How AI is used to generate fake product listings and reviews on marketplaces

Artificial intelligence has long been a business tool: it is used to write product descriptions, process images, and respond to customers. But the same technologies are also being increasingly used by unscrupulous sellers - generative AI allows them to prepare a product listing in a few minutes that is almost indistinguishable from the original and generate dozens of plausible reviews. Let’s examine how this works and why brands should reconsider their approach to intellectual property protection.

Counterfeits have become harder to identify.

Previously, violations could be identified by exact matches of names or photos, as counterfeit listings gave themselves away through low-quality images, description errors, and sloppy presentation.

Today, using generative AI, sellers can create professional content in minutes that is nearly on par with the original. AI rewrites product descriptions and modifies images based on original materials so that the listings do not appear to be direct copies of one another. Meanwhile, the trademark and brand identity elements are altered only slightly, remaining recognizable to consumers; consequently, a simple search for duplicates no longer detects these listings.

Now, identifying violations requires analyzing a multitude of more subtle indicators.

What it looks like in practice

Let’s take a typical scenario that brand protection teams face today. An unscrupulous seller takes photos and descriptions of an original product and runs them through a generative model: the text is rewritten to be close to the original but with different wording and structure; in the photos, the angle, background, and color scheme are slightly altered, and the logo on the packaging is blurred or cropped out. Formally, the listing is not an exact copy - neither the text nor the image matches the original byte-for-byte - and while the customer is essentially looking at a counterfeit, they involuntarily associate it with a well-known brand.

Next, this listing is multiplied: several different versions of the description and photos are generated, published under the names of different sellers, and then "boosted" with reviews - dozens of short but differently phrased comments about product quality, fast delivery, and accuracy to the description. No automatic marketplace filter searching for duplicate text or matching images flags such a listing: every copy is unique at the level of phrasing and pixels, even though the entire group of listings clearly originates from a single source.

To find such a group of listings manually, a brand protection specialist cannot look for an exact match but must instead correlate indirect signs: similar image composition, linguistic patterns characteristic of a specific seller, and recurring patterns in reviews. In practice, this means that a single case like this takes many times longer than it did a few years ago, and the number of such cases is steadily increasing.

Reviews you can't trust

A separate issue is the automatic generation of reviews. Modern language models create large volumes of natural-sounding comments, varying in style and length, which are difficult to distinguish from real ones - and this is not just an assumption, but a result confirmed experimentally.

A 2025 study showed that people are no longer able to distinguish real reviews from AI-generated ones - recognition accuracy averaged only 50.8%, which is practically equivalent to random guessing. Moreover, language models themselves also fail at this task and show results no better, and sometimes worse, than humans. In other words, neither humans nor algorithms can currently reliably separate a real review from a generated one, meaning that automated marketplace filters relying solely on text analysis catch such fakes only slightly more effectively than random selection (see Meng W., Harvey J., Goulding J., Carter C. J., Lukinova E., Smith A., Frobisher P., Forrest M., Nica-Avram G. “Large Language Models as 'Hidden Persuaders': Fake Product Reviews are Indistinguishable to Humans and Machines”).

Such reviews are used to artificially boost trust in a product and promote listings in marketplace search results. As a result, a high product rating is no longer a guarantee of quality, and negative reviews can be a tool for unfair competition. For the consumer, this means that the usual benchmarks for choosing a product are no longer effective.

What this means for rightsholders

The use of AI makes counterfeiting not only more widespread but also more convincing, which means risks for brands are growing:

• there are more violations and they are harder to track manually;

• fake product listings are harder to distinguish from authentic ones even upon close inspection;

• inflated ratings and fake reviews nudge consumers toward purchasing counterfeits, leading to disappointment with low-quality goods that are associated with the brand itself rather than the infringing seller.

Ultimately, both sales and reputation suffer, and the damage to reputation may not manifest immediately.

Why system monitoring is needed

Periodic manual checks of marketplaces can no longer keep up with the scale of the problem: employees would have to analyze hundreds of product listings, compare images, and track changes in near real-time. Even a strong team will inevitably miss some violations.

Consequently, companies are switching to continuous automated monitoring with prompt documentation of violations, filing of complaints, and analysis of repeat instances of illegal trademark use. Solutions like ZIPDetect help not only identify suspicious listings but also establish the entire brand protection process - from monitoring to legal support.

The sooner a company implements such tools, the higher the chance of detecting a violation in time and maintaining customer trust.