What the engagement evidence actually supports

Two things are documented well enough to plan against. Neither is the multiple you were promised.

Engagement falls as audience size rises. eMarketer’s June 2025 round-up carries it in full sentences: “Nano-influencers maintain the highest engagement rate across influencer categories on Instagram at 6.23%” and, separately, “Nano-influencers have the highest overall engagement rate (2.53%), and engagement decreases as follower count increases, with mega-influencers averaging .92%, per a HypeAuditor report cited by Shopify” (eMarketer, 24 June 2025, read 18.08.2026).

Editorial note — read those two sentences as two different measurements. The 6.23% is Instagram-only and carries no in-line attribution on the page. The 2.53% and .92% are cross-platform and arrive through a three-link chain: a vendor study, relayed by a retailer’s blog, relayed by trade press. Subtracting one from the other is not a finding, it is a category error — and it is exactly how the “nano beats mega by 7×” line got manufactured. Use the direction. Do not build a gap out of numbers from two different methodologies, and note that every figure here measures human creators only: none of it says anything about synthetic ones.

(brand-reported, against the brand’s own baseline). The clearest buyer-side statement on record concerns the Dubai restaurant WooHoo. Marketing Week reports, in the passage relaying its chief brand officer: “Beyond the successful Aitana López collaboration, working with AI influencers has driven 2-3x higher engagement rates compared to standard branded content for WooHoo, as well as stronger video completion rates, particularly on reels and short-form storytelling” (Marketing Week, 20 February 2026, read 18.08.2026; the executive is Gökhan Girmez of WooHoo-owner Gastronaut Hospitality).

Editorial opinion — the comparison class is the whole story. That is one brand measuring a format against its own previous content. For that brand it is the right comparison and a genuinely useful one. As an industry average it is worthless: no sample, no window, no denominator, and a baseline nobody outside the company can see. “For this brand, against its own baseline, the format outperformed” is defensible. “AI creators get 2–3× the engagement” is a different sentence — and it is the one that ends up in decks.

The same article shows why campaign numbers rarely settle anything. The WooHoo × Aitana López New Year’s Eve collaboration on Instagram “saw more than 105,000 total views, a total reach of 46,000, and 1,300 plus engagements” — figures the piece reports inside the passage on the AI-influencer agency Pixel.ai, whose co-founder Lewis Davey it quotes throughout, rather than sourcing them to the brand.

Our arithmetic, not the publisher’s. 1,300 engagements is 2.8% of that reach and 1.2% of those views. Same campaign, same source, two results on opposite sides of most published benchmarks depending on which denominator you pick. An engagement figure without its denominator is decoration — and when a vendor quotes you one, “over what?” is the entire audit.

Cost: two price structures, and only one of them has a rate card

For human creators, price is per post, tier-driven and published. The two most-cited grids are live and internally consistent — but they do not share a vocabulary, so the table below aligns them by follower band and keeps each guide’s own label inside the cell.

Table 2. Instagram per-post benchmarks for human creators

Sources: Meltwater (bylined Chris Hanson, dated 18 June 2026) and Shopify, both read 18.08.2026.

Editorial note. Look at the two “macro” rows. Shopify’s macro is 501K–1M; Meltwater’s is 100K–500K. A quote presented as “the standard macro rate” means two different things depending on which guide the seller read. Make every quote name its follower band before it names its tier — that single question removes most of the ambiguity in an influencer negotiation.

On the AI side there is no equivalent grid. One persona has published economics end to end, and that is the entire public record.

Euronews reported the numbers for Aitana López, the Spanish AI model run by The Clueless: “The virtual model can earn up to €10,000 a month, according to her creators, but the average is usually around €3,000”; “She earns just over €1,000 per advert”; and “In a year and a half, she has gained more than 343,000 followers on Instagram” (Euronews, published 22 November 2023, updated 27 December 2024, read 18.08.2026).

Put the two blocks side by side and the intuition behind “AI is cheaper” gets a shape rather than a coefficient: a persona above 343K followers earning just over €1,000 per advert sits well below the published human band for that follower count. But this is one disclosed case against a rate card — different market, different year, creator-side earnings rather than a brand-side invoice, and a follower count that was a December 2024 snapshot. It is a reason to demand a cost breakdown, not a multiplier to put in a model.

