What the Evidence Actually Says About Performance
Virtual influencers are often declared successful because people notice them. The more useful question is what kind of response that notice represents: novelty, engagement, trust, a favorable attitude, or an intention to act. This chapter focuses on the comparative evidence behind those outcomes so that readers can judge a virtual influencer without treating one visible signal as proof of overall marketing performance.
What a comparative study can—and cannot—tell us
A single post can look persuasive. It may attract a large audience, collect comments, or become a talking point for a brand. But an observed reaction does not tell us what would have happened with another endorser, a different message, or no such stimulus at all. Comparative studies are useful because they place a virtual-influencer endorsement beside a control condition and ask whether the measured response differs.
The strongest evidence supplied for this question is a meta-analysis: a statistical synthesis that combines results across studies rather than relying on one campaign or one famous persona. It covers 210 independent experimental studies from 110 papers, 643 effect sizes, and a total sample of 91,058 consumers. Its inclusion rule matters. The retained studies compared a virtual-influencer endorsement with a control and provided enough information to calculate an effect. Studies that only compared two kinds of virtual influencer, or only measured a virtual influencer’s characteristics without a comparative consumer outcome, were excluded.
That design gives the findings a useful form: relative effects. If the virtual-influencer condition scores lower on credibility than its comparison condition, the result does not mean that every virtual influencer is inherently untrustworthy. It means that, across the relevant studies, the virtual condition tended to receive a lower credibility response under the conditions tested. The comparison is the denominator of the conclusion. Remove it, and a precise research result becomes an exaggerated slogan.
The outcomes were also kept separate. Novelty refers to whether the endorser, message, or advertisement feels unusual or unexpected. Engagement refers to interaction or involvement, such as likes, shares, comments, or time spent. Credibility concerns perceived trustworthiness and expertise. Attitudes capture evaluations of the advertisement, brand, or product. Behavioral intention is readiness to perform a behavior such as buying. Attractiveness concerns the persona’s perceived physical appeal or likability.
These measures sit at different points in a customer relationship. Novelty is often an early perceptual response. Engagement indicates that someone interacted, but not necessarily that they agreed or will return. Attitude indicates evaluation, but not necessarily action. Behavioral intention is closer to a decision, though it is still not the same as observed purchase. A campaign dashboard that combines all of these into “influence” hides the path from attention to value.
The second source, a PRISMA-based systematic review of virtual influencers, serves a different purpose. It reviews 73 articles selected from an initial pool of 188 and organizes recurring themes around evolution, comparative effectiveness, trust, authenticity, engagement, and ethics. It is valuable for showing how the field describes the problem and where questions recur. It is not a replacement for the meta-analysis’ pooled comparative estimates. Its tables of accounts and social-media activity are descriptive illustrations, not controlled estimates of what a representative virtual influencer will achieve.
The distinction between the two kinds of evidence is important. A meta-analysis can estimate an average pattern under a defined set of comparisons. A systematic review can map concepts, methods, and gaps across a wider literature. Neither one is a live return-on-investment test. Neither can settle how a particular audience on a particular platform will respond to a new persona. The right use of the evidence is therefore conditional: establish a baseline, identify the variables that may change it, and avoid extending the baseline beyond its design.
The meta-analysis also does not describe one uniform experiment repeated hundreds of times. The studies vary in their products, messages, participants, platforms, and comparison conditions. Multiple effect sizes can come from the same study, and the analysis accounts for that nested structure rather than pretending that every observation is independent. This improves the estimate, but it does not make the underlying situations identical. A pooled result is best read as a pattern across a research landscape, not as a guarantee for the next campaign brief.
Applicability depends on what a new campaign shares with that landscape. A short exposure to a stylized character is not the same test as a long-running account with repeated audience interaction. A product whose value can be explained in a sentence is not the same as one that depends on embodied experience. Nor is a social-media interaction necessarily evidence of a downstream purchase. Those differences do not invalidate the pooled baseline; they specify the questions a manager must answer before transferring it. A useful follow-up test would preserve a meaningful comparison while varying the persona’s anthropomorphism, the product context, and the outcome window, rather than measuring only immediate reactions.
The average pattern: novelty up, persuasion mixed
The pooled results are clearest when read as a profile rather than a ranking. Virtual influencers produced significantly more perceived novelty than their comparison conditions. The reliability-corrected, sample-size-weighted correlation for novelty was 0.217. The value is not a percentage of people persuaded; it is a pooled association expressed in the analysis’ correlation metric. It tells us that artificial distinctiveness is a recurring feature of the virtual-influencer condition.
Engagement tells a different story. The average correlation was −0.052, and the difference from the controls was not statistically significant. That is consistent with the conclusion that virtual influencers can be about as effective as the comparison conditions at generating engagement overall. It does not mean that they engage every audience equally, or that a virtual account is interchangeable with a human one in every platform environment. The underlying effects are heterogeneous: the Q statistics reported for the outcomes indicate that the observed variation is greater than what random sampling alone would explain.
