ONCE Social Short-Form Video Marketing Research · Social Short-Form Video Marketing Research 07/10
Research article / evidence comparison and creative analysis. Research cut-off: October 5, 2026.
A short video can generate no click yet leave a model name in memory, prompting a later search. Another can collect comments and saves while leaving viewers unable to name the brand. Treating both outcomes as “content performance” hides one of the hardest parts of short-form video marketing: discovery may happen in the feed, while the decision migrates to search, a brand site, a marketplace, a dealer, a private conversation, or an offline visit.
Our central argument is that a short video should not only compete for an in-platform action. It should leave “searchable residue”: a sufficiently clear relationship among brand, category, model, use problem, principal claim, and evidence that can travel with the viewer beyond the feed. Interest without a retrievable name easily becomes an unattributed impression. A name without verifiable evidence may lose credibility as soon as search begins.
Searchable residue is ONCE’s original synthesis of the evidence reviewed here, not a universally validated theory. It divides cross-platform decision design into four tasks: discovery, naming, evidence, and handoff. The point is not to assume that everyone will search. It is to decide whether the likely next step for this product is immediate purchase or leaving with a question.
Six adjacent outcomes that should not be collapsed
Dashboards often arrange views, searches, saves, clicks, and conversions into one funnel. Real decisions need not occur in that order.
| Observable outcome | What it primarily indicates | What it cannot prove on its own |
|---|---|---|
| Exposure and viewing | The content had an opportunity to be seen and watched | The viewer understood the product or formed purchase intent |
| Comments and saves | The viewer chose a public response or deferred action | The viewer will return, search, or buy later |
| Branded or category search | A discovery was reformulated as an active question | Which video caused incremental search or which brand captured it |
| Click | The viewer attempted the platform-designated next step | Comparison, verification, or purchase was completed |
| Platform-attributed conversion | An action was recorded within observable attribution rules | The result was incremental or all off-platform outcomes were captured |
| Off-platform lead or purchase | The decision finished with a brand, retailer, dealer, or offline channel | Without matching or an experiment, which exposure caused it |
This distinction extends our first study on recommendation, exposure, and platform objectives: the platform conditions who gets a chance to see the content, while the brand must decide what the encounter leaves behind. It also extends our fourth study on brand memory and correct attribution: remembering a sequence is not the same as retrieving the right brand when a need arises.
No click does not necessarily mean no effect. A click does not automatically mean incremental sales either. The useful first step is not to place every metric in one larger dashboard. It is to specify the estimand: are we trying to estimate added search, qualified visits, leads, or purchases, or merely describe one observed path?
What an 84-million-user experiment establishes—and what it does not
Yitian (Sky) Liang, Xinlei Chen, Shengnan Han, Jinglong Zhang, and Yubo Chen published a ByteDance field experiment in Marketing Science in 2025. Working with an automobile brand, the researchers randomly assigned more than 84 million users to treatment and control groups; only treatment users could be exposed to the short-form video ads. The experiment ran for seven weeks, followed by roughly a month and a half of conversion observation. “Conversion” meant providing personal information such as a name and phone number, not purchasing a car.
The brand suspended its other digital advertising during the experiment. The researchers matched ByteDance users to brand conversion records by phone number, allowing them to compare leads occurring within and beyond ByteDance. Exposed users were more than eight times as likely to convert outside ByteDance as within it, and incorporating off-platform outcomes materially changed the estimated effect and cost calculations. The formal article also argues that incomplete information sharing between platform and brand can distort evaluation and targeting decisions. The journal record and abstract are available from INFORMS; the public preprint supplies additional method detail but is explicitly labelled as not peer reviewed.
Its causal identification is substantially stronger than an ordinary attribution report, but its external-validity boundary is equally important. This was one automobile advertiser, one paid campaign, a particular ByteDance environment, and a lead outcome. It does not establish the same spillover for organic content, consumer goods, software, industrial equipment, or local services. The final article also discloses that two authors had previously been an employee or intern at Ocean Engine, while the others declared no relevant interests. Turning the eight-times figure into a universal media-planning multiplier would convert a valuable experiment into a new myth.
The defensible conclusions are narrower. First, platform-only data can miss genuine off-platform outcomes. Second, whether and how much it misses must be estimated in the brand’s own context through matching or experimentation—not borrowed from someone else’s multiplier.
When people move from scroll to search, the cue hierarchy can change
Edith Shalev, Coby Morvinski, Meyrav Shoham, and Ellie J. Kyung’s “Search or Scroll,” published online in 2025 and appearing in the 2026 volume of the Journal of Consumer Research, analyses more than 27 billion following decisions by roughly eight million users across Goodreads, Yelp, Twitter/X, and Instagram. It combines four large platform datasets with pilot studies and an online experiment.