The tooling underneath does have list prices, and they are smaller than most quotes imply. Google’s Gemini API price list puts Veo 3.1 video generation at $0.40 per second on Standard (720p and 1080p) and $0.10 per second on Fast at 720p (ai.google.dev, read 18.08.2026 — note that Veo 3 and Veo 2 are marked deprecated on that same page, so a vendor quoting Veo 3 is quoting a shut-down model). ElevenLabs lists Creator at $22/month and Pro at $99/month, with multilingual synthesis inside those plans rather than as a separate localisation SKU (elevenlabs.io/pricing, read 18.08.2026; the $11 figure visible on that page is a “First month 50% off” promotion, not the rate).

Our arithmetic, not a vendor’s. At those list rates, the raw generation for one 30-second clip is roughly $12 on Standard and $3 on Fast, before retries. If the quote for that clip is four figures, the gap is not “AI costs” — it is direction, iteration, brand-look conformance, rights and margin. That is the negotiable part, and a single-number quote is designed to hide it.

The contract modifiers behave identically for both options, because they are commercial rather than technical. impact.com’s guidance: “Most influencers charge an additional 20 to 50 percent of their base rate for usage rights” and “Most creators charge between 20 and 100 percent of their base rate” for exclusivity, depending on duration and brand size (impact.com, read 18.08.2026). No published AI-specific corridor exists for either. The argument that a studio generating variants at near-zero marginal cost should charge less for rights is real — as a negotiating argument, not as a benchmark.

Editorial opinion. The structural difference is amortization, not discount. A human integration is a transaction: one fee, one airing, repeat next month. An owned character is a fixed setup spread across everything it will ever publish, plus a running cost. One or two deliverables and the setup never amortizes — the AI route will not save you money. A series across markets and the amortization does the work, with localisation genuinely cheap because multilingual voice sits inside the same subscription. We are not printing our own per-channel setup or operating figures here: those are first-party numbers, and they belong in a conversation where the counterparty can interrogate the method.

Speed and control: the part that is not a metric

Editorial opinion. This is where the two options differ most, and none of it fits in a benchmark table — which is probably why the industry keeps arguing about engagement instead.

A human creator brings an audience relationship built without you and continuing without you, a calendar, an agent, a body that gets ill, opinions that exist off the clock, and a voice the audience trusts precisely because it is inconsistent in human ways. Time-to-publish is a function of other people’s schedules; consistency of look is a function of who shot what, when. A synthetic character brings a fixed visual language that does not drift between shoots or seasons, availability that is a scheduling question rather than a negotiation, and every word of its personality authored by someone on a payroll.

The commercial consequence is stated precisely by the same buyer: AI influencers “function more like licensed digital IPs rather than individuals”, and “In many cases, this makes them more cost-efficient long-term, especially when considering scalability, content reuse and the absence of recurring appearance fees. That said, high-quality AI talent still requires meaningful investment in creative development and technology” (Girmez, Marketing Week, 20 February 2026).

Editorial opinion. Read the second half of that quote as carefully as the first. “Licensed digital IP rather than an individual” is the real product difference: it changes who signs, what the contract covers, how reuse is priced and what remains yours when the campaign ends. It is not a synonym for “cheaper”, and the same executive says so in the next breath.

Risk: it does not disappear, it changes owner

Editorial opinion. The standard pitch is that a synthetic character cannot have a scandal. The record says it can — and when it does, the brand has no “they said it themselves” distance to stand behind.

In May 2019, Calvin Klein ran a #MYCALVINS video in which Bella Hadid kissed Lil Miquela. After accusations of queerbaiting the brand issued a public statement: “We understand and acknowledge how featuring someone who identifies as heterosexual in a same-sex kiss could be perceived a queerbaiting… As a company with a longstanding tradition of advocating for LGBTQ+ rights, it was certainly not our intention to misrepresent the LGTBQ+ community. We sincerely regret any offense we caused” (Washington Blade, 20 May 2019, read 18.08.2026). Nothing in that episode came from an individual’s behaviour. Every element was authored: the concept, the pairing, the script.

