When dozens of sellers stream the same commodity at the same price, such as the same orchard's oranges or the same white-label face cream, buyers have almost no reason to prefer one seller over another. Competition then turns into a price war. Live-streaming commerce, where a host sells products on real-time video while viewers chat and buy inside the stream, should in principle help here: it can show a product being used. The best experimental evidence says it mostly does not. A host eating the fruit on camera barely moved credibility or purchase intent, because viewers assume the demonstration was staged. A host broadcasting from the orchard itself did move them. That signal works because it is expensive, and most small sellers cannot afford it.
We propose that live interaction can make a cheap demonstration hard to fake. The mechanism is that the viewer, not the seller, chooses the test. A seller who can pass a rehearsed test proves little. A seller who passes whatever an unknown viewer demands in real time ("cut the third orange from the left," "rub the cream on the back of your hand and wipe it with a tissue") must have a product that passes nearly any test. The plan is a live, incentive-compatible experiment that tests this claim. It also asks the question that decides whether the idea can scale: will viewers who did not make the request themselves believe the demonstration, given that most live chat is already assumed to be paid shills?
Core Claim
In a market for identical goods, a quality signal matters only if it lets a buyer tell sellers apart, and a buyer believes it only if a low-quality seller could not copy it. The claim has three parts. First, a viewer-chosen, real-time test restores the credibility that scripted trials lose, and it raises the price premium a buyer will pay to stay with that seller rather than switch to an identical cheaper listing. Second, the effect disappears for bystanders when the request comes from an anonymous chat user. Third, it returns when the platform certifies that the requester was drawn at random from real viewers. If the third part holds, the useful intervention is a platform feature (certified random test requests) rather than a streamer behavior, and small sellers get a cheap substitute for costly field broadcasting.
Where This Sits in Prior Work
The closest work is How to stand out in the "lemon market"?. Yao and Li apply signaling theory, the economics of actions only honest sellers can afford, to live commerce. In a scenario experiment with 215 consumers, field streaming raised perceived quality, seller credibility and purchase intent against a studio broadcast (P = 1.2 × 10⁻⁶). A live eating trial raised perceived quality only marginally (P = 0.047) and did not significantly raise credibility (P = 0.248) or intent (P = 0.147). The authors blame a "scam hangover" from publicized fake eating broadcasts. This paper both motivates and constrains our design. It motivates it because it shows that cheap demonstrations fail. It constrains it because any new signal must beat staging suspicion specifically. Our design differs in three ways. It uses actual synchronous streams where Yao and Li used static screenshots. It gives control of the test to the viewer. It measures behavior with real money, not only Likert ratings.
Two findings from How Live Streaming Changes Shopping Decisions in E-commerce supply the second constraint. Only 17.5% of surveyed users trusted live comments, citing paid posters and bots. The same survey found that 41.7% bought because a live demonstration showed the product was useful. These results pull against each other. Demonstrations persuade, yet anything routed through chat is suspect. That is exactly why a request from another chat user may fail, and why we test platform certification as the fix. The same study's "efficiency paradox", fatigue from repetitive scripted streams across rooms, is the consumer-side face of service homogeneity. A viewer-driven test is by construction unscripted.
On the homogeneity side, Fang Chen and Yufan Liu describe how converging offerings drive switching costs toward zero and push firms into price wars. We make that outcome measurable as the price gap at which a buyer switches away from a seller. Measuring Similarity between Brands using Social Media Content shows that official brand content is nearly indistinguishable across competitors and that real differentiation lives in what customers produce. Our proposal takes the same stance, applied to a live stream: the viewer's own contribution carries the information that seller-authored content cannot. Xiaowei Zhou et al. show that Geographical Indications, legal certification of origin, command lasting premiums. We treat that as a competing, institution-backed route to credibility.
The recommender-system line of work, LARM, HarmonRank, LSEC-GNN and KuaiLive, addresses a different part of homogeneity: which rooms get shown. LARM, for example, reduces traffic concentration on superstar hosts. These systems route attention among sellers. They do not change whether a buyer who arrives believes the seller. We deliberately do not propose another ranking objective, for reasons given below.
