Why visual search treats your product photo as the query
When someone types "black leather chelsea boots," a search engine matches text to text. When they point a camera, the system generates an embedding — a numeric fingerprint of the image — and looks for catalog images whose fingerprints sit nearby. Your product photo is not supporting evidence for the listing. It is the index entry.
That changes what "a good product photo" means. A dramatically lit, heavily styled hero shot can be beautiful and still fail, because the model cannot cleanly separate the product from the scene. Meanwhile a boring, evenly lit shot on a clean background matches reliably — and that is the shot most brands treat as an afterthought.
Optimized for humans only
- One hero angle, heavily styled
- Moody, directional lighting
- Product cropped or partially occluded
- Busy or textured background
- Color graded for mood, not accuracy
Optimized for humans and machines
- 4-6 angles including back and detail
- Even, neutral lighting with soft shadow
- Full product visible, generous margins
- Clean background on the primary image
- Color matched to the physical product
The good news: you do not have to choose. The lifestyle and editorial images can stay exactly as they are. Visual search leans hardest on the primary image, so the fix is usually adding one clean, high-resolution, accurate shot per SKU rather than reshooting a whole catalog.
The image specs that actually matter
Marketplace minimums and visual-search-friendly specs are not the same thing. Amazon will accept a 1,000px image; a matching model gets meaningfully better signal from 2,000px and up, because fabric weave, stitching, hardware, and grain survive the downsample into the embedding.
| Attribute | Bare minimum | Visual-search target |
|---|---|---|
| Longest edge | 1,000 px | 2,000-3,000 px |
| Aspect ratio | Anything | 1:1 square, consistent catalog-wide |
| Product fill of frame | Unspecified | 80-90%, with even margin |
| Primary background | Any | Pure or near-pure white / neutral |
| Images per SKU | 1 | 4-6 distinct angles |
| Format | JPEG | WebP or AVIF served, JPEG/PNG master retained |
| Compression | Whatever the CMS does | Avoid aggressive presets — artifacts destroy texture |
Plenty of stores shoot at 3,000px and then let an image pipeline serve a 600px, quality-60 WebP to every surface — including the one search crawlers fetch. Check what is actually delivered, not what is uploaded.
Multiple angles matter more than most sellers expect. A shopper photographing a bag from behind on a train will never match a catalog that only has a three-quarter front shot. Back, side, top, and a close detail crop each create a separate chance to be the nearest neighbor.
Retouching choices that make or break a match
Editing decisions that read as "polish" to a human can erase exactly the features a model keys on.
Over-smoothing texture
Heavy noise reduction and skin-style smoothing flatten knit texture, leather grain, and brushed metal into a plastic surface. Two different sweaters that both got smoothed to a matte blob become genuinely harder to tell apart.
Cutting out the shadow entirely
A hard cutout floating on pure white loses the depth cue that tells a model where the object ends and how it sits in space. A soft natural or contact shadow keeps the silhouette readable and still looks clean.
Color grading away accuracy
If a shopper photographs a forest-green jacket and your catalog image has been warmed two stops toward olive for a campaign look, the color histogram argues against a match — and the customer who does buy is more likely to return it.
Keep an unretouched master of every SKU. When a matching model or marketplace spec changes, you can re-derive clean variants from the original instead of re-editing an already-edited file.
This is where AI retouching earns its keep: tools like Retouchable can produce a clean-background, correctly-proportioned, color-consistent primary image plus extra angles from existing shots, which is the practical way to bring a legacy catalog up to visual-search standard without rebooking a studio.
The metadata layer still counts — as a tiebreaker
Pixels get you into the candidate set. Metadata decides which candidate wins, and whether the result shows a price, availability, and a buy path instead of a bare thumbnail.
The checklist is short and mostly mechanical:
- Product schema with
image,name,brand,color,material,offers.price, andavailability. List every angle in theimagearray, not just the hero. - Descriptive alt text that names the attributes a camera cannot infer — material, colorway name, fit.
- Readable file names:
ridge-parka-forest-green-back.jpg, notIMG_4417_final_v3.jpg. - Merchant feed alignment — the image in your product feed should be the same clean primary you optimized, not an older asset.
- Crawlable images: no lazy-load pattern that hides the src, no blocking in robots.txt, stable URLs.
One more: if AI tools touched the image, keep the provenance metadata intact. Disclosure expectations for AI-assisted product imagery are tightening across major markets, and stripping metadata during export is the most common way brands lose that record without meaning to.
A practical audit you can run this week
You do not need a platform to start. Pull your 25 best-selling SKUs and work through them.
- Photograph your own products with Lens. Shoot each item from three angles on a phone and search. If your own listing does not come back first, that is your baseline problem, stated precisely.
- Log what is served. For each SKU, record the delivered pixel dimensions, format, and file size of the primary image — not the uploaded master.
- Count angles. Flag anything under four distinct views. Back and detail shots are the usual gaps.
- Check color truth. Compare the on-screen image against the physical product in daylight. Note anything visibly shifted.
- Validate schema. Run the listing through a structured data test and confirm every angle appears in the
imagearray. - Fix the primary first. Clean background, full product, accurate color, 2,000px+. That single image carries most of the weight.
Visual search skews toward items people encounter in the wild — apparel, footwear, furniture, accessories, anything someone photographs on a stranger or in a friend's apartment. Start there rather than working alphabetically.
Re-run the Lens test after the fixes. It is the rare SEO change where you can verify the outcome yourself in about thirty seconds per product.