AI product descriptions: how to prevent invented product features
Check AI-generated product descriptions against sources, gaps, variants and omitted conditions across sales channels.
Article language: English

In this article
The bag had two pockets and an adjustable strap. After its description was generated, it also had water resistance, responsibly sourced material and the ability to accompany its owner on every adventure. Adventures may be literary licence. Water resistance and material provenance require evidence that the product sheet did not provide. The model has just expanded the offer without asking whether the company intends to deliver on it.
When generating catalogue descriptions, check every product claim, not just the language. Reordering sentences and changing the tone may remain editorial work. Adding a material, resistance, dimension or item included in the set creates new information. The boundary can be thin, especially when a promise hides in an ordinary adjective.
A sound process lets you trace a public sentence back to an approved source. If the source is missing, the result should preserve that gap or be sent for clarification. Filling it elegantly does the catalogue no favours: a customer buys a particular item, not an average bag like every other.
Establish the source of features before the model starts writing
Consider a fictional bag for exhibition materials. Its sheet confirms two pockets, an adjustable strap and the dimensions of the main compartment. It says nothing about water resistance or a coating. A photograph shows colour and shape but does not establish material specifications. If the image was also edited with AI, check the limits of product photo editing (in Polish) before using it as a basis for the description. This input supports only a limited set of claims.
The team should identify the authoritative source: an approved sheet, a product record or another controlled document. If two documents conflict, do not ask the model to decide which is "more reliable". Assign the discrepancy to the person responsible for the product and retain it in the process.
Distinguish a confirmed value, missing data and a feature that is absent. These are different states. An empty "water resistance" field means neither yes nor no. A claim about resistance needs an appropriate source; otherwise omit it or send a question for completion.
Define the variant boundary too. A larger bag may have a different number of pockets, and a set with a cover may contain different items from the basic version. The model must not borrow a feature from a neighbouring item merely because their names begin the same way. The input needs to identify the particular item and its scope.
In its current explanation of Sidekick content generation, Shopify warns that a suggestion may contain benefits not supplied or facts about similar products. It recommends careful review before publication. This is a specific risk described by the tool provider, not proof that every generated text will be wrong.
Design your own trial so the model has an opportunity to make that mistake. Include an incomplete sheet and a similar product with a richer description. Check whether the result preserves the distinction or automatically makes the offers alike. An overly perfect data set may conceal behaviour editors encounter every day.
In the product catalogue content model (in Polish), record features as identifiable fields and assign responsibility for updates. A free-form paragraph can still be input, but it is harder to distinguish a missing parameter from an omission by its author. More structured information makes its checks easier to design.
When sources conflict, retain both entries and their versions. If the sheet says two pockets and an older description says three, selecting the newest file date may not be enough: the file may merely have been exported again. The product owner should confirm the current feature. Until then, the process may prepare the rest of the description but should not turn the conflict into an apparently certain number.
Do not put all the manufacturer's knowledge into one long prompt without identifying the current item. The material may include different models, old versions and general category descriptions. The process needs to provide the right context before expecting fidelity. "Do not invent anything" does not by itself repair conflicting input.
Assess sentences by the information they add
Start with a simple distinction: a feature, a description of use derived from it, and a claim requiring additional evidence. "The bag has two pockets" refers to the sheet. "Two pockets let you separate small items" explains a use that still needs assessment against the construction. "Keeps documents safe in the rain" adds a different promise.
Do not treat every benefit statement as an innocent metaphor. "Easy to clean" can shape expectations about care. "Universal" can suggest compatibility beyond the confirmed dimensions. "Durable" sounds ordinary, but the company needs to establish what it intends to convey and on what basis.
During review, mark claims and their sources. A public footnote for every catalogue word is unnecessary. What is needed is an internal way to check which record or document supplies the information. A sentence without a basis can be removed, narrowed or held for clarification.
Check negatives too. The sheet may state that a cover is not included. A generated "bag with a practical cover" is not better style; it changes the offer. Likewise, "adjustable" should not appear on a version with a fixed-length strap.
Add traps of meaning to the test set: a compartment dimension rather than the overall dimension, the number of internal pockets rather than all pockets, and an accessory ordered separately. Each requires the relationship to be preserved, not just the number copied. Check what each value describes: the compartment measurement must still refer to the compartment, and the internal pocket count to that same part of the bag.