Research Program
We will build a small live-commerce testbed: a streaming page with working chat, a buy button, and a request queue. Trained confederate hosts sell two homogeneous products. The first is fresh citrus from a single supplier, which replicates Yao and Li's category. The second is a white-label moisturizer, which extends the study to a cosmetic and addresses their stated category limitation. Every seller sells the identical physical product at an identical posted price, so any difference in outcome comes from the signal alone.
Participants join in cohorts of about twenty per live session, recruited from active live-commerce users. Each receives an endowment and makes consequential decisions: an actual purchase at the posted price, and a willingness-to-pay elicitation using the Becker–DeGroot–Marschak procedure. In that procedure the buyer states a maximum price, a random price is drawn, and the buyer purchases if the draw falls below their stated maximum, so truthful reporting is optimal.
After the stream, a switching stage makes homogeneity concrete. Each participant faces a series of choices between the watched seller at price p and an identical listing from an unseen seller at p − d, with d increasing. The d at which the participant switches is the price premium the stream earned. We will also collect Yao and Li's three rating scales so results are directly comparable. We add a rating of "how easily could this demonstration have been staged?" as the hypothesized mediator. Watch time before leaving the stream is recorded passively.
The conditions are between-subjects and randomized at the session level:
- Studio pitch. No demonstration.
- Scripted trial. The host tests the product in a pre-planned way.
- Field broadcast. The host streams from the orchard or the production line.
- Own-request test. The participant's own submitted request is performed live.
- Unattested other-viewer test. A request is shown with an ordinary chat username and performed live.
- Platform-attested random-viewer test. The same kind of request is shown with a platform badge stating that the requester was drawn at random from verified viewers of this session, then performed live.
- Asynchronous own-request test. The participant's request is collected before the session, and the stream shows a recorded clip of it being performed.
- Matched price cut. A studio pitch plus a discount equal to the median premium measured in the pilot.
In conditions 5 and 6, the requests actually come from randomly chosen cohort members, and requesters are excluded from bystander analyses, so there is no deception about the request's origin. Host scripts, product lots, session times (evening, matching the 6–10 PM peak documented in KuaiLive) and stream length are held fixed. Host identity is crossed with condition so that host charisma cannot masquerade as a signal effect.
The sample target is about 100 bystanders per condition per product, roughly 1,600 bystanders in total, plus the requesters. Yao and Li's N = 215 across three cells left the trial effect underpowered. Around 100 per cell gives about 80% power to detect a standardized difference of 0.4 between two cells at α = 0.05, and we pre-register that as the smallest effect worth acting on.
Baselines
The naive reference is the studio pitch. Every other condition is compared against it on purchase rate, willingness to pay, switching premium and credibility, using the same product lot, price, host pool and session schedule.
The strongest existing method is field broadcasting, the cost-increasing signal from Yao and Li. That paper reports significance at P = 1.2 × 10⁻⁶ against studio broadcasting for perceived quality, credibility and purchase intent. It reports no effect size or means that we can quote. We therefore re-estimate the field effect in our own testbed rather than compare to a published number. Field broadcasting gets the same stream length and host pool as the other conditions. The viewer-chosen test "beats" it in a useful sense if its switching premium falls inside a pre-registered equivalence margin of the field premium, or above it, at a small fraction of the production cost. We log setup cost per session for both conditions. Outright superiority is not required, because the point is cost substitution.
The scripted trial replicates the risk-increasing signal that Yao and Li found ineffective. It is the direct foil for our mechanism. The own-request test must beat it on credibility and switching premium. The matched price cut stands in for the reviewer's "just be cheaper" alternative. The signal counts as escaping the commodity trap only if it raises willingness to pay and the switching premium, not merely conversion at a lower margin.