Editorial opinion. So the honest framing is a transfer, not an elimination:

• With a human you carry reactive risk — off-duty conduct, personal scandal, a schedule slip, a parallel account you do not control. It is managed through relationships, contracts and morality clauses.
• With a synthetic character you carry authored risk — every storyline is your editorial decision, and every representation choice about gender, race, sexuality and lived experience is made by a brand team rather than by a person speaking for themselves. It is managed through process: a review step before publication with a named senior owner for the character’s narrative.

A brand that switches to synthetic and skips that review step has not removed risk. It has removed the alibi.

Disclosure: the legal load is asymmetric, and it has dates

This is the axis most comparisons skip, and the only one where the difference between the two options is written down in law.

The FTC’s Endorsement Guides put the duty on the advertiser regardless of who — or what — the endorser is. The endorser “could be or appear to be an individual, group, or institution”; “Advertisers are subject to liability for misleading or unsubstantiated statements made through endorsements or for failing to disclose unexpected material connections between themselves and their endorsers”, and advertisers should “(1) Provide guidance to their endorsers…; (2) Monitor their endorsers’ compliance; and (3) Take action sufficient to remedy non-compliance and prevent future non-compliance” (16 CFR Part 255.0(b) and 255.1(d), current text via eCFR, read 18.08.2026). The FTC’s business-guidance FAQ restates the same programme operationally: “Instruct members of the network on their responsibilities for clearly and conspicuously disclosing their connections to you… Periodically search for what members of your network are saying; and Take appropriate action if you find questionable practices” (FTC, read 18.08.2026).

(a negative one). We searched the current text of Part 255 for provisions specific to virtual, synthetic or computer-generated endorsers: zero occurrences of “virtual” and zero of “computer-generated”. The obligations attach to the advertiser, the endorsement and the material connection — not to the endorser’s ontology. Anyone citing a dedicated FTC rule for AI personas should be asked for the section number.

The EU rule is the one that changes the calendar. Article 50(4) of the AI Act: “Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated” (EU AI Act, Article 50, read 18.08.2026). On timing, the correction most articles get backwards: the Article 50 obligations start on 2 August 2026 — the penalty regime (Chapter XII) has applied since 2 August 2025 (Article 113).

Platform labelling is already live and works differently on each surface:

• YouTube — labels “may appear in the expanded description for AI content that is non-photorealistic or animated”, and “For AI content that is photorealistic, a label in the video player may also appear” (YouTube Help, read 18.08.2026).
• Meta — detects AI images through C2PA/IPTC metadata, but relies on creator self-disclosure for AI-generated video and audio and can penalise non-disclosure (Meta).
• TikTok — described itself as “the first video sharing platform to put Content Credentials into practice” (TikTok Newsroom, 9 May 2024) and states: “To date, we’ve labeled over 3 billion videos as AIGC using a combination of Content Credentials, creator labeling tools and our invisible watermarking technology” (TikTok Newsroom, 10 July 2026).

Editorial opinion. Compliance overhead is therefore a real budget line on the synthetic side and close to zero on the human side — but it is not a penalty box. Platforms are neutral toward disclosed AI content and act against concealment; the expensive failure mode is a brand running a synthetic character as if it were a person and being corrected by a platform, a regulator or a journalist. Disclosure designed in from the first post costs a design decision. Disclosure retrofitted after a correction costs the campaign.

Trust: what the research supports, and what it does not

The uncanny-valley effect for virtual influencers is documented experimentally. Gutuleac, Baima, Rizzo and Bresciani, “Will virtual influencers overcome the uncanny valley? The moderating role of social cues” (Psychology & Marketing, 28 February 2024, DOI 10.1002/mar.21989, metadata confirmed via Crossref 18.08.2026), report that “highly anthropomorphized VIs may elicit a greater sense of uncanniness among consumers” and that “in the presence of social cues, the influence of anthropomorphism on uncanniness is attenuated”.

Editorial opinion. Read practically, that is a design instruction rather than a warning off the format: photorealism is not the safe default it looks like. A stylised character — the virtual (CGI) persona Noonoouri is the reference point here — asks less of the viewer than a near-human one that is slightly wrong, and social cues (responsiveness, visible personality, interaction) soften the effect.

