TL;DR for operators

Generative AI makes it technically possible for a small team, or even a disciplined solo operator, to build a virtual influencer: a consistent face, voice, backstory, content calendar, visual style, and interaction pattern. That is the easy part. The harder part is making the persona commercially useful rather than merely photogenic.

The relevant lesson from virtual-influencer research is not “AI avatars get attention.” It is more specific: audiences respond to cues that make synthetic characters feel socially embedded, human-like, and trustworthy. In one study of virtual influencers, posts showing the influencer with companions generated higher engagement, with the authors estimating a 27% lift in field data, and experiments linking the effect to perceived humanness and trust.1

For businesses, the practical implication is simple and mildly inconvenient: do not treat an AI persona as a cheaper stock photo with captions. Treat it as a governed publishing system. The asset has to include a character bible, disclosure rules, visual continuity, a review workflow, brand-fit testing, and measurement against business outcomes. Otherwise you have not built Lil Miquela. You have built a mannequin with a posting schedule. Congratulations, sort of.

Mascots used to be expensive. Now they need governance.

A brand mascot used to be a controlled object. It appeared in campaigns, packaging, maybe a few carefully scripted videos. It did not wake up every morning and improvise opinions about skincare, crypto, coffee, labour rights, or your competitor’s latest product launch.

Virtual influencers changed that. They took the mascot out of the campaign folder and gave it a social feed. Lil Miquela became the obvious reference point because she looked like a person, behaved like a creator, collaborated like an influencer, and remained, inconveniently for philosophers and conveniently for managers, owned infrastructure.

Generative AI changes the next layer of the story. It does not merely reduce the cost of making posts. It reduces the cost of maintaining identity across formats: image, caption, short video, comments, newsletters, livestream scripts, customer replies, and multilingual variants. The cost centre moves from production labour to editorial control.

That is the useful shift. A virtual persona used to require a studio-like setup: creative direction, 3D modelling, copywriting, production, scheduling, community management, and analytics. Many of those functions still matter, but AI compresses the workflow. A smaller team can now generate drafts, visual concepts, story arcs, and variants at a pace that would previously require several people pretending the content calendar was a personality.

The misconception is that the advantage is “near-zero content cost.” That is only half true, and the half that gets people into trouble. AI lowers the marginal cost of producing more content. It does not lower the cost of knowing what should be said, why the audience should care, whether the persona should say it, and what happens when synthetic charm meets legal liability.

The paper’s real lesson is trust, not cheaper CGI

The most useful research anchor here is not a paper about image generators. It is a paper about how virtual influencers become persuasive.

Rizzo, Berger, Villarroel Ordenes, and Pozharliev study what drives virtual influencer impact by examining “companion presence”: whether a virtual influencer appears alone or with other people in an image.1 The design matters because virtual influencers have a built-in credibility deficit. Audiences know, or eventually discover, that the character is controlled. The character may be stylish, consistent, and tireless. It is also not a person. Tiny problem.

The study combines large-scale analysis of social posts with controlled experiments. Its central finding is that showing a virtual influencer with companions can increase engagement and product choice. The mechanism is not visual clutter or simple popularity theatre. The authors argue that companion presence makes the virtual influencer seem more human; that perceived humanness increases trust; and trust improves response.

This is a more useful claim than “make the avatar pretty.” Pretty is abundant. Human context is harder. A virtual persona standing alone in a luxury apartment can look polished and dead. A persona interacting with recurring friends, collaborators, customers, or other characters starts to acquire social evidence. The audience is not just evaluating the face. It is evaluating the world around the face.

For operators, that changes the brief. The question is not only:

What should this persona look like?

It is also:

Who appears with this persona, what relationships do they imply, and what trust signals does the content repeatedly create?

That is where the business value begins to look less like avatar design and more like media-system design.

What the evidence supports, and what it does not

The research does not prove that every brand should launch a synthetic influencer. It does not prove that AI-generated creators outperform human creators. It does not prove that audiences enjoy being quietly manipulated by a corporate imaginary friend, though some seem remarkably tolerant of it.