The authors distinguish goal-directed “search” from experiential “scroll” orientations. In the following decisions they examine, search-oriented settings placed a higher premium on communicator credibility, while scroll-oriented settings placed a higher premium on likability. Inducing participants to search rather than scroll within X also changed the relative weight given to those cues. The open article is available from Oxford Academic.
The transfer must be disciplined. The outcome is whom to follow, not what to buy. Goodreads, Yelp, X, and Instagram do not automatically represent contemporary Douyin, Xiaohongshu, WeChat Channels, TikTok, Reels, or Shorts. The study nevertheless offers a testable mechanism: as people shift from experiential browsing to goal-directed inquiry, their evaluation criteria may change. Warmth, pacing, and novelty that work in the feed may not carry a comparison. Model numbers, specifications, applicable conditions, limitations, independent evidence, and service information may become more important.
A short-form video should therefore ask not only how to stop the scroll, but whether a departing viewer can use its cues to find the correct object. This does not require turning every video into a specification sheet. It requires the decisive relationship to be nameable, repeatable, and verifiable.
The searchable-residue model: discovery, naming, evidence, handoff
1. Discovery: connect the problem to the product
Discovery is not merely an exaggerated opening. It is the moment viewers recognize that the content relates to a task they have. For a familiar, low-risk category, a visual contrast, use moment, or price may suffice. For a higher-risk or unfamiliar category, discovery may also need to define the problem, identify who experiences it, and show where the product enters.
If the opening is a spectacle unrelated to the product and the brand is appended at the end, viewers may remember the content without forming a query. A practical discovery check is to remove the end card and ask whether viewers can name the category problem, rather than merely retell the plot.
2. Naming: give interest a phrase it can carry
Naming does not mean repeating the brand. It creates a minimum searchable unit: “brand plus model,” “brand plus functional category,” “problem plus technical term,” or “product plus use context.” The unit must be specific enough to retrieve and simple enough not to depend on perfect recall of advertising copy.
A model name flashed briefly, a product replaced by a nickname, or a function expressed only through generic adjectives adds friction to migration. The opposite extreme—stuffing every keyword into subtitles—competes with comprehension. A stronger approach synchronises name and action: when the product performs its defining action, a legible name, sound, subtitle, and object converge on one memory anchor.
3. Evidence: anticipate the first search-stage objection
Evidence does not mean shrinking a detail page into a short video. It gives the core claim a minimum visible basis. Once viewers leave the feed, especially for higher-consideration products, they may ask: under what conditions does this work, what changes relative to an alternative, and does the brand acknowledge limits?
A short video can establish first-layer evidence through a real action, comparison, interface, scale relationship, material detail, or explicit limitation, then hand full specifications and testing conditions to the next page. Evidence must match the claim: atmosphere does not demonstrate performance, and a use demonstration does not establish long-term reliability.
4. Handoff: make the next interface continue the same question
Handoff is not synonymous with a “buy now” button. It may be a brand-and-model search, product-page visit, case study, dealer inquiry, save-for-comparison, or message to another decision-maker. The production team should know what the next interface contains before shooting, so that names, claims, images, and landing-page evidence remain continuous.
ONCE’s cross-border e-commerce video solution separates content tasks for discovery, comparison, and purchase. The page describes a production scope; it is not independent research evidence for this article. Its practical value is to show why one production can create different information densities for different decision stages instead of forcing a single cut to do everything.
Low-risk products and high-risk decisions should not share a default path
For familiar, low-cost, low-risk goods, viewers may buy within the platform using very little information. Reducing handoff steps and preserving a price and purchase entry may matter more than prompting external search. Yet even a low-priced product needs a retrievable name when category differences are confusing, the brand is unfamiliar, or the claim is easy for competitors to appropriate.
High-value, infrequent, or risky decisions usually involve more evidence and more people. A viewer considering a vehicle, enterprise software, industrial equipment, health-related product, or durable good may discover, search, read reviews, ask a colleague, contact sales, and finally convert offline or on another platform. A longer path does not confine short video to “top-of-funnel awareness.” It changes its assignment from completing persuasion alone to handing the right question to the next interface.
ONCE’s public Oraimo folding e-bike case offers material for a creative analysis. We read the page text and visible page information; we are not treating it as a validated growth experiment, and no public conversion data are available. The page links “fold into the car boot—arrive—unfold and ride—reach the pier,” turning portability from an adjective into a repeatable use sequence.