(adjacent evidence, different population). The largest preregistered evidence we could locate on how people judge disclosed AI use is not about creators at all: Reif, Larrick and Soll, “Evidence of a social evaluation penalty for using AI” (PNAS, 8 May 2025, DOI 10.1073/pnas.2426766122, metadata confirmed via Crossref 18.08.2026), find across “four preregistered experiments (N = 4,439)” that “people who use AI at work anticipate and receive negative evaluations regarding their competence and motivation”. That studies employees and job candidates, not brand audiences, so it does not transfer directly. We cite it because the honest state of the evidence is “the nearest rigorous finding points at a cost to disclosure”, not “disclosure is free”.

Editorial note. What does not exist, as far as we could establish: any study comparing trust or purchase intent between a virtual and a human creator at campaign scale with a stated method. Looi & Kahlor (Journal of Interactive Advertising, 2024, DOI 10.1080/15252019.2024.2313721) is the closest — a mixed-method Instagram comparison reporting greater engagement for the human influencers — and it is not campaign-scale. The familiar claim that AI creators cannot sell in finance, health or parenting is a plausible mechanism argument — and this article marks it as an argument, because we could not find the measurement behind it.

Run the comparison yourself: a protocol instead of a benchmark

Editorial opinion / our practice. Since the public comparison does not exist, the useful move is to generate a private one. This is a small, cheap test — one that a brand can run in a quarter and that produces a number nobody can argue with, because it is measured on your audience, in your category, against your own baseline.

Six rules that make the result mean something:

1. Fix the denominator before you start. Decide now whether engagement rate is over reach, over views or over followers, and write it down. The WooHoo figures above land at 2.8% or 1.2% purely on that choice.
2. Compare against your own baseline, not an industry average. The only comparison with a controlled variable is your synthetic content against your own standard branded content, on the same account or on accounts of comparable size.
3. Match the creative brief, not the creator. Same product claim, same offer, same call to action, same format and length. If the AI version is a polished 30-second spot and the human version is a phone selfie, you have measured production value.
4. Match the window and the posting slot. Same days of week, same time band, same number of posts, minimum two weeks per arm to survive one algorithmic bad day.
5. Disclose in both arms. Labelling is mandatory on the synthetic arm anyway (see above), so building the comparison on an undisclosed arm gives you a number you can never act on — and a compliance problem.
6. Log cost the same way on both sides. All-in per published asset: fee or setup amortized over planned volume, plus production, plus revisions, plus rights and exclusivity modifiers, plus the internal hours spent reviewing.

Table 3. What to log in each arm

Editorial opinion. Two arms, one quarter, one fixed denominator. That will beat every sourced number in this article for your specific decision — and it is the only comparison you will be able to defend when someone asks where it came from.

The decision grid

Editorial opinion / our practice. No thresholds and no multipliers below, on purpose. Every row is a question about your campaign; the reasoning is the deliverable.

Table 4. Which option the argument favours

Editorial opinion. The framing that survives contact with the evidence is not “replace” and not “AI wins the top of the funnel”. It is a portfolio question: a synthetic channel is a media property you own, a human creator is a relationship you rent. Brands that can run their own unit economics keep both and stop asking which one has the higher engagement rate — because nobody has published a comparison worth the argument.

What is not in this article, and why

Editorial note. Earlier versions of this piece carried a six-parameter comparison table with an engagement multiple, two campaign performance figures for a named automaker, a per-violation fine table, a return-on-investment range for owned AI channels, and a payback horizon for a character build. All of it is gone — from the body, the tables, the FAQ, the meta description, the social cards and the structured data.

The reason is uniform: we opened the pages those figures were attributed to and the figures are not there. Not different numbers — no such measurement on the page. Several of the cited domains have since been blacklisted in our own source registry as vendor pages presented as research. Where a removed figure was our own — channel costs, payback, portfolio results, production capacity — it stays out until the owner of those numbers signs the exact wording, because a first-party figure is the most checkable and most expensive kind of error a portfolio company of a listed group can publish.

The takeaway

A human influencer and an AI creator are not two grades of the same product. One is an audience relationship you rent for a post; the other is an asset you build, amortize and are wholly accountable for. The public record supports that split and remarkably little else: one vendor study with an undisclosed control group, one persona’s disclosed earnings, one brand’s comparison against its own baseline, one hero campaign with no published performance numbers, one experimental finding about uncanniness, one adjacent finding about the cost of disclosure, and a dated set of legal obligations that only one side carries.

