The evidence problem specific to AI creators

The founding claim of this category comes from HypeAuditor’s 2021 study of virtual creators: “Virtual Influencers have engagement rates almost three times higher than the engagement rates of real influencers”, drawn from a sample described as “the top 129 virtual influencers on Instagram at the moment” (HypeAuditor). Treat this as a directional pattern from a vendor study, not as a benchmark: the size and construction of the human control group is not disclosed on the page, the methodology is not published, and the post carries a later update stamp on the same URL — so anyone citing it is citing a moving document. Per-tier percentages circulating under this study’s name are downstream retellings; they are not on the source and they are not in this article.

Editorial opinion. Even taken at face value, the pattern measures the wrong thing for a performance buyer. A novel format collects curiosity engagement: comments asking whether she is real, shares captioned “look at this”. ER is a leading indicator of attention, not of purchase intent, and the novelty component decays as the format stops being novel. That is why the cases below are read for what they measure, not for what they claim.

Case 1. Aitana López: the only AI-creator economics documented end to end

Euronews reported the numbers on Spain’s best-known AI model directly: “The virtual model can earn up to €10,000 a month, according to her creators, but the average is usually around €3,000”, and “She earns just over €1,000 per advert”. On audience, “In a year and a half, she has gained more than 343,000 followers on Instagram”. She was built by “Rubén Cruz, her designer and founder of the agency The Clueless”, and by the time of publication “she has recently become the face of Big, a sports supplement company” (Euronews, 27 December 2024).

A note on the follower number. 343,000+ is a snapshot dated December 2024, not a property of the character. Any follower count inside a case study — this one included — is a measurement with an expiry date, and a competitor can date-stamp a stale figure in ten seconds. Cite the date, not just the number.

What an advertiser should take from it. The economics, not the ER. At roughly €1,000 per branded post, a creator with a premium fashion audience prices like a nano or micro creator while delivering the production consistency of a studio: the visual language is fixed once and does not drift between shoots, seasons or haircuts. That is the actual structural difference from a human creator of the same audience size, and unlike an engagement multiple, it is visible in the deliverable before you pay for it.

Our practice. In our own portfolio the consequence of that consistency is testing hygiene rather than a cheaper post: several integrations a month share one visual language, so a creative A/B test compares the variables you set rather than the noise of two different shoots. We are not publishing our internal conversion or engagement figures here — those belong in a commercial conversation where the counterparty can interrogate the methodology.

Case 2. WooHoo Dubai: a brand-reported number, labelled as one

(brand-reported, not independently verified). Gökhan Girmez, chief brand officer at Gastronaut Hospitality — owner of the Dubai restaurant WooHoo — told Marketing Week that working with AI influencers has driven engagement rates 2–3× higher than the brand’s standard branded content, alongside stronger video completion rates, particularly on Reels and short-form storytelling. He also framed the commercial mechanics: AI influencers behave contractually more like licensed digital IP than like individuals, which he presents as more cost-efficient over time given scalability, content reuse and the absence of recurring appearance fees (Marketing Week, 2026).

Why this case is stronger than the campaign write-ups it usually appears in. It is an on-the-record statement by a named executive at the brand that ran the activity, in a trade publication that puts its name on it — not an anonymous agency deck. That does not make it a benchmark. No methodology, sample or measurement window is published, and the comparison is against the brand’s own prior content, which is exactly the right comparison for a brand and exactly the wrong one for an industry average. Read it as: for this brand, against its own baseline, the format outperformed. That sentence is defensible. “AI creators get 2–3× the engagement” is not.

Case 3. BMW iX2 × Lil Miquela: the campaign whose ER does not exist

The “Make It Real” campaign for the all-electric BMW iX2 paired the brand with Lil Miquela — a virtual (CGI) character, not an AI system — described by the agency as having a “social audience of more than 8.2 million people” across platforms. The campaign’s award record and credits are on the agency’s own case page: 1x CLIO Award, 1x Epica Award and 3x Lovies, directed by Stefanie Soho with production by BWGTBLD (Monks).

(a negative one, and the most useful sentence in this article). That same page contains no engagement rate, no impression count and no interaction count. We checked it. Two different numeric versions of this campaign circulate in the industry — a flattering one attributed to the agency and a damning one attributed to critics — and neither appears on the agency’s own case page or on any primary campaign page. They are not reproduced here in either direction, because a number with no primary is not evidence of a good campaign or a bad one.