Through the searchable-residue lens, that sequence could leave three query levels: the category “folding electric bicycle,” the context “fits in a car boot,” and the brand and model. It does not prove folding speed, range, or compatibility with every vehicle; those claims need specifications and testing material at the search stage. The case demonstrates how a function can become a nameable action, not a sales outcome.
Search can also create attribution errors: a counterexample
Sarah Moshary’s 2025 Management Science article uses a field experiment at an unnamed e-commerce platform that blocked all sponsored-search advertising for 3% of visitors. In that platform and period, sponsored search diverted business from organic listings and reduced the total volume of platform-intermediated transactions, although advertising revenue still made the arrangement profitable for the platform. The formal abstract and study record are available from INFORMS.
This is not a short-video study and not a universal verdict on branded search. It matters as a counterexample: a search click may sit close to conversion while still redistributing demand that would have occurred organically. Observing that viewers later searched and clicked an ad does not, from path correlation alone, reveal how much incremental demand the video created or how much incremental value the search ad contributed.
Cross-platform design must therefore solve two different problems: what cues the content sends away with the viewer, and how measurement distinguishes new action from channel reallocation. The first is a creative problem; the second is causal. Combining them tends to award all credit to the channel that is easiest to record.
A minimum falsifiable test of decision migration
The following is a proposed experiment. It has not been run.
Create two versions for the same product. Version A keeps brand, duration, production quality, target audience, budget, and landing destination constant but states the core function abstractly. Version B adds one searchable residue under the same conditions: when the decisive action occurs, it presents the brand, model or category name and one visible item of evidence. Do not simultaneously change the opening, music, edit rate, and offer, or the result cannot be attributed to the cue.
Pre-register one primary outcome. A low-risk good might use incremental in-platform purchases or qualified clicks. A higher-risk product might use treatment-versus-control changes in branded search, direct visits, qualified leads, or dealer bookings. Secondary outcomes can include unaided brand and model recall, claim comprehension, comparison-page visits, and negative reactions. Search volume should, where possible, be separated by random audience, geography, or time. If person-level matching is unavailable, state the boundary rather than force an aggregate trend onto one video.
The experiment can fail in informative ways. Version B may raise branded search while lowering completion; improve model recall while making an inaccurate claim more portable; or shift existing category search toward a brand without increasing purchase. Failure does not simply mean the content failed. It identifies which transition changed: discovery, naming, evidence, or handoff.
Conclusion: a short video need not complete the decision, but it must hand over a usable question
The commercial effect of short-form video does not always occur on the platform where it is watched. For an impulse good, the shortest path may be the best one. For a product that requires comparison and verification, forcing an immediate click or purchase can misrepresent how the decision actually migrates.
Searchable residue reframes the job from chasing one visible action to leaving a portable relationship among brand, category, model, claim, and evidence. It does not guarantee search, and search does not guarantee sales. It provides falsifiable questions: after scrolling away, can the viewer name it? Can search retrieve the correct object? Can the next page verify the claim? Does the handoff continue the same question?
If you are planning a product-content system that must serve feed discovery, off-platform search, and purchase comparison, ONCE can begin with the product, market, principal risk, and existing destination pages. We can then define what each version must leave searchable before deciding the script, action design, asset tree, and validation metrics.
Sources and reading scope
- Yitian (Sky) Liang, Xinlei (Jack) Chen, Shengnan Han, Jinglong Zhang, and Yubo Chen, “Is the Money Spent on Short-Form Video Social Platforms Worth It? The Role of Advertising Spillover in a Large-Scale Randomized Field Experiment on ByteDance,”Marketing Science 44(5), 2025, 1125–1144, DOI: 10.1287/mksc.2023.0575. We read the formal record and abstract plus the public preprint’s methods, results, and limitations; the preprint is not peer reviewed, and bibliographic claims follow the final journal version.
- Edith Shalev, Coby Morvinski, Meyrav Shoham, and Ellie J. Kyung, “Search or Scroll: How Credibility versus Likability Premiums Shape Consumers’ Following Decisions,”Journal of Consumer Research 53(2), 2026, 281–304, DOI: 10.1093/jcr/ucaf038. We read the open full article and do not extend its following outcome into a purchase claim.
- Sarah Moshary, “Does Sponsored Search Advertising Augment Organic Search? Evidence from an E-Commerce Platform,”Management Science 71(11), 2025, 9687–9709, DOI: 10.1287/mnsc.2024.04792. We read the formal abstract and experiment scope and do not generalise one platform’s result to all search advertising.
- ONCE’s cross-border e-commerce video solution, Oraimo folding e-bike case, and previously published series articles were read on October 5, 2026. The case is used for creative analysis only; no undisclosed conversion, budget, or experiment result is claimed.
