Friday, November 10, 2023
Amazon Rekognition uses OCR to extract text from product images for Rufus AI indexing. Most sellers fail because their image text has insufficient contrast ratios, font sizes below 14pt, or resolution under 1000px wide, making it unreadable to AI systems.
Most sellers think Rufus only reads their bullet points and product descriptions. They're missing a critical piece of the system: Amazon Rekognition, the AWS computer vision service that powers image analysis across Amazon's platform.
Rekognition is the AI layer that extracts text from your product images through optical character recognition. When customers ask Rufus questions about your product, the system doesn't just pull from catalog fields. It's also reading the text overlays, infographics, comparison charts, and benefit callouts in your image stack.
According to testing documented in seller resources, Rufus actively pulls images into conversational responses when they contain relevant information. If the AI can't read your image text, that content is invisible to Rufus regardless of how great it looks to human shoppers.
From our work with 7-figure sellers at Atomic: We've audited hundreds of product image stacks and found that approximately 80% have text that Rekognition cannot reliably extract. The most common issue isn't design quality, it's technical readability for computer vision systems.
Here's what happens behind the scenes when Rufus processes your listing:
If Rekognition fails at step 2 or 3, your image text never makes it into the index. Rufus can't cite information it never received.
Rekognition uses deep learning models trained on millions of images to detect and extract text. The AWS documentation for text detection explains that the service looks for both printed text and scene text (text in real-world environments).
The detection process analyzes:
Rekognition assigns a confidence score (0-100%) to every piece of detected text. Low-confidence detections get filtered out before they reach Rufus. If your image text scores below Amazon's confidence threshold due to poor contrast, small fonts, or visual noise, it's effectively invisible to the AI.
Testing with actual product images shows that text with confidence scores below 85% often gets excluded from indexing. This means your infographic might look perfect to human shoppers but score too low for AI extraction.
Rekognition struggles with text that doesn't have strong contrast against its background. The WCAG accessibility guidelines recommend a minimum contrast ratio of 4.5:1 for normal text, but for reliable AI detection, you need at least 7:1.
Common failures:
Text smaller than 14pt at standard viewing resolution frequently fails OCR extraction. Even if humans can read it when zoomed in, Rekognition processes images at fixed resolutions and can't reliably detect sub-14pt text.
This is particularly problematic for:
Amazon allows images as small as 500px on the longest side, but Rekognition performs best with images at least 1000px wide. Smaller images reduce the effective pixel density of text, making character recognition less accurate.
OCR systems are trained on standard fonts. Script fonts, heavy stylization, text effects (shadows, glows, outlines), and artistic typography significantly reduce detection accuracy.
Fonts that work well:
Fonts that fail frequently:
Rekognition needs clean separation between text and background. Busy patterns, gradients, photographic backgrounds, and overlapping design elements confuse the detection algorithm.
| Element | AI-Unreadable (80% of Sellers) | AI-Readable (Optimized) |
|---|---|---|
| Contrast Ratio | 3:1 to 4:1 (light text on light background) | 7:1 or higher (stark contrast) |
| Font Size | 10-12pt (too small for reliable OCR) | 16pt minimum (14pt acceptable for simple text) |
| Image Resolution | 500-800px wide (Amazon minimum) | 1500-2000px wide (optimal for text detection) |
| Font Style | Script, decorative, thin weight | Sans-serif, medium/bold weight |
| Background | Gradients, patterns, photos | Solid colors, simple backgrounds |
| Text Effects | Shadows, glows, outlines, transparency | Flat, solid text with no effects |
| Layout Complexity | Multi-column, overlapping elements | Clear hierarchy, ample whitespace |
| Rekognition Confidence | Below 85% (often excluded from indexing) | 90%+ (reliably indexed for Rufus) |
Before creating new images, test your existing image stack using the AWS Rekognition demo console. Upload each product image and check the "Text in Image" detection results.
Look for:
When creating infographics or text-heavy product images:
Contrast:
Typography:
Layout:
Resolution:
Remember that extracted text becomes searchable content for Rufus. Structure your image text to answer common customer questions:
Instead of: "Premium Quality"
Write: "Made with Food-Grade Stainless Steel"
Instead of: "Perfect Size"
Write: "Fits Standard Kitchen Counters (12" × 8")"
Instead of: "Multi-Use"
Write: "Ideal for Coffee, Tea, Smoothies & Protein Shakes"
Specific, descriptive text in your images gives Rufus more material to cite when answering customer questions.
Amazon allows up to 9 images. Allocate them based on Rufus impact:
Prioritize text-readable infographics in slots 3-6 where Rufus is most likely to pull visual evidence.
You don't need expensive software to test Rekognition readability. AWS provides a free tier that includes text detection:
Look for confidence scores above 90% for all critical text elements. Anything below 85% may not make it into Rufus's searchable index.
If Rekognition misses text or shows low confidence:
Retest after each change until all critical text scores above 90% confidence.
Find out if your Brand is invisible to Amazons Rufus AI discovery tool and how to close the Gaps