Lesson for the brand. When a mega-scale virtual creator fronts a hero campaign, treat agency-reported performance figures as marketing collateral until you see the export. What that campaign demonstrably bought is on the record and is not an engagement rate: an awards shelf, a named director, a production company, industry press and a permanent portfolio case. If the objective is cultural placement, that is the deliverable. If the objective is conversion, none of it is.

Case 4. KFC’s Virtual Colonel: 151 million impressions of a joke

The agency’s own case page describes the Virtual Colonel as “The world’s first branded virtual influencer”, built to “parody the lifestyle of Instagram influencers” — posting from the gym and from private jets, “with each post complete with brand sponsorships and captions spoofing some real-life influencers”, in partnership with Dr Pepper, Old Spice, TurboTax and Casper, across an account with 1.3 million Instagram followers. The reported outcome: “Virtual Colonel generated over 151 million earned impressions—not bad for someone who doesn’t exist” (Wieden+Kennedy). The work is listed in the Webby Awards winners index for Best Use of Social Media, 2020 (Webby Awards).

Editorial opinion — read the intent before you read the impressions. This campaign was designed as satire of influencer culture, by a brand with the licence to be absurd, running for two weeks. 151 million earned impressions is a real, agency-reported attention number and it is also a number that says nothing about sentiment, brand lift or intent — and copying its format without its irony is how a brand ends up in a comment section it did not budget for. Percentages of negative sentiment attached to this campaign circulate widely and trace back to no primary source we could verify; they are not in this article. The durable lesson does not need them: volume of engagement and quality of engagement are separate measurements, and a case study that reports only the first has told you only half of what happened.

What this means for the mega-versus-niche decision

Our arithmetic, not eMarketer’s. Take the two cross-platform figures above at face value for a moment — 2.53% for nano, 0.92% for mega. To generate the same number of interactions, a mega audience has to be roughly 2.75× the size of the combined nano audience you would otherwise buy. That is arithmetic on two published figures, not a measured campaign result, and it ignores price entirely. Its only use is to show why the tier argument exists at all: at the top of the audience curve you are paying for reach and cultural weight, and the interactions arrive as a by-product.

Editorial opinion. Three things a niche creator — human or AI — structurally does better: measurable performance work where the outcome is a signup, install or sale; always-on publishing at a cadence a comparable human would price out of; and categories where a personal scandal is the expensive risk. Three things a mega still does better: mass awareness in a compressed window; categories with a high trust bar, such as medical and pharma, where credentials rather than reach carry the message; and moments where the campaign needs to be part of a cultural conversation rather than adjacent to one.

We are not putting a ratio on any of those six statements. The published comparisons of AI-creator versus human-creator earnings and engagement that would let us do so failed source verification — the pages cited for them do not carry the figures — so this section is a judgment about mechanisms, marked as one.

A five-step framework for the choice

Editorial opinion / our practice. This is how we frame the decision with advertisers. It contains no benchmark numbers on purpose: every step is a question about your campaign, not a threshold from someone’s report.

1. Objective. Top-funnel awareness in a fixed window puts macro and mega on the table. Mid- and bottom-funnel performance work almost always favours a pack of niche creators, because you need repetition and variation, not a single moment.
2. Budget shape, not budget size. One hero placement is a bet with a single outcome. A pack is a portfolio with a distribution of outcomes, and it produces reusable assets. Decide which risk profile you are buying before you decide how much to spend.
3. Category. In verticals where engagement is structurally low, stop using ER as the selection criterion and select on trust signals and audience fit instead. In regulated and high-trust categories, an expert human is still the format. In luxury and fashion, brand fit dominates every metric.
4. Platform. Short-form platforms reward frequency and format-native production; that favours creators who can publish often and consistently. Longer-form video rewards depth and watch time, which is a different casting decision entirely.
5. Geography and language. A single market usually favours a local creator with genuine cultural fluency. Multi-market campaigns are where a synthetic creator’s ability to run consistently across languages becomes the actual argument — not its engagement rate.

Checklist: how to read an AI-creator case study

• [ ] Open the cited link and search the page for the exact number. If it is not there, the case does not exist as evidence.
• [ ] Is the comparison against the brand’s own baseline or against an industry average? Both are legitimate; they are not interchangeable.
• [ ] Is sentiment reported, or only volume?
• [ ] Was there paid amplification inside what is presented as organic performance?
• [ ] Is the claimed return actual revenue, or earned media value?
• [ ] Is the source a primary page — brand, agency case, awards entry, regulator, filing — or an aggregator repeating an aggregator?

