What you'll find in this article
- How automatic text detection works in Google Ads, OCR and computer vision, official sources and inferences clearly separated
- Exactly what the system detects: promotional text, watermarks, logos, prices, badges, QR codes
- The 20% rule for Responsive Display Ads, officially confirmed, often misread
- Shopping disapprovals for text on image, "Promotional overlay" and "Text on image" explained
- Patterns observed directly on live campaigns, clearly flagged as inferences
- How to structure creatives to pass detection without losing your message
You upload an image to your Google Ads campaign with the product price in plain sight, "€49.90, Today Only". The system rejects it or lowers the asset score. You remove the price, re-upload, and the problem disappears.
No human looked at the image. An automated system scanned every pixel to identify what was written on it. This system runs across all Display, Shopping, Performance Max, and Demand Gen campaigns that use images. It works silently, often without any error message, and Google has never published a full technical description of how it operates.
This article keeps official documentation and direct observations strictly separate.
How Automatic Text Detection Works: OCR and Computer Vision
Google's ad review system uses two distinct technologies to find text in uploaded images. Both appear in Google's public technical documentation, though neither has been described in detail specifically for ad review.
The two technologies that make up the detection system:
The process is fully automated. For the vast majority of assets, even those with clear text overlays, the system decides on approval or rejection with no human in the loop. A person only steps in if the advertiser files an appeal, and that covers a small share of cases.
What the System Detects Exactly
The official documentation lists the overlay categories the system flags. The section below combines confirmed information with patterns observed directly. Each source is clearly labelled.
Officially confirmed categories
- Overlaid promotional text, prices, discount percentages, claims ("Today Only", "Free Shipping"), calls to action ("Buy Now", "Learn More")
- Watermarks and photo signatures, brand logos, copyright notices, photographer credits
- Corporate logos placed on top of the product
- Badges and graphic stickers, "NEW", "SALE" icons, star ratings built into the image
- Contact details, phone numbers, URLs, email addresses
What is NOT flagged as a problematic overlay
✅ Text that is part of the real scene
A licence plate on a car, text printed on a T-shirt, a logo on a photographed product, a sign in the background of a lifestyle shot. These are part of the subject, not digital overlays. The system may detect them but does not treat them as a violation.
✅ Brand name built into the design
When the brand name is a design element, not a watermark on top, but part of the visual layout, the system tends not to flag it. Text woven into the composition differs from text sitting above the image.
❌ Promotional text added in post-production
Any text placed on top of the product image after shooting, prices, discounts, claims, CTAs. The system catches both the content (via OCR) and the overlay nature (position, contrast). For Shopping this almost always triggers a disapproval. For Display it lowers the asset score.
⚠️ Small text at the image edge
A small copyright line in the bottom-right corner or a minimal watermark in a peripheral zone. In practice the system seems more tolerant here. This is a pattern observed, not a published exception, so treat it as a data point, not a safe harbour.
The 20% Text Rule for Google Display Campaigns
The 20% rule is the most-cited, and most-misread, requirement for Display image creatives.
The most common misconception: many advertisers assume the 20% rule covers the final ad, the version the user sees, with Google's headline and description on top. It does not. The rule covers the uploaded image only. The text Google adds in the assembled ad is handled by a separate system and is not counted in this calculation.
One more factor worth noting: the 20% check runs on the cropped version of the image, not just the original. If Smart Cropping converts a landscape image to square, the text area share can shift. Text that covered 18% of the landscape original might cover 28% of the square crop.
Text Disapprovals: Display vs Shopping vs PMax
The detection system behaves differently across formats, and so does the consequence when it fires.
| Format | Consequence | Error code | Recoverable |
|---|---|---|---|
| Google Shopping (Merchant Center) | Immediate product disapproval, product stops showing in ads | "Promotional overlay on image" / "Text on image" | Yes, fix the image and update the URL |
| Responsive Display Ads | Reduced distribution, lower asset quality score, no formal disapproval | Warning in Asset Report | Yes, replace the asset |
| Performance Max asset group | "Low" rating in Asset Report, reduced reach on sensitive placements | Low rating, no explicit error | Yes, update the asset |
| Demand Gen / Discovery | Reduced distribution without formal disapproval | Asset Report warning | Yes, replace the asset |
The key difference between Shopping and Display: Shopping issues a formal disapproval that cuts product visibility immediately. Display never formally disapproves, but quietly reduces distribution. The damage builds slowly and often stays invisible in standard reports unless you actively monitor individual asset ratings.
Patterns Observed Directly: Cases Not in the Documentation
The system tells readable text from decorative text, but the line is not fixed. Images with highly stylised type, distorted decorative fonts, layered italics, elaborate handwriting sometimes pass without being flagged. OCR likely struggles with fonts that look very different from standard characters. That said, this is not a reliable workaround, results vary font by font.
Detection seems more sensitive on white backgrounds than on dark or textured ones. Light text on a dark background is harder for standard OCR to catch. In practice, images with light text on dark backgrounds tend to receive fewer penalties than images with dark text on light backgrounds, same font size, same content. Use with caution before building it into your creative process.
Text inside geometric shapes gets flagged more consistently. A claim placed inside a circle, a hexagon, or a coloured rectangle triggers detection more reliably than the same text on an open background. The system appears to recognise "text inside a shape" as a strong signal of an artificial overlay, a badge looks like a badge, regardless of what it says.
Google Shopping's detection is stricter than competing platforms. The same images, same amount of text on top, consistently get flagged on Google Shopping while passing on social ad platforms. Google Merchant Center introduced progressively tighter image requirements from 2022 onward, with further tightening in 2024–2025.
How to Structure Creatives to Pass Detection
The goal is not to fool the system, it is to route the commercial message through the right channels so the image itself stays clean.
- Move promotional text out of the image. Prices, discounts, claims, and CTAs belong in the ad's headline and description fields. Google places them on the image in the final format, the user sees the same result, but the system does not count them as overlays on the uploaded image.
- For Shopping: keep the main image clean and use
additional_image_linkfor lifestyle variants. The primary image (image_link) must have no overlays. Additional images viaadditional_image_linkface slightly looser requirements. - Turn on Automatic Image Improvements to fix existing overlays at scale. Google Merchant Center can automatically remove watermarks and overlays from your main image. Enable it under Growth → Manage Programs. It works well on the most common overlay types but will not handle every case.
- Check detection before launch with Google Vision AI. The Google Cloud Vision Text Detection API returns a bounding box for every text element it finds. Add up those boxes and divide by the total image area to get an estimate of text coverage. It is not identical to Google's internal calculation, but it is the closest publicly available tool.
- For Display banner assets, apply the visual-fifth check. If the text in the image takes up more than roughly one fifth of the frame by eye, it is likely over the 20% limit. A rough guide, not a measurement tool, but useful as a fast pre-upload check.
- Update the image URL every time you correct an overlay. Keeping the old URL after editing the image means waiting up to 6 weeks for a re-crawl. Adding even a version tag to the URL (e.g.
?v=2) triggers re-processing within 24–72 hours.