Karaca tested AI product descriptions on 50 German pages and reported 7% higher conversion

A German shopper choosing between two dinnerware sets does not want a lecture on porcelain. She wants to know whether the set survives a family table, fits the cupboard and feels worth the price. Karaca had descriptions written for that shopper. In January it let an AI system write a second version and made the two paragraphs compete for the sale.

The test was small on purpose, with fifty product pages, one German customer segment, one month and an equal traffic split. Cypien, which generated the new descriptions, reports that conversion rose from 1.28% to 1.37%. That works out to about nine additional purchases per 10,000 visits if the rates held at that volume. The company has not published how many visitors actually took part.

If you advise ecommerce teams, the design is the part to copy. Karaca left the catalogue alone, never asked a model to rearrange the store and did not treat a pile of generated words as progress. It changed one decision surface and watched whether more visitors bought.

Fifty products in one market

Karaca began in Istanbul in 1973 and sells cookware, tableware, small appliances and home products. Its German business runs an online shop and physical stores under Karaca Porzellan Deutschland. The experiment ran on 50 product detail pages in that market from 1 to 31 January 2026.

Here is the situation in marketing terms. Karaca had paid to bring a visitor to a product page, and the page then had to do the persuading. The existing description may not have spoken clearly to the reasons a German home shopper picks one item over another. Rewriting every page by hand would take months, and a store wide rollout would blur the picture, because promotions, stock changes and traffic quality would all be moving at once.

Karaca and Cypien picked a segment they called Home Enthusiasts. Cypien generated new German descriptions around what that group is likely to care about and what pushes it to buy. The published design points to one shared treatment for the segment, with no separate persona for each visitor. That treatment went up against Karaca’s existing copy.

The model wrote while the holdout judged

Cypien says traffic to the chosen pages was divided 50/50 whatever the acquisition source. The control group saw the original description and the treatment group saw the AI version. Prices, product selection and the rest of each page were meant to stay the same.

For an operator, the useful system here goes beyond the language model that produced the copy. It includes the delivery layer that served a variant to half the traffic, kept a control group intact and tied each exposure to add to cart and purchase events. Without that plumbing, Karaca would have owned a faster content workflow and no reliable way to say whether it helped.

Cypien also says the 50 products passed checks meant to screen out stock problems, campaigns and other marketing activity that could distort conversion. The public account skips the detail of those checks, but the intent is sound. If the treated product is discounted while the control product sells out, the model stops being the interesting variable.

Think of it as German ecommerce localisation with a measurable counterfactual. The system did more than translate, since the copy was written around a defined customer segment. It also did less than live one to one personalisation, because every treated visitor saw the same Home Enthusiasts version. That gap is good news for smaller teams. A segment level content test can be reproduced without building a live intent model.

Nine more buyers in every ten thousand visits

Cypien reports conversion of 1.37% for the AI copy against 1.28% for the original descriptions. That is a 7% relative lift. The absolute gap is 0.09 percentage points, which works out to about nine extra conversions per 10,000 exposed visits if the rates held at that volume.

Add to cart rate moved from 3.16% to 3.39%, a relative rise of 7.3%. Bounce rate fell from 23.20% to 17.72%, which Cypien expresses as a 23.6% improvement. Both upstream measures point the same way as conversion, though purchase is the number the business banks.

According to the case page, the treatment received 50.03% of traffic and the control 49.85%. Those shares look balanced, yet they total 99.88%, and the account gives no reason for the missing 0.12 percentage points.

Because both versions ran in the same month, this beats a before and after claim. The control soaks up much of the seasonality, campaign mix and market noise. Even so, the case is far from settled. Cypien gives no visitor count, number of purchases, confidence interval, test power or revenue per visitor. A small sample could produce these exact rates with wide uncertainty, while a large one would make the gap persuasive.

A result published by the company that wrote the copy

Cypien designed the test, generated the treatment and published the result. Cypien’s LinkedIn post thanks Karaca team members, and Turhan Koryak, Karaca’s international ecommerce manager, is linked to the project. The public evidence still comes from Cypien. Karaca has not released the raw data, the statistical output or its own account of the rollout decision.

The page also leaves the assignment mechanics open. It does not say whether a returning visitor always saw the same version, whether exposure was randomised by user or by session, or whether purchases were deduplicated across devices. The choice of products adds one more limit, since results from 50 selected items may not travel to low traffic products, technical appliances or categories where specifications matter more than lifestyle language.

The treatment is also a bundle. It mixes AI generation, German localisation and a Home Enthusiasts framing, so the experiment shows that the finished package beat the existing copy and leaves the source of the gain open. Better German, a different emphasis, clearer structure or a plain fresh editorial pass could each explain it.

The next test should separate translation from persuasion

A cleaner follow up product page A/B test would run three groups. The first keeps Karaca’s existing copy. The second gets a professional localised rewrite that improves clarity without any customer persona. The third gets the Cypien treatment written for Home Enthusiasts. Revenue per visitor would be the main metric, with conversion, add to cart rate and average order value showing where any gain appeared.

Karaca could fix the test window and minimum sample in advance, keep assignment steady for returning shoppers and report confidence intervals by product category and device. Stock availability, discount exposure, page speed and return rate belong beside conversion as guardrails, because better copy that attracts the wrong buyer can lift checkout and then send the product back a week later.

The lever for your own team is to treat AI as a cheap variant factory. Limit the first experiment to a coherent product set, keep half the traffic on the old experience and let paid behaviour decide the rollout. Galeria followed the same discipline when it tested AI search on 1.5 million shoppers before committing to a new cost structure. Karaca’s version is smaller, and the decision rule matches.

Cypien’s numbers suggest that one paragraph changed enough minds to matter, and the missing sample size decides how closely anyone should copy it. The most useful thing the model could produce next is a test report with the denominator attached.

Experiment detail	What the reporting shows
Constraint	Karaca wanted more German product page visitors to buy without changing traffic acquisition.
Intervention	Cypien generated German descriptions for a Home Enthusiasts segment across 50 products.
Comparison	A 30 day 50/50 traffic split between original and AI generated descriptions, whatever the traffic source.
Reported result	Conversion rose from 1.28% to 1.37%, add to cart from 3.16% to 3.39%, and bounce rate fell from 23.20% to 17.72%.
Limits	Published by Cypien, with no visitor count, purchase count, confidence interval, test power, revenue per visitor or assignment details.

Sources and disclosure

• Cypien’s Karaca International case study, covering the 1 to 31 January 2026 test
• Cypien’s practitioner post naming Karaca collaborators
• Karaca’s official company history and German legal notice
• Karaca international ecommerce manager Turhan Koryak on LinkedIn

Cypien produced the AI product descriptions, designed the experiment and published the performance figures.

FAQ

Compare more measured experiments in the Signal Index, browse the Growth Signal archive, or use the benchmark tools in the IC Lab.