How much does AI product photography increase conversion?

No reliable figure was found as of October 2026. The only numeric conversion result found, +17.27% from an Alibaba test, compared AI-designed items with human-designed items, so the product changed as well as the image. It reports no sample size.

Is a 30% conversion lift claim from a vendor believable?

It is unverified until the vendor shows its sample size, baseline, test design and confidence interval. A real 30% lift at a 2.5% conversion rate needs about 7,780 visitors per variant to detect, so a vendor with enough traffic can publish that evidence.

Do AI backgrounds beat plain product images?

In advertising tests run by Taboola on its own platform, generated backgrounds raised click-through by about 15%, within a range of about 4% to 40%. The paper gives no conversion figures, and the setting was ads, not product pages.

Do AI product images reduce returns?

This is not established. An Alibaba paper reports 7.9% lower returns for AI-designed items against human-designed items in the same stores, but product design differed as well as imagery. No independent study of returns on AI photos of the same product was found.

How long should an A/B test of product images run?

It should run until it reaches the sample calculated in advance, in whole weeks. At a 2.5% conversion rate and a target of a 10% lift, a store with 5,000 daily visitors to the tested pages needs about 26 days.

Does AI product photography increase sales? 2026 evidence

Is there independent proof that AI product photography increases sales?

No independent, peer-reviewed field test showing that AI product photography raises sales or conversion was found in this review, dated 31 October 2026. The closest evidence is company-run tests that mostly measure clicks, one preprint with mixed results, and a large field study that did not touch images.

Absence of proof is not proof of no effect. An AI image may well outperform a hanger photo or fill a gap where no on-model image exists. That has not yet been shown with data a reader can check.

Study	Who ran it	What changed	Result	Limit
Czapp et al., RecSys 2024	Taboola, on its own platform	Generated backgrounds on product images in retargeting ads, catalogues mostly apparel	About 15% higher click-through over originals; about 4% to 40% across later tests	Ads, not product pages. Conversions said to improve, with no figures.
Chen et al., WWW 2025	JD.com with university authors	Generated ad images, online A/B test, over 60 million impressions	2% higher click-through	No conversion or sales figures.
Lin et al., KDD 2026 ADS Track	Alibaba with Shanghai Jiao Tong University	AI-designed items shown on AI models, six-month test across 17 categories, against human-designed items in the same stores	Search click-through +17.81%, recommendation click-through +13.35%, conversion +17.27%, returns -7.9%	Product and image both differ. No sample size or significance reported.
Pec, Proszowska and Rzewnis, preprint, August 2026	The three authors; affiliations not in the record read	Paired Vinted listings, original photos against AI images, two seasonal waves	AI better on attention in autumn and winter; originals better in spring and summer, especially on conversion-related measures	Not peer-reviewed. Only the abstract was read; it gives no effect sizes.
Ghose, Lee, Todri and Adamopoulos, 2025	NYU Stern and Emory	Fully AI-made ads against expert-made ads, lab and field	Click-through up to 19% higher; disclosure cut effectiveness by up to 31.5%	Working paper on ads, not product pages.
Fang et al., arXiv, version of June 2026	Cross-border online retailer and academics	Seven randomised workflows, up to 13.7 million consumers each, September 2023 to June 2024	Sales effects from none detectable to +16.3%, mainly through conversion, with no worse returns or ratings	No workflow generated or changed product images. Treat it as evidence about chatbots and text.

Why do click-through gains not prove sales?

Click-through measures interest before the product page, so it cannot show whether a shopper bought or kept the item. A striking image can raise clicks and then disappoint on arrival. Among the image studies above, only the Alibaba paper gives conversion and return figures in the text read for this page, and its items were also newly designed by AI.

Expect small effects even in good tests. A 2026 paper by Hut and Masoero describes 67 historical creative A/B tests at a large e-commerce service: it calls the percent impacts small, and the outcomes approximately balanced across harmful, inconclusive and launch.

Vendor statements add less than they appear to. Zalando's vice president of content solutions told Reuters in May 2025 that AI content drives greater customer engagement, with no figure attached. A December 2025 guest post by Minimist on the Charity Retail Association site describes controlled tests of AI model images on click-through, conversion and returns, but gives no sample sizes, durations or effect sizes. Figures of this kind cannot be checked and should not guide a budget.

What is known about product images and conversion in general?

Product images strongly shape buying decisions, whether or not AI made them. In usability testing by Baymard Institute, which sells UX research, 56% of users investigated product images as their first action on a product page (2018), and 42% tried to judge a product's size from its images (2017, repeated in Baymard's March 2026 update).

On-model images fill a known gap. Baymard's article of 2020 says cut-out images do not show fit, length or proportion, and its 2026 update reports that 23% of benchmarked sites provide no human-model image for apparel, accessories and cosmetics.

Model size matters. Zhang, Ikonen, Eelen and Sotgiu (Journal of the Academy of Marketing Science, 2024) ran three studies and eight experiments. Thin models deterred shoppers who wear larger sizes, and models of the shopper's own size improved online purchase decisions.

Professional images can raise demand. Zhang, Lee, Singh and Srinivasan (Management Science, 2022) studied 7,423 Airbnb properties over 16 months and found about 8.98% higher occupancy with verified professional photos than with host photos. The setting is property rentals, not retail.