It does support a narrower and more operationally useful point: virtual influencers are not judged only as images. They are judged as social actors. Their effectiveness depends on cues that make them interpretable as credible participants in a social environment.

Claim Evidence Business meaning Boundary
Virtual influencers can drive engagement and choice Companion presence increased engagement in field data and improved product choice in experiments.1 Visual storytelling should include social context, not just solo glamour shots. The effect is about specific content cues, not a guarantee that synthetic personas work everywhere.
Trust is a key mechanism The study links companion presence to perceived humanness and trust.1 Persona design should prioritise credibility cues, relational continuity, and behavioural consistency. Trust may vary by category, audience, culture, and disclosure context.
Influencer marketing depends on content features, not just follower count Broader influencer research finds that originality, sponsor salience, fit, and content choices shape effectiveness.2 AI personas need a content strategy, not merely a posting engine. Human-influencer findings do not transfer perfectly to synthetic characters.
GenAI lowers barriers to content production Research on GenAI use by creators shows that non-experts can remix, repackage, and scale content across modalities.3 Smaller teams can now operate persona-led media channels. Lower barriers also increase spam, misinformation, sameness, and reputational risk.
Disclosure affects trust and compliance Studies of social-media endorsement disclosures find persistent under-disclosure and user confusion.4 Synthetic personas need clear paid-relationship and synthetic-identity policies. Disclosure rules differ by jurisdiction and platform.

The point is not to worship the avatar. The point is to understand the production and trust mechanics behind it.

The AI persona stack is an operating system, not a profile picture

A credible AI persona has layers. Most failed attempts only build the top one.

Layer What it contains Why it matters Common failure
Identity system Name, backstory, worldview, interests, taboos, language habits Keeps the persona coherent across formats The persona sounds like a brand deck wearing sunglasses
Visual continuity Face rules, wardrobe logic, settings, recurring objects, camera style Makes the character recognisable and less obviously stitched together Every post looks like a different intern prompted it
Social world Friends, customers, collaborators, rivals, places, rituals Creates human context and trust cues The character exists alone in aesthetic captivity
Content pillars Education, entertainment, product use, commentary, community response Prevents random posting disguised as experimentation “We post daily” becomes the entire strategy
Governance Disclosure, claims review, escalation, prohibited topics, brand safety Reduces legal and reputational surprises The avatar says something clever, false, and expensive
Measurement Engagement quality, conversion path, sentiment, retention, assisted revenue Separates audience curiosity from business value Vanity metrics are mistaken for market traction

Generative AI helps across every layer. It can generate backstory variants, caption drafts, multilingual posts, image prompts, comment replies, A/B hooks, and campaign calendars. But the system still needs editorial constraints. Without them, the persona becomes smooth, energetic, and strategically homeless.

The best use case is not “replace the influencer.” It is “create a controlled media asset that can be tested, iterated, and governed.” That distinction matters. Human creators bring biography, lived experience, audience history, and reputational texture. Synthetic personas bring consistency, controllability, and scalability. These are different assets. Treating one as a cheap imitation of the other is how brands get both the ethics and the economics wrong.

The business case is control plus iteration, not magic engagement

A synthetic persona can be valuable when the brand needs repeatable communication in a defined niche. Think financial education, product onboarding, language-localised tutorials, gaming communities, beauty experimentation, B2B explainers, or fictional brand worlds.

The strongest business case usually has three conditions.

First, the brand needs frequent content, not one heroic campaign. AI personas become more useful when the workflow requires ongoing presence: daily posts, weekly explainers, recurring customer education, seasonal launches, and community replies. If the campaign only needs one polished video, hire a studio and spare everyone the “metaverse ambassador” paragraph.

Second, the category benefits from controlled voice. A virtual persona can stay on-message, avoid personal scandals, and follow compliance rules if the governance is real. That is attractive in regulated or reputation-sensitive sectors, although those sectors also raise the review burden. The machine does not remove compliance work. It merely creates more material for compliance to inspect. Progress, with paperwork.

