Customer Reviews in Chat: Ratings From the Catalogue, Not Invented Voices

How a shop assistant states ratings only from the catalogue, answers the genuineness question from stored text, and what Section 5b UWG requires.

In plain language

Before buying, many people still ask about reviews. The assistant answers from the shop's catalogue: with the average star rating and, where it exists, the number of votes. If that data is missing, it gives none. It is not meant to make up voices or project numbers. It does not pull in review sites run by others. It only says reviews come from real buyers if the shop has stored a text about that. The shop itself stays responsible for that check.

24 October 2026 13 min read By XICBOT

Key points

The rating is often the last question before a purchase. An assistant answers it from the catalogue: the star average and, where the catalogue keeps it, the number of votes. If the field is missing, the line on the card is missing.
No invented voices, no “most customers praise”, no extrapolation and no comparison with other review sites. Review texts are not in the catalogue; if they reach the knowledge base via crawled product pages, the assistant generally reproduces them with a citation.
For anyone who makes consumer reviews accessible, information on whether and how they ensure that the reviews come from actual buyers or users is deemed material under Section 5b(3) UWG; withholding it can be unfair under Section 5a(1) UWG. The assistant answers the genuineness question only from the text the shop has stored for it.
Claiming that reviews come from actual buyers without reasonable and proportionate checks is listed in the annex to the UWG under no. 23b, in the list of practices that are always inadmissible towards consumers. The assistant repeats such a statement only if the shop has stored a text describing its check; whether that check is sufficient is the shop’s responsibility.
In the 2021 European sweep, before the current rules on reviews applied, authorities could not confirm for 144 out of 223 (European Commission) of the websites checked that enough was being done to ensure reviews are genuine.

The last question before buying

Anyone who asks about an item in the chat has often almost made up their mind. Price, delivery time and size are settled, the product card is in the conversation, and the cart button is one movement away. This is typically the moment the question about the rating comes up. It is less a request for new information than a request for reassurance: have others had good experiences with the item? When buying a gift it weighs more, because the buyer will not use the product herself and has to rely on other people’s judgement. How an assistant advises in that situation is described in the article on gift advice in chat.

For the assistant this question is trickier than it looks. A language model can effortlessly produce a sentence that sounds like customer feedback, and that very ability is the problem here. An invented voice in a sales situation is not a harmless embellishment but a statement about other consumers that nobody made. The well-meant summary is just as difficult: “The reviews are mostly positive” sounds neutral, yet it is an evaluation the assistant can only provide if it has the individual reviews. As a rule it does not. What it has are one or two numbers in the catalogue, and it should not say more than that.

The rating is a field, not a phrase

What the assistant says about ratings as a rule comes from a catalogue field or from a text the shop has stored. Anything that is in neither is not intended for its answer, even when the question sounds as if it wants quick reassurance.

What the catalogue holds – and what it does not

The catalogue of a shop assistant is a store of its own next to the knowledge base, holding the fields product cards are built from: title, image, price, crossed-out price, features, rating, delivery time and availability. For the rating that means at most two values: the star average and, where the catalogue keeps it, the number of votes cast. The texts of the reviews themselves are not in it. In XICBOT, the product cards show the rating as a line of its own on the card, next to price, delivery time and the cart button. The table shows what becomes of each field and what happens when it is empty.

Catalogue field	What the card shows	If the field is missing
Star average	A line of the form “average out of 5”, written in the language of the conversation	The rating line is dropped; no estimate is filled in
Number of votes	In brackets after the average, if it is greater than zero	The average stands alone, without saying how many votes it is based on
Review texts	Nothing, they are not a catalogue field	No substitute in the form of a summary
Ratings from other sites	Nothing, the catalogue only holds the values the shop supplies	No comparison and no reference to other averages

Which of the two numbers arrives depends on the route by which the range gets into the catalogue. For a shop based on Shopware Community Edition, the sync takes over the star average; the number of votes appears on the card only if the catalogue keeps it, for instance because it is supplied together with the range. The average goes out as a bare number and only becomes a sentence in the chat window, because whether it is written with a comma or a point depends on the language of the conversation. Which catalogue fields are worth having for a shop assistant in the first place is described in the article on product data for the shop assistant.

What the assistant makes of it – and what not

The rule is simple, and that is exactly why it holds: the assistant states the rating as it stands in the catalogue. The card builds the line from the catalogue field, not from the answer text, so the number sits in a place that does not depend on any wording. XICBOT’s shop assistant reads the catalogue of a shop based on Shopware Community Edition for this in the Pro package, and its answers are bound to the sources available to it and as a rule carry a source reference. What that looks like in a conversation is described in the article on citing sources in chat. For questions about ratings, this leads to five things an assistant refrains from.

No invented voices: no “One customer writes …” and no quote that appears in no source.
No summary without a basis: no “most customers praise the workmanship” as long as the assistant does not have the individual reviews.
No extrapolation and no rounding up: an average does not become “almost full marks”, and missing votes do not become “only good ratings so far”.
No comparison with other review sites: the assistant does not quote averages from platforms that are neither in the catalogue nor in the knowledge base.
No comparison without both values: “better rated than” requires an average in the catalogue for both items. The comparison card then puts both lines side by side.

