Italian product pages for AI assistants: what E-GEO shows and what it cannot tell you

Data and reports · Version 1.0 · Published October 9, 2026 · Updated October 9, 2026 · by Alessio Di Rubbo

When an AI assistant recommends a product, it picks from a few candidates and puts them in order. E-GEO, a study by researchers at MIT and Columbia University, measured which rewrites of the product description move an item up that ranking. The most useful result for sellers runs against intuition: marketing techniques lower the position, while a factual, well-organised description written in the shopper’s own words raises it. The study, however, was carried out in English only. This page summarises what it measured, what it suggests for Italian product pages, and where the data ends and my own inferences begin.

What the study measured

The authors collected 13,747 shopping requests written in natural language, taken from a Reddit forum about durable products, and paired each one with 10 products from Amazon¹. The requests are long, about 59 words on average, the way people write to an assistant: budget, intended use, experience with previous products.

For each request, a language model ranks the 10 products from first to tenth place. The study rewrites the description of one of the ten and measures how many places it moves. To avoid results that hold for a single system only, the ranking was computed with five different models: GPT-5, Claude Sonnet 4.5, Gemini 3 Flash, DeepSeek V3.2 and Llama 4 Maverick. All figures on this page are averages over 2,000 test requests, expressed in places out of 10.

What lowers the position

The authors tested 15 rewriting strategies based on common GEO advice, all carried out by the same model, GPT-4.1. Eleven of the fifteen did worse than a plain request to “optimise the description”². Some pushed the product down sharply:

Only four strategies kept pace with the plain request: adding frequently asked questions, formatting with headings and lists, stressing distinctive advantages, and a format that imitates model output. Even the best of the four gained less than two tenths of a place.

Two results contradict common beliefs. Longer text does not help: the correlation between added length and improvement is practically zero, and GPT-5, one of the two most effective rewriters, shortened descriptions by about 14 words on average. A bulleted list helps, but only a little: about a quarter of a place on average³.

What raises it

The authors then let an algorithm revise the rewriting instructions step by step, based on the results. Starting from fifteen very different strategies, the algorithm arrived at the same pattern every time, which the study calls a “universally effective” strategy⁴:

• factual, concrete language, with no unsupported claims and no references to competitors;
• clear sections: key features, benefits, typical use cases;
• each feature tied to a need that shoppers actually express;
• the keywords and synonyms shoppers use, worked in naturally and without forced repetition;
• answers to likely questions: durability, ease of use, compatibility, value for money.

The study also checked whether these gains amount to manipulation. Fake reviews, extreme superlatives and hidden instructions aimed at the model were all tested: the strongest models, GPT-5 and Claude, did not reward them, even with a single defensive instruction⁵. The authors’ conclusion is that, in this setting, a product rises through better communication, not through tricks.

What the study cannot tell you

The authors state the limits clearly, and two of them concern anyone selling in Italy⁶.

It is English-only. Requests and product pages are in English, with a mostly North American audience. The authors write that their conclusions should not be assumed to transfer to other languages without further evidence. No figure in E-GEO says how an assistant treats an Italian product page.

It measures only the final ranking. The study starts from the point where the product is already among the 10 candidates and measures only how its position changes. It does not measure whether the product gets into the candidate set at all. For a translated product page, that is the delicate step: if Italian shoppers describe the product in different words from the page, the product may never reach the ranking. This is my inference, not a finding of the study, but it follows from the way the authors themselves describe how these systems work.

Two further limits are worth keeping in mind. All products come from Amazon, and the study simulates assistants with the models instead of observing ChatGPT or Gemini in real use. It is a controlled test bed, not a market measurement.

What I take from it for Italian product pages

What follows has not been shown by E-GEO for Italian. It is how I would apply its results, knowing where a translated product page usually departs from the pattern that works.

Use the words Italian shoppers use, not the translation of the German or English name. This is where a correct translation can fail. A German “Akku-Staubsauger” or an English “cordless vacuum” translated literally becomes “aspirapolvere a batteria”, while many Italian shoppers search for “scopa elettrica”. The study rewards the terms and synonyms shoppers use: on a localised page they have to come from Italian, not from the source text.

Remove the advertising tone, even if the source has it. Advertising style lost the most places after the creative short story, which the authors included on purpose as a negative control. If the source page reads like a slogan, a faithful translation carries the same flaw into Italian.

Do not leave foreign words in for flavour. The study tested foreign expressions added deliberately to suggest sophistication, not German or English words left behind in a translation. The analogy is mine: a “Must-have” or “Highlights” left on the Italian page gives no information to the shopper or to the model reading it.

Do not add or lose facts in localisation. The winning pattern keeps the facts of the original page and adds no promises. In localisation the risk runs both ways: a technical detail lost in translation, or a claim added to make the text sound more persuasive.

Build the page the way the pattern does. An opening sentence that says what the product is and what it is for, features as a list with clear labels, use cases, two or three questions with answers. It is also the structure that makes the page easier for a person to read.

Frequently asked questions

Does E-GEO apply to Google?

No. The study measures the ranking that language models produce when ordering products, not placement in Google’s results. Some recommendations overlap with good SEO practice, such as the searcher’s own terms and a clear structure, but the study does not test that.

Should I rewrite all my product pages with an AI model?

The study shows that the result depends heavily on the model and on the instructions: with plain instructions, some models made the pages worse. For Italian there is an extra caution, because a model readily rewrites into Italian that is correct but sounds translated. If you use a model, the instructions should follow the pattern described above and the output should be reviewed by a native speaker.

How much do FAQs on the product page help?

On their own, very little: as a standalone strategy they produced only a minimal gain. They matter as part of the overall pattern, where they answer the questions shoppers actually ask about durability, compatibility and use.

Is there a similar study in Italian?

Not that I know of. E-GEO publishes its data and code, so an Italian experiment is technically possible. Until one exists, the recommendations for Italian remain reasoned inferences, not measured results.

Sources

• Bagga, P. S., Farias, V. F., Korkotashvili, T., Peng, T., Wu, Y. (2026). E-GEO: A Testbed for Generative Engine Optimization in E-Commerce. Preprint arXiv:2511.20867, version 2 of 14 July 2026.

Notes

1. Bagga et al. (2026), section 4 and table 2.
2. Bagga et al. (2026), section 5.3 and table 5. Averages across the five models, with GPT-4.1 as the rewriter.
3. Bagga et al. (2026), section 5.2.
4. Bagga et al. (2026), section 5.5.
5. Bagga et al. (2026), section 6.1.
6. Bagga et al. (2026), sections 3.2, 5.1 and 7.1.

Further reading

• GEO glossary for Italian content
• Do AI systems cite different sources in Italian and German?
• Italian SEO and GEO for brands and agencies