Third, the audience must accept stylised mediation. Some communities enjoy fictional guides, mascots, avatars, and synthetic characters. Others want human expertise, lived experience, or professional accountability. A synthetic skincare model can demonstrate a routine. A synthetic doctor giving treatment advice is a lawsuit rehearsing in public.

The ROI logic should therefore be framed around operating leverage:

  • lower cost per content variant;
  • faster localisation;
  • faster testing of angles, formats, and hooks;
  • stronger consistency across channels;
  • reduced dependency on external creator schedules;
  • reusable persona assets across campaigns.

That is a serious case. It is just not the lazy version where the avatar becomes famous because the brand bought a Midjourney subscription and a dream.

Companion presence is really a design principle

The companion-presence finding deserves special attention because it points to a general principle: synthetic personas need social proof built into the narrative environment.

A human influencer brings a visible social world by default. They have friends, colleagues, messy apartments, travel delays, old posts, offhand references, and the occasional regrettable lunch photo. A virtual influencer has none of that unless someone designs it.

That absence is costly. A synthetic character without relationships can feel frictionless in the wrong way. Too clean. Too available. Too obviously optimised. The result is not trust but suspicion, or worse, indifference.

Companion presence helps because it answers an implicit audience question: does anyone else in this world treat the persona as real? The companion does not need to be another synthetic influencer. It could be a founder, a customer, an artist, a recurring fictional side character, a product expert, or even a community member featured with permission. The point is that trust often emerges through relationships, not declarations.

For operators, this suggests a more disciplined content brief:

Instead of asking Ask this
“How do we make the avatar look premium?” “What social context makes the persona believable?”
“How many posts can we generate?” “Which recurring situations teach the audience how to interpret the character?”
“Can we automate replies?” “Which replies require human review because they affect trust, claims, or customer expectations?”
“Can this replace influencers?” “Where is control more valuable than lived authenticity?”

The cheap version of AI persona marketing generates content. The useful version designs repeated trust cues.

Disclosure is not a footnote for lawyers to find later

Synthetic personas create two disclosure problems at once. The first is the standard influencer-marketing issue: if there is a material connection between the endorser and the brand, it should be disclosed. The second is synthetic identity: the audience may also need to know that the persona is not a human being.

The FTC’s guidance on endorsements focuses on clear disclosure of material connections, and broader endorsement rules have increasingly had to account for fabricated or virtual endorsers.5 Separately, empirical work on affiliate-marketing disclosures shows how often disclosure is missing or poorly understood, even in ordinary human creator contexts.4 Add synthetic media and the ambiguity does not improve itself out of courtesy.

The operational answer is boring, which is usually a good sign. Put disclosure rules into the workflow:

  • label synthetic identity clearly where it matters;
  • disclose paid relationships in the content, not in a hidden policy page;
  • avoid fake personal experience claims;
  • maintain a claims-review process for regulated categories;
  • archive prompts, approvals, and final posts for auditability;
  • decide which interactions require human escalation.

This is not just defensive lawyering. Disclosure can be part of positioning. A brand can say: this is our virtual product educator, our fictional market guide, our AI-assisted design muse. Audiences are not automatically offended by fiction. They are offended by being treated as if they are too inattentive to notice.

Where this works, and where it probably disappoints

AI personas work best when the audience wants continuity, style, utility, or entertainment more than autobiography. They are well suited to product education, fictional brand worlds, gaming, youth culture, fashion experimentation, software onboarding, and repeatable explainer content.

They are weaker when credibility depends on lived experience. Parenting advice, medical judgement, trauma stories, political persuasion, financial recommendations, and luxury authenticity all create higher trust burdens. A synthetic persona can assist communication in those spaces, but replacing accountable human expertise with an attractive simulation is not strategy. It is a compliance department’s origin story.

There is also a sameness problem. Generative AI makes it easy to produce competent content in large quantities. That means competent content becomes less scarce. If every brand has a polished AI ambassador with flawless skin and a suspiciously balanced content calendar, audiences will not reward the fifteenth one for existing. Novelty decays. Distinctiveness remains expensive.