None of these sources tested AI-generated images. AI may deliver more on-model, size-varied or in-scale images at lower cost, and that is a reasonable hypothesis, not a finding. Baymard adds a caution: computer-generated images should supplement real in-scale photography, not replace it, a view published in 2017 before current image models.

Add on-model images with Virtual try-on

What evidence would prove that AI photos raise sales?

Proof requires a randomised test of the same product, at the same price, in the same period, with only the image changed. A before-and-after comparison fails this standard, because season, promotions and traffic also change.

The comparison must be named. AI against a phone photo, AI against a missing on-model image and AI against a professional shoot are three different questions, and a win on the first says nothing about the third.

The outcome must be money, not attention: orders, revenue per visitor and return rate. An image that lifts orders and also lifts returns may lose money.

The result must be checkable: published sample size, confidence interval and test duration, from someone without a commercial interest, and repeated. Kohavi, Deng and Vermeer (KDD 2022) warn that some misleading ideas about A/B testing have been promoted by tool vendors and agencies.

How many visitors does an A/B test of product photos need?

At a 2.5% conversion rate, detecting a 10% lift needs about 64,124 visitors per variant, or 128,248 in total. The 2.5% is an example, not a benchmark. Use your own rate.

The calculation below uses the standard two-proportion approximation, with a 5% significance level, two-sided, and 80% power. Those settings give (1.96 + 0.84) squared, which is 7.84. Then n per variant = 7.84 × [p1(1 − p1) + p2(1 − p2)] ÷ (p2 − p1) squared. This is the same family as the 16σ²/δ² rule of thumb described by Zhou, Lu and Shallah (2023).

Worked example: p1 is 0.025 and a 10% lift makes p2 0.0275. Then p1(1 − p1) is 0.024375 and p2(1 − p2) is 0.026744, which sum to 0.051119. The difference squared is 0.00000625. So n = 7.84 × 0.051119 ÷ 0.00000625 = 64,124, rounded up.

A simulation of 200,000 tests at that size, run for this page, detected a true 10% lift 79.9% of the time and gave a false positive 5.0% of the time when no lift existed.

Small tests mislead. With 5,000 visitors per variant at the same 2.5% rate, the simulation found a true 10% lift only about 12% of the time. That size is adequate only for a lift of about 38%, far larger than most creative changes produce.

Baseline conversion	To detect +5%	To detect +10%	To detect +20%	To detect +30%
1%	636,287	162,908	42,642	19,802
1.5%	421,994	108,028	28,269	13,123
2%	314,847	80,588	21,082	9,784
2.5%	250,559	64,124	16,770	7,780
3%	207,700	53,148	13,896	6,445
5%	121,983	31,196	8,146	3,773

How long will that test take?

Duration is the total sample divided by daily visitors to the tested pages. For 128,248 visitors in total, the examples below apply. Run whole weeks whatever the arithmetic says, because weekday and weekend shoppers differ.

Daily visitors to tested pages	Days needed	Weeks
200	642	91.6
1,000	129	18.3
5,000	26	3.7
20,000	7	0.9, so run at least one full week

Does the test also measure returns?

A test sized for conversion usually cannot measure returns. The worked example yields about 1,603 orders in the control group. Detecting a return rate that moves from 20% to 22% needs about 6,500 orders per variant: 7.84 × (0.16 + 0.1716) ÷ 0.0004. Plan a longer test, or treat returns as a warning signal only.

How do you run a valid product photo test?

A valid product photo test changes one image set, splits real visitors at random, and fixes its size before it starts. Follow these steps in order.

1. Write one question, such as: does adding an AI on-model image beat our current flat lay for these products?

2. Randomise by visitor, not by session. Zhou, Lu and Shallah note that repeat observations from one user are correlated, so session-level calculations can leave a test underpowered.

3. Keep price, title, stock, promotions and page layout identical across variants.

4. Calculate the sample from your own conversion rate, and stop only when you reach it. Stopping early at the first good result inflates false positives.

5. Record orders, revenue per visitor and return rate by variant, and read the confidence interval, not only the p-value.

6. Repeat a surprising win on a different product group before changing the whole catalogue.

Create test images with Product studio

When is a conversion test the wrong tool?

A conversion test is the wrong tool when a store cannot reach the required sample in a sensible time. At 200 visitors a day, the worked example takes 642 days, and season, stock and pricing will change long before it ends.

Smaller sellers should decide on cost, accuracy and coverage instead. Ask whether a listing lacks an on-model image, a back view or a scale reference, and whether an AI image can supply it accurately. The sibling page on whether AI product photography is worth it covers that decision.

A traditional shoot may remain the better purchase for a hero campaign or fit-critical garments, because this evidence cannot show that AI matches it. Where a product is unique or second-hand, keep the real photo: the Pec preprint found originals ahead on conversion-related measures in one of its two waves.

Check a listing image with the free product photo checker

Checklist

Name the comparison: AI against which current image, on which products.
Use your own conversion rate in the formula, and set the sample before launch.
Randomise by visitor, and keep price and page content identical.
Run whole weeks, and do not stop at the first good result.
Track orders, revenue per visitor and return rate, not clicks alone.
Treat any uplift claim without sample size and confidence interval as unverified.
Check every test image against the real product before it goes live.

Sources

Checked 31 October 2026. Rules change; the official pages are the final reference.

Does AI product photography increase sales?