 analysed videos and 96% of sales at the market average (1.01×, 0.91–1.12). Photo slideshows (10 videos) sell at 0.88× (0.65–1.18).

In the cross-market model, filmed live action has the largest positive association with being a top seller (odds ratio 2.54), followed by a visual pattern-break combined with a review or testimonial (1.92), no visible AI (1.71) and boosting with ads (1.59). Showing the product in the first second (0.62), an education angle (0.59) and a product close-up paired with a bold claim (0.60) are associated with lower odds. The model's AUC of 0.64 (range 0.57–0.70) indicates modest ranking ability, and because it was not trained only on sales-blind analyses these odds ratios should be read as indicative.

6. Results: the season ahead

The category's seasonal index shows October at about a normal month (102), followed by a drop to 86 in November, the lowest month of the year, and a partial recovery to 94 in December. Sales then rise through January (103) and February (108) to a peak of 122 in March.

A secondary high appears in June (112), while July and August (both 93) and September (96) run slightly below average. The pattern suggests that the creative evidence gathered in September comes from a slightly below-average month, ahead of a soft late autumn.

7. Discussion

For creators and brands, the most defensible takeaway is that bold-claim openings are associated with higher sales per view in this category, and they already carry a quarter of analysed sales. Beyond that, the leading formats, reviews, GRWM and talking head, sell close to the market average, and the gaps between formats are within noise. Live-action filming is the norm rather than a differentiator within the category, although the cross-market model associates it with higher odds of being a top seller.

These results are consistent with prior work finding that content features explain only part of short-video performance, with creator features carrying most of the signal [3]. A modest AUC is what that literature would lead us to expect from creative attributes alone.

Several alternative explanations apply. Bold claims may be used more often by experienced sellers or for products that already convert well, so the hook may proxy for creator or product quality. Ad boosting, which the model associates with higher odds, could also concentrate on particular hooks. And because the sample is top sellers in a single month, selection on success may compress differences between groups.

8. Conclusion

In one month of top-selling Womenswear & Underwear videos, bold-claim openings sell at about 1.3× the market average (range 1.05–1.60), while formats and production styles show no difference clearly separated from the average. Several promising signals, notably question openings, rest on too few videos to act on.

With November historically the category's weakest month (86) and March its strongest (122), these findings should be revisited over a longer window and with a model trained only on sales-blind analyses before they are treated as stable.