Two ablations isolate the key ingredient. The asynchronous own-request test keeps the viewer's control of the test but removes real-time execution. It separates "unpredictable to the seller" from "live." The unattested other-viewer test keeps live execution and viewer choice but removes the bystander's reason to believe the requester is genuine. The attested condition then tests whether platform certification restores that belief.
Why not just…?
Why not just cut the price? For identical goods, a lower price will usually raise conversion, so the price cut is a real competitor on sales. It cannot raise the price a buyer will pay to stay with a seller, though, and Fang Chen and Yufan Liu describe where that road ends: collapsing margins. Our matched-discount cell settles the question empirically. If it matches the viewer-chosen test on willingness to pay and switching premium, the signal adds nothing and we will report that.
Why not just stream from the field? Field streaming works, per Yao and Li. Its cost is the reason it works, and cost is what small sellers cannot bear. Michael Huang et al. document the same capital barrier for rural operators in a neighboring setting. This alternative is a genuine threat only if the viewer-chosen test proves weaker than the field broadcast. The field-versus-attested comparison, together with logged costs, measures the trade-off directly instead of assuming it.
Why not just add a certification badge or origin label? Legal origin labels do carry premiums (Xiaowei Zhou et al.). They exist only for certified regions, however, and they certify origin rather than the condition of the particular lot on camera. Certification is closest in spirit to our attested condition, since both borrow the platform's credibility. If reviewers insist, a static-badge cell is the natural ninth condition. We expect it to sit between the scripted and attested tests. If it matches the attested test, the "live" part of our claim is unnecessary.
Why not just re-rank feeds for diversity among rooms selling the same product? A redundancy-penalizing ranker built on semantic embeddings of the kind LARM produces could reduce the sameness viewers see. Standard diversity re-ranking would probably do this about as well as any new objective, which is why we did not propose one. Showing viewers more varied rooms, however, does not make any room more believable, and believability is the bottleneck identified by the 17.5% chat-trust figure and the failure of the scripted trial.
Decisive Checks
The idea stands or falls on four comparisons, each pre-registered:
- Mechanism. The own-request test must exceed the scripted trial on credibility and on switching premium, with the staging-ease rating mediating the difference. If it does not, a viewer's control over the test is not a credible signal, and the proposal fails at its core.
- Scaling. Among bystanders, the attested random-viewer test must approach the own-request test while the unattested test falls back toward the scripted trial. This pattern would show that chat distrust, not the demonstration itself, is what blocks scaling, and that platform certification fixes it. If the attested and unattested conditions are equal, certification adds nothing.
- Synchrony. The asynchronous own-request test tells us whether "live" matters at all. If it performs as well as the live version, the contribution narrows to seller-unpredictable testing. That is still useful, but it is no longer a claim about synchronous commerce, and we will say so.
- Generality. Every contrast is estimated separately for citrus and moisturizer. This checks whether the effect holds beyond the fresh-food category that Yao and Li studied, or whether it depends on how easily a product's defects can be seen on camera.
References
- How to stand out in the “lemon market”? application of signaling theory in live streaming commerce
- How Live Streaming Changes Shopping Decisions in E-commerce: A Study of Live Streaming Commerce
- HEYTEA Consumer Consumption Experience, Brand Identity and Research on the Relationship of Brand Loyalty
- Performance Research of Chinese Patent Medicine Enterprises to Innovative Medicine Research and Development Strategy Transformation
- [Paper] Measuring Similarity between Brands using Social Media Content
- Analysis on the Influence of Geographical Indication Protection on International Competitiveness of Shaanxi Apple
- LLM-Alignment Live-Streaming Recommendation
- HarmonRank: Ranking-aligned Multi-objective Ensemble for Live-streaming E-commerce Recommendation
- Leveraging Tripartite Interaction Information from Live Stream E-Commerce for Improving Product Recommendation
- KuaiLive: A Real-time Interactive Dataset for Live Streaming Recommendation
- Exploration of the Application of Virtual Digital Human Technology in Rural Cultural and TourismDigital Transformation Pathways from the Perspective of Scene Theory