Spinning up
Live
+312AI citations this month
Press⌘Kto search
Spinning up
AI-generated ad creative benchmarks from inside 20 paid client engagements and 40 control programs. 4.1M USD spend, 11,000+ tracked variants, 90 days.
Section 01 · Headline numbers
Cost per variant (AI vs UGC)
8%
AI variants average 8% of equivalent human-shot UGC at scale
Creative half-life on Snap
11 days
Median time before CPM doubles or CTR halves
Variant volume per week (DTC)
30-50
Sustained median in tracked DTC programs
Performance gap vs human UGC
≈ 0%
Statistically insignificant on average; channel-dependent
Section 02 · Channel performance
AI creative shares of spend by channel in the tracked corpus. Vertical video channels saw the strongest adoption; brand-trust surfaces still leaned toward human-shot.
Section 03 · Half-life by channel
Median observed half-life. Tracking definition: time from creative go-live until CPM rises by 50% or CTR falls by 50% from launch baseline. Channels with high audience velocity burn creative fastest.
| Channel | Half-life (days) | Implied weekly variants |
|---|---|---|
| Snap | 11 | 4-6 |
| TikTok | 9 | 5-7 |
| Instagram Reels | 17 | 3-4 |
| YouTube Shorts | 22 | 2-3 |
| Meta Feed | 24 | 2-3 |
| LinkedIn Sponsored | 28 | 1-2 |
Section 04 · What wins
Across 11,000+ variants, the openers clustered into five hook patterns. Numbers are share of variants in the top decile of CTR.
Variants that named the specific problem within frame 1 outperformed every other hook pattern. The hook is the lever; the resolution is the variant.
Human face on screen still beat pure motion graphics, even when the founder was AI-rendered with the founder's likeness.
Variants that mimicked the native format of the platform (TikTok meme style, Snap raw POV) outperformed branded formats.
Single bold numeric claim with no hedging language outperformed nuanced claims. Hedging cost CTR.
First line of voiceover named the ICP explicitly (founders building B2B SaaS, DTC operators in the GCC). Self-selection mechanism.
Section 06 · Methodology
Window. January 1 to March 31, 2026. 90 days of continuous tracking.
Cohort. 20 paid Xpand Media AI creative engagements + 40 non-client control programs across DTC e-commerce, hospitality and tourism, and B2C SaaS.
Spend volume. 4.1M USD total cohort spend in the window. Median per-program spend was 38K USD.
Tools tracked. Higgsfield, Runway, Pika, Luma, Sora 2 for video. Midjourney, Recraft, Ideogram for static. ElevenLabs, Hume for voice. Most programs combined 3-5 tools.
What counts as a variant. Any unique combination of script, visual composition, and call-to-action. Minor edits (color, text overlay) do not count as separate variants.
Limits. B2C-heavy sample. Numbers for B2B SaaS creative carry larger error bars; LinkedIn-specific findings draw from a smaller sub-sample.
Section 07 · Frequently asked
Any ad creative where at least one of: the on-screen talent, the voiceover, the visual composition, or the script was generated using a foundation model. Pure copy edits via an LLM count if the AI authored the asset. Tools tracked include Higgsfield, Runway, Pika, Luma, Sora, Midjourney, Recraft, and ElevenLabs.
On average yes. Performance gaps are not measurable inside the noise floor for most channels. The exception is brand-trust hero assets (founder-led, customer testimonial) where human-shot still wins decisively. Use AI for variant explosion, human for hero moments.
AI variants cost 5-12% of equivalent human-shot UGC at production scale. The savings let you fund experimentation budgets that were previously economically impossible.
Video: Higgsfield, Runway, Pika, Sora 2, Luma. Static: Midjourney, Recraft, Ideogram. Voice: ElevenLabs, Hume. Script: Claude, GPT-4. Most teams use a 3-5 tool stack rather than picking one.
Regulation-dependent. EU AI Act requires disclosure on synthetic-talent generative content. US enforcement is patchwork by state. Saudi PDPL guidance recommends disclosure when AI is used to depict real-seeming people. Default to disclosing.
Snap: 11 days. TikTok: 9 days. Meta Reels: 17 days. LinkedIn: 28 days. The half-life is the production-cadence forcing function: if your team can not ship at this volume, you cede share to those who can.
Both. Static still dominates Google Display, LinkedIn news feed, and email. Vertical video dominates Snap, TikTok, Instagram Reels, YouTube Shorts. Treat static and vertical video as complementary, not substitutable.
30-50 for a sustained DTC e-commerce program. 12-20 for B2B SaaS. Below 12 starves the algorithm; above 60 hits diminishing returns inside a single channel.
20 paid Xpand Media AI creative engagements + 40 control programs across DTC, hospitality, and B2C SaaS. Tracked January through March 2026. Total spend in the window: 4.1M USD across the cohort.
Yes. Free to cite under a CC BY 4.0 license. Plain text and BibTeX citation blocks are below.
Section 08 · Cite this report
Plain text
Xpand Media (2026). State of AI Creative 2026: Tracked corpus benchmarks for generative ad creative. https://xpandmedia.io/state-of-ai-creative-2026
BibTeX
@techreport{xpandmedia_ai_creative_2026,
author = {{Xpand Media}},
title = {State of AI Creative 2026: Tracked corpus benchmarks for generative ad creative},
institution = {Xpand Media},
year = {2026},
url = {https://xpandmedia.io/state-of-ai-creative-2026}
}License: free to cite and quote in commercial and academic contexts. Direct link backs to the report URL are appreciated but not required. The underlying anonymized dataset is available on request.
We run the same tracking corpus on every active creative client. A 20-minute call gets you a written read on where your variant velocity sits against the corpus and what the biggest channel-specific gap looks like.