If three or more answers fail, you are holding PR collateral. That is not an accusation of dishonesty; it is a statement about what the document can be used for. Collateral cannot set a KPI.

The cases that are not in this article, and why

Editorial note. Earlier versions of this piece carried two further “wins” of the kind that circulates in every deck on this subject: a luxury-brand collaboration with a headline return-on-investment and production-saving figure, and a large nano-creator programme with a headline return per euro spent. Both were dropped in full — from the body, the headline, the meta description, the social cards and the structured data.

The reason is the same in each case and it is not a matter of interpretation: we opened the pages those figures were attributed to, and the figures are not on them. Not different numbers — no such case on the page at all. Publishing a performance claim about a named public company on that basis is a risk we are not taking, and the checklist above would have failed our own article. Where a figure could not be sourced, this version says so instead of substituting a friendlier one.

The takeaway

A niche creator, synthetic or human, is a cost-per-outcome instrument. A mega creator is an attention instrument. The public case record supports that split, and — read strictly — supports very little else about AI creators specifically: one fully documented set of economics (Aitana López), one brand-reported engagement uplift against a brand’s own baseline (WooHoo), one award-winning hero campaign whose performance numbers do not exist in public (BMW × Lil Miquela), and one enormous earned-impressions number attached to a deliberate parody (KFC).

That is a thinner evidence base than the category’s marketing implies, and knowing exactly how thin it is happens to be the most valuable thing an advertiser can carry into a negotiation. The person across the table is quoting from the same set of cases.

Next reads: how to write a creator brief that works · what an AI-creator integration costs, line by line · the seven-step framework for choosing an influencer · what AI creators actually are.

→ Browse the AI-creator catalogue · Discuss a campaign: advertise@cifratar.ai

Frequently asked questions

What counts as a healthy engagement rate in 2026?

There is no single number, and any answer that gives you one without naming a dataset is guessing. Within one published dataset the direction is consistent: eMarketer’s June 2025 round-up reports nano creators at 6.23% on Instagram, and across all platforms 2.53% for nano against 0.92% for mega — figures it attributes to vendor studies, not to its own research. Use them as direction, not as a contractual target, and never compare a number from one vendor’s methodology with a number from another’s.

Do AI creators get better engagement than human ones?

Not provably, at the level of precision the industry claims. The pattern comes from a 2021 vendor study of virtual Instagram accounts whose methodology and human control group are not published; the per-tier percentages that circulate under its name are downstream retellings that do not appear on the source page. What is documented is narrower: one brand, WooHoo in Dubai, has stated on the record that AI influencers outperform its own standard branded content by 2–3×. That is a brand’s comparison against its own baseline, not an industry benchmark.

Can I trust the engagement rate an agency reports for an AI campaign?

Only with the export behind it. The clearest example is the BMW iX2 campaign with Lil Miquela: two contradictory numeric versions circulate in the industry, and neither appears on the agency’s own case page, which carries the awards, the director and the production company but no performance metrics at all. Ask for a raw data export or an independent audit, and treat any figure whose primary page you cannot open as absent.

What did the BMW × Lil Miquela campaign actually deliver?

Publicly and verifiably: a campaign for the all-electric iX2 with a virtual (CGI) character whose social audience the agency puts at more than 8.2 million people, one CLIO Award, one Epica Award, three Lovies, direction by Stefanie Soho and production by BWGTBLD. What it did not deliver publicly is any engagement, impression or interaction figure — none appears on the agency’s case page.

How much does an integration with an AI creator at Aitana López’s level cost?

Euronews reported just over €1,000 per advert, average monthly earnings around €3,000 and up to €10,000 in a strong month, on an Instagram audience of more than 343,000 followers as of December 2024. Economically that prices like a nano or micro creator with a premium fashion audience. Note that the follower figure is a dated snapshot, not a fixed property. The full cost structure is in our integration cost breakdown.

Why does this article have no engagement-rate table by niche?

Because we could not build one that would survive a source check. Every widely quoted niche breakdown we tested either contradicts the figures attributed to it or comes from a vendor content page with no published methodology. An unsourced table is the most portable kind of misinformation — it gets screenshotted and separated from its caveat — so this version omits it rather than publish one we would have to defend.

Where does an AI creator lose to a human today?

In categories where credentials carry the message rather than reach — medical and pharmaceutical work in particular — and in cultural moments built on a personal story arc. We are not attaching a ratio to that gap: the published comparisons of AI versus human creator earnings and engagement did not survive source verification, so this is a judgment about mechanisms, not a measurement.