Looking real is free now. Proof isn't.
AI can generate a convincing customer photo for a few cents, so looking authentic no longer proves anything. What still works is proof a shopper can check, and most stores are already sitting on it without surfacing it.
Looking real used to be the whole trick
The original case for customer photos was simple, and we made it in an earlier post: a studio shot says this is what we want it to look like, a customer photo says this is what it actually looks like. The customer photo won because it looked like it came from a real room, a real phone, a real person.
That argument has quietly lost half its force. Generating an image of a candle on a slightly messy kitchen counter, shot at an amateur angle in imperfect afternoon light, now costs a few cents and about eleven seconds. The visual grammar of authenticity is available to anyone now: the crooked framing, the blown-out window, the dog in the background. That includes a store selling a product that doesn't exist.
Looking real is free now. Customer photos still work, but the thing that makes them work has moved. If you're still optimizing for "looks authentic," you're optimizing for a signal that no longer carries information.
Shoppers don't detect AI. They detect error.
Most merchants assume shoppers can smell an AI image and recoil. Survey work through 2026 keeps finding something less flattering and more useful: in a blind comparison, most shoppers can't reliably pick the synthetic image out of a lineup, and when they're told afterward that an image was AI-assisted, the majority reaction is a shrug.
What does damage trust is visible inaccuracy. Fabric that drapes in a way fabric doesn't. A color that shifts between the third and fourth image. A handle meeting the pot at an angle no manufacturing process would produce. Shoppers rarely conclude "this image is AI." They conclude "something is off with this seller." That's the more expensive verdict, because it generalizes to your shipping estimates and your return policy too.
So the aesthetic of realness has been commoditized, and the premium has moved to something harder to manufacture.
The suspicion has moved to the text
While everyone was watching the images, the credibility of review text collapsed underneath them.
Written reviews are now the cheapest content in existence. A model will produce two hundred plausible, varied, specific-sounding reviews for a product it has never seen, complete with the small complaints that make a review feel honest. Shoppers know this, and studies of review-reading behavior consistently find a large minority who start from the assumption that some meaningful share of what's in front of them is fake, AI-written, or paid for.
The five-star rating has stopped functioning as evidence. It functions as a claim, and claims are what shoppers learned to discount.
Which leads to the shift worth internalizing: a photo attached to a verified order is now doing the verification work the star rating used to do. Not because it's prettier or more informative, but because it's the part of the review with a receipt behind it.
Four kinds of proof you can actually ship
"Be authentic" is not an instruction. These are, roughly in order of effort-to-payoff.
1. Purchase linkage, and say it out loud. A submission tied to a real order is the strongest and cheapest signal available, and most stores already have it and never surface it. The second half is the critical one: an unlabeled proof signal is not a proof signal. A gallery photo with no badge is indistinguishable from a stock image someone dropped in. The same photo carrying "Photo from a verified order · March 2026" is a different object. Label it where the shopper is looking, not in a tooltip.
2. Context that's expensive to fake. Ask for the specific, not the beautiful. A hand in frame. The product in a room with other objects in it. Visible wear after two months. A generic hero-angle shot on a clean background is precisely what a model produces for free, which makes it the category where your customer's photo has the least advantage. The photos worth collecting are the ones a generator has no reason to produce.
3. Provenance metadata, with realistic expectations. Content Credentials (the C2PA standard, backed by Adobe, Microsoft, and a growing set of camera makers) attach a signed manifest to a file recording what device or software created it, whether generative AI was involved, and what edited it since. Recent phones can sign at capture; major AI image tools embed credentials marking their output as generated.
Be precise about what that proves. A valid manifest proves this file came from this device or tool and hasn't been altered since it was signed. It does not prove the scene in front of the lens was real. You can point a credential-signing camera at a screen. It's a provenance record, not a truth oracle.
And there's a trap worth knowing before you build on it: most upload pipelines destroy the manifest. Any step that recompresses or re-encodes an image strips the C2PA container along with EXIF. A CDN transform does it. So does a thumbnail generator, or most naive resize code. If provenance matters to you, verify it at ingest and store the result as a property of the submission. Don't assume the file a shopper downloads still carries it. It almost certainly doesn't.
4. Disclosed incentives. Merchants hide the discount code assuming disclosure undermines the review. The evidence points the other way: shoppers told that a reviewer received 10% off, who then see a mix of positive and critical reviews, trust the set more than an unexplained wall of five stars. Concealment is what reads as manipulation. And disclosure happens to be the law.
The compliance line, briefly
The FTC's Rule on the Use of Consumer Reviews and Testimonials has been in force since October 21, 2024, and carries civil penalties per violation, where each non-compliant post can count separately. It has no AI-specific provisions and doesn't need them: AI-generated reviews are fake reviews, and the rule already covers those.
Three working rules cover almost everything a normal store does:
- Reward the submission, never the sentiment. Offer the discount for posting a photo, never for posting a positive photo, and never contingent on the rating. It's a one-line change in most review-request emails, and it's the difference between an incentive and a purchased review.
- Disclose the incentive where the review appears. Not on the terms page. Next to the review.
- Never selectively remove negative content. Suppression is explicitly covered by the rule, and it's the failure mode a moderation queue drifts into on its own. Our moderation workflow draws the line where it belongs: reject for safety and relevance (wrong product, product not visible, someone in frame who didn't consent), never for tone.
Don't polish it into the uncanny pile
Here's the operational consequence most likely to save you from an own goal.
When a customer photo comes in slightly dark, slightly crooked, shot against a cluttered counter, the instinct is to fix it. Auto-crop to square. Bump the exposure. Run it through an upscaler. Normalize the color so the gallery looks like a gallery.
Every one of those steps moves the photo toward the synthetic-looking pile and erases the signal you spent a discount code to acquire. Worse, upscalers and generative "enhance" features introduce exactly the artifacts that shoppers read as something is off with this seller: smeared texture, invented detail, plasticky skin. Imperfection is load-bearing now. Straighten a horizon if you must; leave the rest alone.
What this means for your gallery
Things you can do this week, none of which require a redesign:
- Change your request email from "leave us a review" to "show us where it lives." You're asking for context, not praise.
- Disclose the reward inline with incentivized submissions.
- Audit your image pipeline for what it strips, and record provenance at ingest if you plan to use it at all.
- Check that mildly critical submissions still appear. If none ever have, your moderation is suppressing, whether or not anyone decided to.
The earlier post called UGC a trust-density problem rather than a content-volume one. That's still right, but the definition of density has changed. It used to mean how believable does this look. Now that believability is free to manufacture, it means something narrower and more durable: how much of what you're showing a shopper could they actually check?
Every disclosed incentive, every unretouched photo tied to a real order number is a claim with a receipt behind it. Those are the only claims that still cost something to make, which is exactly why they're the only ones still worth anything.