In December 2023, we celebrated our one millionth product review. What made some people happy led to a kind of overload for others. That's because a million product reviews mean that, on average, there are far more than 20 reviews for every product featured on Migipedia. And who reads 20 product reviews? Almost nobody, as we know from web analysis evaluations. 50% of the community members that look at reviews on the product detail pages read the first five. 80% of the rest read ten reviews. Almost nobody reads more than 20 reviews. The growing number of product reviews is important for our product managers, but simply no longer manageable for many community members. Product reviews should not only help us optimize products and design our product ranges, but also support our customers in choosing which products to buy. And because this is becoming more and more difficult as the number of reviews increases, I would like to present you with a solution in this update.
Machine-generated summaries of product reviews
Artificial intelligence (AI) has been on everyone's lips for some time now. Migros has also looked into the subject, and is starting to use this technology here and there. To this end, secure, confidential access to the highly trained GPT-4 models was created within the Migros network. As soon as a product has five reviews, our AI springs into action and summarizes the reviews. Care is taken to ensure not only positive but also negative aspects are included in the summaries. As soon as five more reviews are added, the summary is rewritten. Very importantly, the summaries are generated anonymously, i.e. your usernames and other personal data aren't taken into account or passed on.
Summaries can be reported
So does the machine create good summaries of the product reviews? Take a look at some of the product detail pages and see for yourself. We were definitely pleasantly surprised during our tests. Nevertheless, it may happen that the summaries, which are ultimately based on your reviews, somehow don't make sense. I'm sure you'll discover one or two in the coming days and weeks. So we can make improvements here, we have also integrated a "report function". You may already know this function from other places on Migipedia. Reported summaries are checked by us to determine where - and above all why - the artificial intelligence failed. 😉
Best regards,
Philipp
All replies (7)
Good new feature. I would also find it helpful if changes over time in the ratings were taken into account in the summary (or are they already?). Sometimes products are suddenly much better or worse after a change of manufacturer. Or something was criticized years ago that has since been rectified. And finally, I don't really care what the quality of the pumpkins was like five years ago, for example.
Good new feature. I would also find it helpful if changes over time in the ratings were taken into account in the summary (or are they already?). Sometimes products are suddenly much better or worse after a change of manufacturer. Or something was criticized years ago that has since been rectified. And finally, I don't really care what the quality of the pumpkins was like five years ago, for example.
Good morning hardid61 Thank you for your positive feedback and your question. The summaries are always generated when a product detail page is accessed for the first time. If 5 new ratings are added, the rating is summarized again in any case. In this way, we ensure that more recent reviews are also taken into account. In addition, ratings are regenerated at least every two weeks to take account of minor changes. This function is still brand new on Migipedia. It cannot be ruled out that the logic will be adapted or refined in the future. Kind regards Philipp
Good morning hardid61 Thank you for your positive feedback and your question. The summaries are always generated when a product detail page is accessed for the first time. If 5 new ratings are added, the rating is summarized again in any case. In this way, we ensure that more recent reviews are also taken into account. In addition, ratings are regenerated at least every two weeks to take account of minor changes. This function is still brand new on Migipedia. It cannot be ruled out that the logic will be adapted or refined in the future. Kind regards Philipp
Thanks for the information Philipp, but my input was less about the selection of reviews the AI is fed with and more about whether the AI itself can take rating trends over time into account. Suppose 20 out of 100 reviews dislike the packaging (or they find the product too sweet or...). Does the AI write the same summary if it is the first 20 reviews that are bothered by the packaging as it does if it is the last 20, i.e. the most recently written reviews, or if the 20 reviews are spread out over time? That would give an indication of whether the problem has been fixed or is current. But that would be something for the future, when the first version has proven itself.
Thanks for the information Philipp, but my input was less about the selection of reviews the AI is fed with and more about whether the AI itself can take rating trends over time into account. Suppose 20 out of 100 reviews dislike the packaging (or they find the product too sweet or...). Does the AI write the same summary if it is the first 20 reviews that are bothered by the packaging as it does if it is the last 20, i.e. the most recently written reviews, or if the 20 reviews are spread out over time? That would give an indication of whether the problem has been fixed or is current. But that would be something for the future, when the first version has proven itself.
Good morning hardid61 Thank you for your clarification. Let me include your input in our further development of the AI topic. Because you're right, mapping changes in this way would definitely be exciting. Kind regards Philipp
Bravo, good solution. I've looked at a few summary evaluations and it's well done (e.g. M-Budget Orange Juice). And I really appreciate the possibility of communicating about any discrepancies.
Bravo, good solution. I've looked at a few summary evaluations and it's well done (e.g. M-Budget Orange Juice). And I really appreciate the possibility of communicating about any discrepancies.
Thank you very much for your positive comment. Sincerely, Philipp
Hi Philipp,
Congratulations on reaching one million product reviews on Migipedia! That’s an impressive milestone. It’s great to see that you’re addressing the challenge of information overload with AI-generated summaries. This should make it much easier for users to get a quick overview of product feedback without having to sift through numerous reviews.
The approach of updating summaries after every five new reviews ensures that the information remains current and relevant. It’s also reassuring to know that the process is anonymous and that there’s a report function for users to flag any issues with the summaries.
How has the community responded to this new feature so far? Have you received any feedback or suggestions for further improvements?
Best regards,