The durable advantage is therefore not the avatar. It is the operating model around the avatar: sharper positioning, better data feedback, more coherent storytelling, faster testing, stronger governance, and a clearer role in the customer journey.

A practical build sequence for a brand-owned virtual persona

A sensible build does not start with a face. It starts with a job.

  1. Define the commercial role. Decide whether the persona is for awareness, education, lead capture, community engagement, product demonstration, or retention. A persona built for memes will not automatically sell enterprise software. Tragic, but true.

  2. Write the character bible. Include backstory, vocabulary, values, humour range, visual style, content boundaries, forbidden claims, and escalation rules. The character bible is the difference between a persona and a recurring accident.

  3. Design the social world. Create recurring contexts: colleagues, customers, locations, rituals, collaborators, events, and community interactions. The companion-presence research makes this more than decoration; it is part of trust design.1

  4. Build the production workflow. Use AI for drafts, variants, image concepts, localisation, and scheduling support. Keep human review for claims, sensitive replies, final tone, and strategic direction.

  5. Test content types separately. Measure educational posts, product demos, lifestyle posts, companion posts, founder interactions, and direct-response posts differently. Aggregated engagement is where weak strategy goes to hide.

  6. Disclose cleanly. Make synthetic identity and commercial relationships understandable. Do not bury the truth under a confetti cannon of hashtags.

  7. Connect metrics to revenue. Track saves, replies, qualified clicks, sign-ups, assisted conversions, sentiment, and repeat engagement. Likes are useful only when they indicate movement through a business pathway.

Conclusion: build the system, not the doll

Generative AI makes virtual personas cheaper to produce. That part is real. The more interesting change is that persona-led media can now be run as a lightweight operating system: identity, visuals, content, interaction, governance, and measurement.

But the research points to a less glamorous lesson. Synthetic influence depends on trust cues. A virtual influencer works not because it is artificial, but because the audience can read it as socially meaningful, consistent, and credible enough for the context. Companion presence is one measurable example of that principle. Disclosure is another. Content originality, brand fit, and clear commercial purpose still matter, because apparently audiences remain annoyingly human.

So yes, a small team can now build its own Lil Miquela-like asset. But the goal should not be to manufacture a fake celebrity and hope the internet applauds the rendering quality. The goal is to build a controlled, useful, transparent media character that earns attention repeatedly and converts that attention into business value.

The face is the easy part. The operating discipline is where the money is hiding.

Cognaptus: Automate the Present, Incubate the Future.

  1. Giana L. C. Rizzo, Jonah Berger, Francisco Villarroel Ordenes, and Rumen Pozharliev, “What Drives Virtual Influencer’s Impact?,” arXiv:2301.09874, 2023. ↩︎ ↩︎ ↩︎ ↩︎ ↩︎

  2. Fine F. Leung, Flora F. Gu, Yiwei Li, Jonathan Z. Zhang, and Robert W. Palmatier, “Influencer Marketing Effectiveness,” Journal of Marketing, 86(6), 2022; Fine F. Leung, Flora F. Gu, and Robert W. Palmatier, “Online Influencer Marketing,” Journal of the Academy of Marketing Science, 50, 2022. ↩︎

  3. Amelia Hassoun, Ariel Abonizio, Katy Osborn, Cameron Wu, and Beth Goldberg, “The Influencer Next Door: How Misinformation Creators Use GenAI,” arXiv:2405.13554, 2024. ↩︎

  4. Arunesh Mathur, Arvind Narayanan, and Marshini Chetty, “Endorsements on Social Media: An Empirical Study of Affiliate Marketing Disclosures on YouTube and Pinterest,” arXiv:1809.00620, 2018. ↩︎ ↩︎

  5. Federal Trade Commission, “Disclosures 101 for Social Media Influencers,” 2019; FTC, “Guides Concerning the Use of Endorsements and Testimonials in Advertising,” revised 2023. ↩︎