Knowing exactly how thin that evidence base is happens to be the most valuable thing you can carry into a negotiation — because the person across the table is quoting from the same short list.

Next reads: what an AI-creator integration costs, line by line · the ER cases that survive a source check · how platform algorithms treat AI content · how we build a character.

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

Frequently asked questions

Do AI creators really get three times the engagement of human influencers?

Not provably. The claim traces to a single vendor study of virtual Instagram accounts whose method and human control group are not published, and the neat per-tier percentages circulating under its name do not appear on the source page at all. The only buyer-side comparison on record is narrower: Marketing Week reports that AI influencers drove 2–3× higher engagement than WooHoo’s own standard branded content, with stronger completion rates on Reels. That is one brand against its own baseline, not an industry benchmark — and it is the strongest evidence in the category, which tells you how thin the category is.

Is an AI creator cheaper than a human at comparable reach?

No published source quantifies the gap, so treat any clean multiplier as a sales claim. What is published: human Instagram rate cards (from “$20-$100” per post at 500–10K followers up to “$10,000 to more than $50,000 per post” above 1M, depending on the guide) and one disclosed AI persona — Aitana López, reported by Euronews at just over €1,000 per advert with more than 343,000 Instagram followers as of December 2024. Those two facts point the same way but are not like-for-like: different market, different year, creator earnings rather than a brand invoice.

Where does a human creator clearly still win?

In categories where credentials rather than reach carry the message — regulated finance, medical and health advice — in campaigns built on a personal story arc, and in single-deliverable campaigns where nothing amortizes. We are deliberately not attaching a ratio to that gap: no study we could locate compares trust or purchase intent between virtual and human creators at campaign scale with a stated method, so this is a judgment about mechanisms rather than a measurement.

What does an AI creator actually change for a brand?

Three things that are not metrics: consistency (a visual language that does not drift between shoots), speed and availability (scheduling rather than negotiation), and contract shape — as WooHoo’s chief brand officer puts it, AI influencers “function more like licensed digital IPs rather than individuals”, which changes who signs, what reuse costs and what remains yours when the campaign ends. The same executive adds that high-quality AI talent still requires meaningful investment in creative development and technology.

Does an AI creator remove reputational risk?

It removes one class and creates another. Off-duty conduct, personal scandal and schedule slippage go away. Authored risk arrives: every storyline becomes a brand decision, as the May 2019 Calvin Klein campaign pairing Bella Hadid with Lil Miquela showed — the brand apologised publicly after queerbaiting accusations, and nothing about that episode came from an individual’s behaviour. Budget a pre-publication review with a named owner for the character’s narrative.

What disclosure rules apply to AI creators in 2026?

The FTC’s Endorsement Guides place liability on the advertiser for undisclosed material connections and require it to guide, monitor and correct its endorsers; the current text of 16 CFR Part 255 contains no provision specific to virtual or computer-generated endorsers. In the EU, Article 50(4) of the AI Act requires deployers to disclose artificially generated or manipulated image, audio and video content; those Article 50 obligations start on 2 August 2026, while the penalty regime has applied since 2 August 2025. On platforms: YouTube surfaces labels in the expanded description and, for photorealistic content, in the player; Meta detects AI images via C2PA/IPTC metadata but relies on self-disclosure for AI video and audio; TikTok says it has labelled over 3 billion videos as AI-generated content.

How do we compare the two options if the industry numbers are unusable?

Run a two-arm test on your own audience for one quarter. Fix the denominator for engagement before you start, compare synthetic content against your own standard branded content rather than an industry average, hold the brief, format, window and posting slots constant across both arms, disclose in both, and log all-in cost per published asset the same way on each side. That produces one defensible number for your category — which is more than the public record offers anyone.

Should we replace human influencers with AI creators?

That is the wrong frame for almost every brand we talk to. A synthetic channel is a media property you own and amortize; a human creator is a relationship you rent per post. They fail differently, they are priced differently, and they carry different disclosure loads. Run both against your own numbers — and be sceptical of any comparison, this article included, that hands you a multiplier instead of a mechanism.