Also check translations of abbreviations and material names if the process covers several languages. An unknown abbreviation should not be expanded into the most likely name. Establish approved equivalents or send the ambiguity to a person. One impressive-looking expansion may then be copied into several channels.
The audit of AI-generated sales proposals (in Polish) illustrates similar limits. There, the agreed scope of a proposal must be preserved. In a catalogue, checks concern public features of many items, often created in bulk. In both cases, fluent language must not change what the company actually offers.

Test the result with no data as well. Given only a name and a photograph, a tool may create a general mood but has no basis for many specifications. The team should know whether such a draft goes for completion or whether generation should not begin at all. Filling an empty field automatically cannot be the only success criterion.
Publishing to different channels involves more than one paragraph
A website description, short product-card text and data sent to an external catalogue may have different format requirements. Their meaning should remain consistent. A shorter version must not remove a condition that changes the offer, such as a note that the pictured accessory is sold separately.
The Google Merchant Center product data specification identifies the structured_description attribute for descriptions created using generative AI. This is a data requirement for that channel. Correctly identifying the text's origin does not establish that its features are true or match the particular product variant.
Check both the content and field mapping. A result for the larger bag must not reach the smaller variant because of an identifier error. A correct paragraph assigned to another product is still incorrect information for the buyer. In the trial, trace the input record to the item actually displayed.
When updating data, use document or product-sheet version comparison (in Polish) to locate changed features. If a cover is removed from the set, check every text that mentions it. Generating a new main description does not automatically remove the old sentence from a short card or message.
The same content may later become a source for a voice AI assistant (in Polish). The invented water resistance may then be spoken in an even more confident voice. Check the feature at its source before further channels start repeating it in agreement.
In evidence for promises in an offer (in Polish), the starting point is the relationship between a claim and its basis. Prepare a similar way to raise doubts in the catalogue. An editor should be able to identify a sentence and request confirmation without searching an entire supplier correspondence alone.
At Aeneia Audit, we can begin by reviewing one content preparation process. Bring approved data, the generation instruction and several public outputs with their identifiers.
Aeneia Audit
Let us examine which sentences follow from product data and which the model added during editing. We will look at incomplete sheets, similar variants and publication of the result in an actual channel.
Assess AI-generated catalogue descriptions
Keep a test that does not reward fluency alone
For a control set, list confirmed features, missing data and content that must not appear. For each difference, also record its consequence for the choice: a different variant, different contents or an unconfirmed use. Editors can then start with sentences that change the meaning of the offer rather than correct every text alphabetically. Assess facts and language separately: smoother style must not compensate for an invented parameter in a single aggregate score. The catalogue owner needs to see both results.
Include the same data with a different stylistic instruction. Ask for a calm version and a more energetic one while keeping the same factual boundaries. Compare claims, not just wording. If a change of tone adds resistance or a new use, the process does not separate editing from expanding the information. Keep the case for checking future brand instructions.
Check omissions too. A description may add nothing false yet remove a restriction without which customers understand the product differently. For our bag, leaving out that the cover is sold separately may matter more than losing an attractive opening. Establish which features and conditions must survive editing.
After improving the instruction, repeat the same cases and add neighbouring ones. If adding a rule about missing water-resistance data makes the model automatically deny water resistance for every product, a new error has appeared. The rule must refer to the source for the specific item, not replace one unsupported claim with another.
When publishing a larger batch, prepare a way to roll back a specific description and find similar errors. If a shared instruction produced an unsupported feature, check other results from that run rather than correcting only the bag you noticed. Version records help establish the scope of review and give the team a way to respond even when a problem surfaces after catalogue content is released.
Record who approves a change and what happens to manually corrected text during regeneration. A later run must not quietly restore a removed promise. For regular work, keep versions of the instruction, data and output that allow the text's provenance to be explained.
In AI quality regression tests (in Polish), retain cases involving missing data, similar variants and exceptions. A new model may write more pleasantly while filling gaps differently. Review should use the same conditions of the offer on which the company already relies.
A good product description helps someone choose an item based on its real features. AI may speed up editing, but the process must preserve the boundary of what is known. A bag does not become rain-resistant because a sentence sounds as if it is. The team should notice the difference before a customer stands under dark clouds carrying paper materials.












