ContentIQ working paperCIQ-WP-2026-55TikTok ShopVersion 118 Oct 2026
TikTok Shop agency creative testing: a practical framework
51,843 videos tracked, 9,285 analysed in depth, 29 categories, 19 Sept–18 Oct 2026
MediaLabs Research · in collaboration with Coherence Limited
Key findings
- 011.5×sales per view for price-or-deal hooks against the market (90% range 1.37–1.58, 492 videos)
- 021.2×sales per view for voice-over B-roll (1.16–1.33) and review or testimonial formats (1.12–1.22)
- 030.96×sales per view for demo or tutorial, the largest format, with 36% of watched-video sales (0.94–0.99)
- 041.04× vs 0.93×sales per view with and without a face in the first second (1.02–1.07 vs 0.91–0.96)
- 051.07× vs 0.92×sales per view with all three of face, text and product in the first second, against only one (1.03–1.10 vs 0.88–0.96)
Contents
Abstract
Brands deciding how to brief an agency or creators on TikTok Shop need a way to test creative that separates what works from who made it. In our analysis of US TikTok Shop videos, price-or-deal hooks sold at 1.5× the market's sales per view (90% range 1.37–1.58) and voice-over B-roll formats at 1.2× (1.16–1.33), while the most common format, demo or tutorial, sold at 0.96× (0.94–0.99). These are observational comparisons, not experiments, because product, offer, creator and distribution were not held constant. We turn them into hypotheses and propose a controlled test plan that keeps product, offer and distribution fixed.
1. Introduction
Short answer: if you are choosing an agency or creator partner to run TikTok Shop creative tests, ask for a controlled design in which one creative element changes at a time while product, offer, price and distribution stay fixed. Our observational data can suggest which elements to test first, but it cannot tell you what will work for your product.
The question is which hooks, formats and opening seconds are associated with higher sales per view on TikTok Shop, and how a brand should turn those patterns into a test it can trust. This matters because creative is one of the few levers a brand and its creators control directly, and testing it badly wastes sample, budget and creator goodwill.
Prior work shows why caution is needed. Early popularity predicts later popularity and attention is long-tailed [1], and in short-video popularity prediction, creator features matter most [2]. The returns to advertising are also hard to measure from observational data [3]. A format that looks strong in our data may simply be used by stronger creators, on better products, with more paid support. This paper therefore reports associations and proposes a test, and does not claim any comparison here is an experiment.
2. Data
Our analysis covers US TikTok Shop from 19 September to 18 October 2026. We tracked 51,843 videos across 29 categories, and our AI analyst watched and analysed 9,285 of them in depth. Watched means analysed, not a count of views.
For each analysed video we recorded the hook type, the format, and whether a face, on-screen text or the product appeared in the first second. We also recorded how many of those three elements appeared together. Each group in the evidence carries a video count and a confidence label, and groups with few videos are labelled medium confidence.
These are US TikTok Shop observations. They do not describe all user-generated content everywhere, other markets, or other platforms.
Daily rankings of top-selling TikTok Shop videos 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, hook source, length, pacing, what appears in the first second and how it is presented (people on screen, voice-over, setting, UGC or studio feel), 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 2026-09-19 to 2026-10-18: 51,843 videos tracked, 9,285 of them watched and analysed.
3. Methods
Sales per view compares a group of videos' revenue per view with the market's, so 1 is the market average and 1.2 means the group sold 20% more per view. Estimates for small groups are shrunk toward the average by empirical Bayes [4], and we report 90% intervals computed on the log scale.
Share of sales is the group's share of sales among the watched videos. It reflects how much of the revenue a creative choice accounts for, which depends on how often it is used as well as how well it sells.
We read a range that sits wholly above or below 1 as a clearer signal, and a range that spans 1 as unresolved. All comparisons are descriptive. We did not control for product, creator, price, offer or paid promotion.
4. Results: hooks differ, but only a few clearly beat the market
Price or deal hooks had the highest sales per view at 1.47× the market (90% range 1.37–1.58, 492 videos), though they carried only 5.6% of sales. Bold claim hooks sold at 1.24× (1.19–1.29) across 1,361 videos and 17% of sales. Result-first hooks sold at 1.19× (1.07–1.31) on 266 videos.
The most used hooks sold near the average. Relatable POV (1,558 videos, 18.5% of sales) was at 0.98× (0.95–1.02), and problem call-out (1,123 videos, 13.5%) at 0.98× (0.94–1.02). Straight into the demo was 0.92× (0.87–0.98) and story opener 0.90× (0.81–1.00).
