Easy to see
Transparent Labs · Protein Coffee
The tub is blank. No logo.
Taste Bench · Harness Evals · Oct 2026
Chat apps make pictures.Omneky makes ads.
52 ad requests across 4 brands · 364 ads.
Last updated:
Ready to run
Same checks for every app.
- 160%
Omneky - 213%
ChatGPTGPT-6 Astra, Medium Reasoning - 38%ChatGPTGPT-6 Sol, Medium Reasoning
- 48%
GeminiGemini 3.6 Flash, High Thinking - 56%ChatGPTGPT-6 Luna, Medium Reasoning
- 66%GeminiGemini 3.6 Flash, Low Thinking
- 72%GeminiGemini 3.1 Pro, High Thinking
A check passes only when more judges pass it than fail it (n/a doesn't count).
- 7
- Apps compared
- 52
- Ad requests
- 364
- Ads judged
- 4
- Brands
Leaderboard
Ready to run
Same checks for every app.
Omneky picks the model, so you don't have to. See Model Evals
Request
52 ad requests in this view.
- Ready
- 60%
- Ad score
- 8.8
Most common failures
Safe zone 40% · Font 2%
- Ready
- 13%
- Ad score
- 7.6
Most common failures
Logo 56% · Safe zone 46% · Product 40%
- Ready
- 8%
- Ad score
- 7.4
Most common failures
Logo 63% · Product 52% · Safe zone 50%
- Ready
- 8%
- Ad score
- 7.2
Most common failures
Product 75% · Logo 50% · Font 35%
- Ready
- 6%
- Ad score
- 7.3
Most common failures
Product 73% · Logo 71% · Safe zone 42%
- Ready
- 6%
- Ad score
- 7.0
Most common failures
Product 85% · Logo 52% · Safe zone 31%
- Ready
- 2%
- Ad score
- 6.8
Most common failures
Product 83% · Logo 63% · Font 35%
Why ads were not ready
Share of that app's ads that failed the check.
| App | Product | Logo | Font | Safe zone | Claims | Language | Gibberish |
|---|---|---|---|---|---|---|---|
| Omneky | 0% | 0% | 2% | 40% | 0% | 0% | 0% |
| ChatGPT(GPT-6 Astra, Medium Reasoning) | 40% | 56% | 29% | 46% | 0% | 0% | 0% |
| ChatGPT(GPT-6 Sol, Medium Reasoning) | 52% | 63% | 33% | 50% | 4% | 0% | 0% |
| Gemini(Gemini 3.6 Flash, High Thinking) | 75% | 50% | 35% | 29% | 4% | 6% | 8% |
| ChatGPT(GPT-6 Luna, Medium Reasoning) | 73% | 71% | 29% | 42% | 4% | 0% | 0% |
| Gemini(Gemini 3.6 Flash, Low Thinking) | 85% | 52% | 29% | 31% | 12% | 4% | 10% |
| Gemini(Gemini 3.1 Pro, High Thinking) | 83% | 63% | 35% | 31% | 4% | 6% | 12% |
Side by side
Pretty is not ready.
The picture can look as good, or better. It still has to be the ad.
Easy to miss
Asarai · Earth Tones
The logo and the button sit outside the safe zone.
Every ad
Same request, every app.
Each row is one ad request. Click an image for the checks.
52 ad requests in this view.
| Request | Omneky | ChatGPT(GPT-6 Astra, Medium Reasoning) | ChatGPT(GPT-6 Sol, Medium Reasoning) | ChatGPT(GPT-6 Luna, Medium Reasoning) | Gemini(Gemini 3.6 Flash, Low Thinking) | Gemini(Gemini 3.6 Flash, High Thinking) | Gemini(Gemini 3.1 Pro, High Thinking) |
|---|---|---|---|---|---|---|---|
Exact copy Asarai Earth Tones Arabic · 1:1 | |||||||
Brand and product only Asarai Earth Tones Arabic · 1:1 | |||||||
Story and feed Asarai Earth Tones Arabic · 4:5 | |||||||
Story and feed Asarai Earth Tones Arabic · 9:16 | |||||||
Pasted prompt Asarai Earth Tones Arabic · 1:1 |
The checks
What ready means
Miss one check, and the ad is not ready.
Product
The real product, in the pack you sell. A lookalike fails.
Logo
The brand's logo, copied exactly. A redraw fails.
Font
The brand's typeface, not a similar font.
Safe zone
Logo, copy, and product stay clear of the edges, where buttons and captions sit.
Language
The words are in the language of the request.
Claims
No claim the brand did not make.
Gibberish
Text is real words, not garbled letters.
Text
The ad has the text it was supposed to have.
Anatomy
People and objects are drawn correctly.
How it's judged
Three anonymous judges. Same bar for every app, including Omneky.
The ChatGPT and Gemini runs used their APIs with the same tools the apps use (web search and image generation). Spot-checks in the ChatGPT and Gemini apps gave similar results.