What you will find in this article
- What SynthID is, Google DeepMind's pixel watermark, and what it is NOT (it is not an EXIF metadata field)
- The distinction between pixel watermarks, IPTC metadata and C2PA metadata, three different signals with different impact on Google Ads
- The complete matrix: which AI tools leave which signal in generated images
- What Google Ads and Merchant Center actually read, official sources and inferences separated
- Behaviours observed on real campaigns, clearly flagged as inferences
- The practical workflow for generating AI images safe for Google Ads without problematic residuals
You generate an image with Midjourney, upload it to AI images Google Ads and it works without issues. The same image made with Photoshop Generative Fill gets flagged by Merchant Center. Same technique, generative AI, different results. The reason is not image quality: it is what remains in the file after generation.
AI tool for Google Ads workflows do not all produce the same type of image from a metadata perspective. Some embed invisible watermarks in pixels. Others insert AI generation information in IPTC metadata. Others still use the C2PA standard to declare content provenance. And some leave nothing detectable, or almost nothing.
This article clearly separates what is officially documented from what emerges from direct file analysis and observation of Google Ads and Merchant Center behaviour.
What SynthID is, and why it is not what many think
SynthID is often misunderstood. The most common confusion is equating it with a visible watermark or an EXIF metadata field. It is neither.
SynthID is currently documented as available for images generated with Imagen, accessible via Vertex AI and Google Cloud. Every image produced by Imagen embeds the SynthID watermark in pixels by default.
The practical question for those using an ai tool for Google Ads is: do Google Ads and Merchant Center use SynthID to detect AI images and apply differential treatment? The documented answer is no, Google has never declared that ad review systems use SynthID as an approval criterion. SynthID is documented as a transparency and attribution tool, not as an advertising filter.
The three types of AI signal in images: pixel, IPTC and C2PA
Before analysing individual tools it is necessary to distinguish between three completely different types of signal, because they have different impacts on Google Ads and Merchant Center.
AI tool matrix: what they leave in Google Ads images
This table summarises the current state (May 2026) for each main tool. Sources are indicated for documented information; for those inferred from file analysis this is explicitly stated.
| Tool | SynthID pixel | IPTC AI Metadata | C2PA | Risk for Google Ads | Source |
|---|---|---|---|---|---|
| Imagen 4 (Google Vertex AI) |
Yes, embedded | Yes (XMP) | Partial | Medium | Official Google DeepMind |
| Adobe Firefly | No | Yes (XMP Adobe) | Yes, embedded | Medium | Official Adobe C2PA |
| Photoshop Generative Fill | No | Yes, relevant | Yes | High for Shopping | Official Adobe + file analysis |
| Midjourney v7 (Discord / API) |
No | No | No | Low | Direct file analysis ⚠️ |
| Flux 1.1 Pro (API BFL) |
No | No | No | Low | Direct file analysis ⚠️ |
| GPT Image 1.5 (ChatGPT / API) |
No | Minimal | No | Low | Direct file analysis ⚠️ |
| Stable Diffusion (self-hosted) |
No | No | No | Low | Direct file analysis ⚠️ |
| Canva AI | No | Yes (Canva metadata) | No | Low-medium | Direct file analysis ⚠️ |
What Google Ads actually reads: Google Ads campaigns and Merchant Center
Merchant Center does not disapprove images because they are AI. It disapproves images that violate specific quality and policy criteria, and the IPTC metadata from Photoshop Generative Fill can trigger these disapprovals indirectly, not because the system "knows" the image is AI, but because the metadata indicates manipulations the system interprets as problematic quality signals.
For Display, Demand Gen, Performance Max and related asset groups, the system is even more permissive: there are no documented cases of disapprovals on Display or PMax assets caused by AI metadata. A Low rating in the Asset Report for visually poor AI images is a separate issue, unrelated to metadata, but tied to the actual visual quality of the image.
Inferences from direct file analysis and observed behaviour
Photoshop Generative Fill: the most critical case for Google Shopping. Analysing with ExifTool the PNG and JPEG files exported after using Generative Fill, you find XMP fields such as xmp:CreatorTool with
value "Adobe Photoshop", photoshop:History with references to generative operations, and in some cases specific C2PA metadata explicitly declaring the presence of AI-generated content. Images with this metadata set received
"Poor image quality" disapprovals more frequently than visually identical images without AI metadata.
Midjourney: clean by default, but variable with the export method. Images downloaded directly from Midjourney on Discord have no relevant IPTC metadata. Images downloaded via API have even cleaner metadata. Verified on Midjourney v6 and v7 output across multiple sessions. Midjourney has no explicit public policy on the absence of watermarks, the situation with future versions could change.
Adobe Firefly: C2PA declared, but practical impact on Google Ads currently limited. I verified on a sample of Firefly images used in Display campaigns that none received disapprovals related to C2PA metadata. Google Ads' review system does not currently appear to read C2PA as a disapproval signal. This could change as C2PA becomes a more widely adopted standard.
Imagen's SynthID: no observable impact on Google Ads today. Images generated with Imagen 4 via Vertex AI used on Display and Performance Max campaigns caused no differential treatment, no disapprovals, no lower ratings, no reduction in distribution compared to Midjourney or Flux images. The distinction between AI identification system (SynthID) and advertising review system (Google Ads) is real and documented by the organisational separation between Google DeepMind and Google Ads.
Practical workflow: how to generate AI images safe for Google Ads
1. Choose the tool based on the destination format. For images destined for Google Shopping (via Merchant Center), the tools with fewest problematic metadata are Midjourney, Flux and GPT Image. Adobe Firefly and Photoshop Generative Fill have more "noisy" metadata for Merchant Center, manageable, but requiring an additional cleaning step. For images destined only for Google Display Ads or Performance Max asset groups, any tool works without documented issues.
2. Apply metadata stripping before any upload. This practice is recommended regardless of the tool. The ExifTool command:
exiftool -all= image.jpg In Photoshop: File → Export → Save for Web (Legacy), which removes IPTC
and XMP metadata by default.
3. Verify metadata after stripping.
exiftool image.jpg | grep -i "creator\|tool\|software\|AI\|generative" If the command returns no
output on these fields, the file is clean of relevant metadata.
4. Do not worry about the SynthID pixel watermark in Google Ads today. SynthID is in Imagen's pixels, it is not removable with metadata stripping. But Google Ads and Merchant Center do not currently use SynthID as a review criterion. Focus attention on XMP metadata (removable) and the visual quality of the image relative to format requirements.