Volume of use and sales per view diverge here. The hooks that account for the most sales are not the ones that sell most per view. This suggests a gap that a test could probe, but it may also reflect which products and creators use each hook.
| Value | Videos (n) | Share of sales | Sales per view | 90% range |
|---|---|---|---|---|
| Price or deal | 492 | 5.6% | 1.47× | 1.37–1.57 |
| Bold claim | 1,361 | 17% | 1.24× | 1.19–1.29 |
| Result first | 266 | 3.7% | 1.19× | 1.07–1.31 |
| Other | 74 | 0.8% | 1.14× | 0.87–1.48 |
| Social proof | 212 | 2.6% | 1.10× | 1.00–1.22 |
| Question | 428 | 4.6% | 1.03× | 0.95–1.12 |
| Unboxing reveal | 287 | 2.9% | 1.01× | 0.92–1.12 |
| Relatable POV | 1,558 | 18.5% | 0.98× | 0.95–1.02 |
| Problem call-out | 1,123 | 13.5% | 0.98× | 0.94–1.02 |
| Straight into the demo | 661 | 6.5% | 0.92× | 0.87–0.98 |
Note. Sales per view is relative to the market (1.00 = average), shrunk toward the average for small groups (empirical Bayes); ranges are 90% intervals on the log scale.
5. Results: formats, with thin data for some
Voice-over B-roll sold at 1.24× the market (1.16–1.33, 561 videos), unboxing at 1.20× (1.10–1.32, 309 videos) and review or testimonial at 1.17× (1.12–1.22, 1,071 videos, 14.1% of sales). Talking head sold at 1.10× (1.05–1.16).
Demo or tutorial was by far the largest format, with 3,163 videos and 36% of sales, but sold at 0.96× (0.94–0.99). Skits sold at 0.89× (0.82–0.96).
Day in the life (1.38×, 16 videos) and live clip (1.19×, 35 videos) look strong, but both are labelled medium confidence and their ranges (0.86–2.21 and 0.82–1.73) include 1. We do not treat them as findings.
| Value | Videos (n) | Share of sales | Sales per view | 90% range |
|---|---|---|---|---|
| Day in the life | 16 | 0.1% | 1.38× | 0.86–2.21 |
| Voice-over B-roll | 561 | 5.8% | 1.24× | 1.16–1.32 |
| Unboxing | 309 | 3.2% | 1.20× | 1.10–1.32 |
| Live clip | 35 | 0.4% | 1.19× | 0.82–1.73 |
| Review or testimonial | 1,071 | 14.1% | 1.17× | 1.12–1.22 |
| Talking head | 987 | 12.1% | 1.10× | 1.05–1.16 |
| Comparison | 260 | 4% | 1.09× | 0.99–1.19 |
| Green screen | 65 | 0.8% | 1.05× | 0.83–1.33 |
| Routine / GRWM | 314 | 3.7% | 1.00× | 0.92–1.09 |
| Demo or tutorial | 3,163 | 36% | 0.96× | 0.94–0.99 |
Note. Sales per view is relative to the market (1.00 = average), shrunk toward the average for small groups (empirical Bayes); ranges are 90% intervals on the log scale.
6. Results: what is on screen in the first second
Videos with a face in the first second sold at 1.04× the market (1.02–1.07, 4,456 videos, 63.5% of sales), against 0.93× (0.91–0.96, 3,122 videos) without. Videos with the product in the first second sold at 1.03× (1.01–1.05, 6,358 videos), against 0.90× (0.86–0.94, 1,220 videos) without.
On-screen text in the first second showed no clear difference: 1.03× without (0.995–1.06) and 0.99× with (0.97–1.01).
The number of elements present was also associated with sales per view. All three of face, text and product gave 1.07× (1.03–1.10, 2,255 videos, 32% of sales). Two gave 0.99× (0.97–1.02) and one gave 0.92× (0.88–0.96). None of the three gave 0.79× (0.63–0.98) but on only 51 videos.
| In the first second | Videos (n) | Share of sales | Sales per view | 90% range |
|---|---|---|---|---|
| Face in the first second: yes | 4,456 | 63.5% | 1.04× | 1.02–1.07 |
| Face in the first second: no | 3,122 | 36.5% | 0.93× | 0.91–0.96 |
| Text on screen in the first second: yes | 5,144 | 69.1% | 0.99× | 0.97–1.01 |
| Text on screen in the first second: no | 2,434 | 30.9% | 1.03× | 0.99–1.06 |
| Product in the first second: yes | 6,358 | 81.6% | 1.03× | 1.01–1.05 |
| Product in the first second: no | 1,220 | 18.4% | 0.90× | 0.86–0.94 |
Note. Each pair splits the same videos; sales per view is relative to the market (1.00 = average), shrunk toward the average for small groups; ranges are 90% intervals.
