Evidence · strength 0 of 5 · searched to 2 Nov 2026
Does AI-written marketing content sell more?
Not enough evidence yet to say whether AI-written marketing content sells more.
What platform experiments found when online shops and advertisers used AI-written product descriptions, ad copy, push messages and ad images: sales, clicks and returns.
VerdictNot enough evidence yet
No good evidence
ImpactNot shown
We show no impact figure until at least two studies qualify.
Evidence strength0 of 5
There is not enough evidence to rate this yet.
Cost per user a year$
Content generation costs little per item at platform scale; the studies do not report it.
Pilot rating under our v0.3 rules. The rules are published on our method page but not yet independently reviewed, so ratings may change as the method is checked and new studies arrive. Our method.
What the studies found
No impact figure yet. Fewer than two studies qualify, so under our rules we do not show how big the effect is. The studies we found are listed below, with what each measured.
How sure are we?
- Starting level2 of 5, from 3 studies graded A, B or C.
- Down one: old models0 of 4 studies show they used a model released since May 2025 (the 18-month window); the rest used older models or do not say which.
- Down one: not independent3 of 4 studies were run, funded or written by the company whose AI was tested.
- Down one: mostly not peer-reviewedMost of the weight is on working papers or preprints, together with another problem above.
- Publication biasnot testable with fewer than 10 studies.
- Result0 of 5. There is not enough evidence to rate this yet.
Behind the average
- Products whose original description had fewer than 50 words (English sub-experiment) (Fang et al. (2026)): Sales rose 6.5% with AI-generated descriptions (p < 0.05)
- Products whose original description had more than 50 words (English sub-experiment) (Fang et al. (2026)): No significant sales increase from adding AI-generated descriptions (50-word cut-off)
- AI-generated images that raters judged 'definitely human-made' (Exner et al. (2025)): Higher click-through than human-made images: +0.2872 log-odds (SE 0.0995, p < 0.01)
- AI-generated images that look more like AI (Exner et al. (2025)): Click-through falls as perceived artificiality rises: interaction -0.5175 log-odds (SE 0.1764, p < 0.01)
How, where and why it works
Why it works
1 of 4. AI fills in missing or thin product information; gains were largest where the original description was short.
For whom and where
2 of 4. Short original descriptions gained most in one trial; AI images that looked artificial lost clicks in another.
What it takes
1 of 4. Most studies were run by the platforms with their own tools, tuned on their own data.
The full cost
0 of 4. No study counts the full cost.
Four separate scores from 0 to 4, never added up: how much the studies tell us about why it works, for whom, what it takes and what it really costs (the EMMIE method).
What to do next
Measure your pilot →If you are trying it yourself
- Use AI to fill gaps in thin product pages first, and check facts and claims.
- A/B test AI copy against your current copy before switching.
If you are rolling it out
- Run randomised tests on sales and returns, not just clicks.
- Label or review AI images; artificial-looking ones did worse.
What it takes
- Gains came from AI tuned to the platform: plainly prompted titles did worse than human ones in one study.
What can go wrong
- AI ad images that look artificial got fewer clicks.
- AI-optimised Google Shopping titles did not raise sales in one trial.
The studies
| Study | Grade | Design | People | AI model (released) | Company role |
|---|---|---|---|---|---|
| Generative AI and Sales Productivity: Field Experiments in Online Retail (Product Description, Marketing Push Message and Google Advertising Title experiments) Lu Fang, Zhe Yuan, Kaifu Zhang, Dante Donati, Miklos Sarvary, 2026. arXiv 2510.12049v6 (working paper) | B | field RCT, deployed work | Product descriptions: 4,772,937 consumers viewing about 45,000 randomly selected platform self-sold products (five language sub-experiments); marketing push messages: 13,715,528 consumers (first day analysed); Google advertising titles: 1,244,016 products randomised | Unnamed GenAI models built and deployed by the platform's own technical team (paper does not name them; the ad-title model was not fine-tuned for e-commerce advertising) | authors employed by vendor: The tools were built in-house by the unnamed partner platform (a leading cross-border e-commerce retailer) |
| AI in disguise - How AI-generated ads' visual cues shape consumer perception and performance Yannick Exner, Jochen Hartmann, Oded Netzer, Shunyuan Zhang, 2025. Author PDF dated 14 January 2025 (working paper) | C | other quasi-experiment, deployed work | 4,633 ads in 1,186 within-campaign quasi-experiments (351 campaigns, 202 advertisers; 29,592 ad-day observations), drawn from 305,121 ads by 7,074 advertisers on Taboola | Stable Diffusion 2 (images, inside Taboola's GenAI Ad Maker) (Nov 2022); OpenAI GPT-3.5 (captions) | none |
| LLMs for Customized Marketing Content Generation and Evaluation at Scale Haoran Liu, Amir Tahmasbi, Ehtesham Sam Haque, Purak Jain, 2025. arXiv 2506.17863v1 (preprint) | C | field RCT, deployed work | Round 1: 3,000 search-ad keywords; round 2: 10,000 keywords, randomised to keyword-specific LLM ad copy vs template-based copy on a major search-engine advertising service | Anthropic Claude Haiku via Amazon Bedrock (version not stated) | authors employed by vendor: Three authors are Amazon employees and the first author did the work as an Amazon intern |
| CTR-Driven Ad Text Generation via Online Feedback Preference Optimization Yanda Chen, Zihui Ren, Qixiang Gao, Jiale Chen, Si Chen, Xubin Li, Tiezheng Ge, Bo Zheng, 2025. arXiv 2507.20227v3 (preprint) | D | field RCT, deployed work | Evaluation A/B test: 1 million items, generated title shown alongside the human-crafted title, 5% of platform traffic for two weeks; training test: 200,000 items, 1 million candidate titles | Qwen-2.5-7B, fine-tuned (GPT-4 used to generate training reasoning) (Sep 2024) | authors employed by vendor: Seven of eight authors work for Taobao & Tmall Group of Alibaba, which built and tested the method on its own shopping platform's ad slots |
Grades: A, a well-run randomised trial in real work; B, a strong natural experiment in real work or a well-run trial on realistic tasks; C, a trial with clear limits or a matched comparison; D, before-and-after, surveys or case studies. How studies are graded.
Does the rating hold up?
- Run without company-run studies1 studies: strength 0, verdict no good evidence, headline band no clear effect.
- Run without working papers and preprints0 studies: strength 0, verdict no good evidence.
How up to date is this evidence?
Method →Evidence searched to 2 Nov 2026. Studies cover models released Nov 2022 to Sep 2024. Rating computed 5 Nov 2026 under Module 15 v0.3 (pilot).
How this page was made: AI research agents searched for studies and read each paper; a second pass checked every number on this page against the paper itself, with the quote and where it appears. The rating is computed by fixed rules, not chosen. No person outside AI Adoption has reviewed it yet. Editorial owner: AI Adoption editors. Spot a mistake? Corrections.
Disclosure: AI Adoption uses AI tools, including Claude from Anthropic, in its research and editorial work. No company pays for or has any say in a rating.