In practical terms, “no significant difference” is not the same as “all designs perform the same.” Imagine a collection of experiments in which some stylized characters perform well, some realistic personas perform poorly, and other studies find little difference. An average near zero can result from genuine equivalence, from opposing effects in different contexts, or from imprecise estimates. The aggregate result tells us not to claim a general engagement advantage. It does not erase the need to ask which design and audience produced which result.
The disadvantages are more consistent on persuasion-oriented and evaluative outcomes. Relative to the coded controls, virtual influencers were on average lower on behavioral intention, with a pooled correlation of −0.130; brand attitude, −0.120; advertisement attitude, −0.192; credibility, −0.234; and attractiveness, −0.160. The signs matter more than any one coefficient taken out of context. The virtual condition was not merely less familiar or less attention-getting. Across the evidence base, it also encountered a comparative penalty when consumers were asked to evaluate the source, the advertisement, the brand, or a possible action.
Those figures should not be translated into claims such as “virtual influencers reduce purchase by 13 percent.” A correlation is not a percentage-point change, and the studies do not all use the same products, audiences, messages, or controls. The values are average comparative estimates. They indicate direction and relative strength within the meta-analytic model, not a forecast for an individual campaign.
Credibility is especially revealing. A virtual persona can be attractive enough to stop a scroll and coherent enough to sustain an account, yet still be judged a weaker source of information. Human influencers can draw on a personal history that makes a recommendation appear grounded in experience. A virtual persona’s history is authored, rendered, and managed. Even a carefully built fictional identity may therefore prompt a different question: is this someone who knows the product, or a commercial character designed to make the product interesting?
That question helps explain why engagement and credibility should not be treated as interchangeable. People may comment because a persona is strange, beautiful, amusing, or technologically impressive. Those reactions are real forms of attention. They become evidence of persuasion only when the research measures a relevant evaluative or behavioral consequence and compares it with an appropriate alternative. The average findings do not deny the value of attention; they show that attention does not automatically carry credibility with it.
The same caution applies to attractiveness. A digital persona can be rendered with visual consistency and aesthetic control, but perceived appeal is not determined by polish alone. Human likeness, cultural expectations, familiarity, and the context in which the image appears can change how the design is read. The pooled disadvantage on attractiveness is therefore a comparative signal, not a claim that synthetic characters cannot be appealing. It is one reason to avoid treating production quality as a substitute for audience fit.
The broader review adds an important field-level observation: human influencers are often described as more relatable and emotionally authentic, while virtual influencers benefit from novelty and controlled messaging. These descriptions align with the meta-analysis, but their evidentiary status differs. A thematic synthesis can reveal a recurring explanation across studies; it cannot turn every recurrence into a common effect size. The reviews converge on a tension, not on a universal rule that human endorsers win every outcome.
The pattern also guards against a common attribution error. Suppose a virtual-influencer campaign receives high engagement. The result may reflect the persona’s novelty, the creative quality of the post, the brand’s existing popularity, the platform’s distribution, or the audience’s prior interest. Without a suitable comparison, we cannot know how much of the response came from the virtual identity. Even with a comparison, an experiment may answer a short-term perception question rather than a long-term retention or revenue question.
The evidence therefore supports three restrained conclusions. First, virtual influencers have a repeatable advantage in making endorsement feel distinctive. Second, they do not show a reliable average engagement advantage over the comparison conditions, so attention should not be sold as automatic outperformance. Third, they face average disadvantages in credibility, attitudes, attractiveness, and behavioral intention, but those disadvantages are not fixed properties of every persona. The heterogeneity is not an inconvenience to be ignored; it is the reason to look for moderators and mechanisms.
A manager deciding from these results should ask which outcome the campaign is trying to move and how close that outcome is to a consequential decision. For awareness, novelty may be a legitimate objective. For a functional claim or a considered purchase, credibility may be the bottleneck. For a launch, engagement may be useful but still insufficient evidence of demand. The same account can therefore be a sensible attention asset and a poor substitute for a human source of lived expertise, depending on the job.
This is also why the average disadvantages do not justify a blanket rejection. An average is a starting point that includes different designs and contexts. If high and low anthropomorphism, autonomy cues, product newness, and message framing change the result, then the overall estimate is partly an average of unlike situations. The next question is not whether the pooled numbers should be believed. It is what psychological route and contextual match can explain their variation.
The apparent contradiction is now easier to state. Virtual influencers can be more novel without being more credible. They can generate interaction without reliably changing attitudes or intentions. The data do not require us to choose between “they work” and “they do not work”; they require us to identify what audiences are evaluating. Once the outcome profile is separated, the mechanism becomes visible: anthropomorphism may determine whether the persona is processed mainly as a credible source or as a novel object.