7. Discussion
For brands and creators, the patterns point to hypotheses worth testing: deal-led and claim-led hooks, voice-over B-roll, unboxing and review formats, and a face and product on screen at the start. They do not show that changing a hook or format will raise sales. Alternative explanations include creator quality and audience size, which matter most in popularity prediction [2], and product and price mix, offer strength and paid promotion. Format and hook choices may also be correlated with category. Observational returns to advertising are hard to read [3], and the same caution applies here.
Recommendation, not a finding: a controlled creative test plan. (1) Pick one product, one offer and one price, and keep them fixed for the whole test. (2) Change one element at a time: hook type first (for example price or deal against bold claim against relatable POV), then format (voice-over B-roll against review against demo or tutorial), then first-second content (face and product on screen against neither). (3) Use the same creator or creators across cells where possible, or assign creators to cells at random. (4) Give each cell the same distribution: same posting window and same paid-promotion setting, or none. (5) Agree before launch the metric (sales per view against a common baseline), the number of videos per cell and the decision rule. (6) Run each cell more than once, and report ranges rather than a single number.
Recommendation, not a finding: when assessing an agency's testing proposal, check that it states what stays constant, how videos are assigned to cells, which metric decides the result, and how it handles cells with few videos. This paper makes no claim about any agency's performance or fees.
8. What this means for brands and founders
- Test price or deal hooks early: they sold at 1.5× the market's sales per view (90% range 1.37–1.58), though the comparison is observational and your offer and product will differ.
- Do not assume the default format is the best one: demo or tutorial carried 36% of watched-video sales but sold at 0.96× the market (0.94–0.99), so include a review or voice-over B-roll cell (1.17× and 1.24×) in your test.
- Put the product on screen in the first second of test videos: those sold at 1.03× against 0.90× without (1,220 videos in the lower group), an association worth checking in your own test.
- Treat day in the life and live clip results as unproven: they rest on 16 and 35 videos and their ranges include 1, so do not base a budget on them.
- Hold product, offer, price and paid promotion fixed across test cells, because creator and distribution differences can produce gaps like those in this data without any creative effect.
9. Conclusion
In our analysis of US TikTok Shop, price or deal hooks (1.47×) and voice-over B-roll formats (1.24×) had the highest well-supported sales per view, and the common demo or tutorial format sat slightly below the market average at 0.96×. A face and the product in the first second were each associated with higher sales per view.
These are associations from one 30-day window, not results of an experiment. The practical use is to turn them into a controlled test with product, offer and distribution held constant, and to let that test, not this comparison, decide what a brand scales.
Methodology
Daily rankings of top-selling TikTok Shop videos 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, hook source, length, pacing, what appears in the first second and how it is presented (people on screen, voice-over, setting, UGC or studio feel), 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 2026-09-19 to 2026-10-18: 51,843 videos tracked, 9,285 of them watched and analysed.
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, views and sales are estimates, not figures reported by TikTok. Observational data from a single 30-day window (19 Sept–18 Oct 2026) in the US. Only 9,285 of 51,843 tracked videos were analysed in depth, and the comparisons rest on the analysed videos. We did not control for creator, product, price, offer, category mix or paid promotion, so creative differences may reflect who used a format rather than the format itself. Groups with few videos (day in the life, live clip, the none-of-three group) are thin and uncertain. Hook and format labels come from an AI analyst and may contain classification error. Nothing here is an experiment or a prediction of results for any brand.
References
- [1]Szabo, G., & Huberman, B. A. (2010). Predicting the popularity of online content. Communications of the ACM, 53(8), 80–88. cacm.acm.org/research/predicting-the-popularity-of-online-content
- [2]Ye, L., Zhang, Y., Wu, Y., et al. (2025). MVP: Winning solution to SMP Challenge 2025 video track. arXiv:2507.00950. arxiv.org/abs/2507.00950
- [3]Lewis, R. A., & Rao, J. M. (2015). The unfavorable economics of measuring the returns to advertising. The Quarterly Journal of Economics, 130(4), 1941–1973. academic.oup.com/qje/article-abstract/130/4/1941/1914592
- [4]Efron, B., & Morris, C. (1975). Data analysis using Stein’s estimator and its generalizations. Journal of the American Statistical Association, 70(350), 311–319.
Cite as
MediaLabs Research, in collaboration with Coherence Limited (2026). TikTok Shop agency creative testing: a practical framework. ContentIQ Working Paper CIQ-WP-2026-55, version 1. https://medialabs-co.com/research/tiktok-shop-agency-creative-testing
Analysis and data: Coherence Research, Coherence ContentIQ. Published by MediaLabs in collaboration with Coherence Limited.
Version history
- Version 118 Oct 2026This version