One qualification belongs here. Review texts are not in the catalogue, but they can reach the material by another route: if the knowledge base reads the shop’s product pages and the reviews sit in the page text there, they become sections like any other text. If they fit a question, the assistant as a rule reproduces them with a citation. Whether that is wanted is up to the shop. If the reviews live at addresses of their own, your own exclusion patterns on the knowledge source keep them out; XICBOT allows up to 50 patterns per source. If they sit in the page text of the product page itself, a pattern can only act on the whole page. If they are loaded later via JavaScript, the crawler does not read them, because JavaScript is not executed while reading. How exclusions are planned and checked is described in the article on exclusions in the knowledge base.

Question in the chat	Answer that invents	Answer from the catalogue
How is this rated?	Customers love it, almost everyone gives full marks.	The rating is on the card: the average and, where kept, the number of votes. I do not have individual reviews.
What do others say about the fit?	Many write that it runs small.	I have no review texts on that. I can give you the measurements from the item description.
Is it rated better elsewhere?	On other portals it even has more stars.	I do not have ratings from other sites. The card shows the rating from this shop.
Which of the two is rated better?	The first one, customers like it more.	Both averages sit side by side on the comparison card. If one of them has no rating, I say so.

Are the reviews genuine? Section 5b(3) UWG

The second question is harder: “Are the reviews genuine?” It is aimed not at the item but at the shop, and unfair competition law has a provision of its own for it. Section 5b of the German Act against Unfair Competition (UWG) lists which information is deemed material; withholding it can be unfair under Section 5a(1) UWG. For price information, the article on prices in chat covers this; for reviews, paragraph 3 applies:

Where a trader makes accessible reviews that consumers have made of goods or services, information on whether and how the trader ensures that the published reviews originate from consumers who have actually used or purchased the goods or services shall be deemed material.

Section 5b para. 3 UWG (Act against Unfair Competition), translated

The provision does not require any particular check. It requires the shop to say whether and how it checks, even if the honest answer is that it does not. This information belongs where the reviews are, for instance on the product page or in an explanation linked from there. The assistant answers the genuineness question only from a text the shop has stored for it. The most reliable form is a question-answer pair in the knowledge base: its wording goes out unchanged, without being rephrased by the model. A maintained text snippet with the same explanation is given priority in the context. If neither exists, the assistant generally says that it has nothing on this; guessing is not intended.

Whether an assistant declines instead of guessing when there is no basis can be measured. In our own measurement series 8 of 8 questions without a basis were declined instead of guessed; eight questions deliberately placed outside the material were measured. For the genuineness question this is the decisive case: a shop that has stored nothing on how it checks reviews as a rule gets a refusal instead of a reassuring invention. The measurement series applies to our own assistant on its knowledge base and to the questions asked there; whether an assistant answers the genuineness question the same way in your own shop only a test question there will show. Why that refusal is a mark of quality is described in the article declining instead of guessing.

Claims with consequences: no. 23b and 23c

Besides the duty to inform, the UWG has an annex listing commercial practices that, according to the wording of the Act, are always inadmissible towards consumers. Two numbers concern reviews directly. The first is aimed at a claim that is easily made in a sales conversation:

the claim that reviews of goods or services originate from consumers who have actually purchased or used those goods or services, without reasonable and proportionate steps having been taken to check whether the reviews actually originate from such consumers

Annex to Section 3 para. 3 UWG, no. 23b, translated

For the assistant this means: phrases such as “verified buyers only” or “all reviews come from real customers” do not belong in an answer unless the shop has stored what measures it takes for this. If the shop has such a check and describes it in a stored text, the assistant as a rule reproduces that text with a citation; additions are not intended. Whether the measures described are reasonable and proportionate is not for the assistant to judge; that is a question for the shop and, if in doubt, for legal advice. The second number concerns fake reviews:

the submission or commissioning of fake consumer reviews or endorsements, as well as the misrepresentation of consumer reviews or endorsements in social media, for the purpose of promoting sales

Annex to Section 3 para. 3 UWG, no. 23c, translated

The German wording can be read in two ways. The submission and commissioning of fake reviews stand without reference to any particular channel; whether “in social media” also refers to the misrepresentation of reviews is left open by the German text. The provision implements an EU directive, and its English version reads “misrepresenting consumer reviews or social endorsements”. According to this, “social” refers only to endorsements; the misrepresentation of consumer reviews is not limited to social media. Recital 49 of Directive (EU) 2019/2161 gives publishing only positive reviews and deleting the negative ones as an example. For an assistant this means: if it reproduces review passages that reached the knowledge base via the product pages, the selection alone can create a one-sided picture. For the design, one conclusion therefore holds: an assistant that invents no voices, reinterprets no reviews and only reproduces what is in the catalogue or in the stored text makes no statement of its own about reviews; which review passages it reproduces still belongs in the checks. How the provision applies in an individual case is a matter for legal advice.

What the 2021 EU sweep found

How widespread the problem was before the current rules can be seen from a coordinated sweep by the European consumer protection authorities. In 2021, authorities from 26 Member States as well as Iceland and Norway examined major websites for misleading consumer reviews: online shops, marketplaces, booking sites, search engines and comparison services. For 144 out of 223 (European Commission) of the websites checked, they could not confirm that the traders were doing enough to ensure that reviews are genuine. 104 out of 223 (European Commission) websites did not inform consumers how reviews are collected and processed. The Commission sums up the first finding as follows:

In 144 out of the 223 websites checked, authorities could not confirm that these traders were doing enough to ensure that reviews are authentic, i. e. that they were posted by consumers that actually used the product or service that they reviewed.

European Commission, 2021 sweep on online consumer reviews (original wording)

The figures describe major websites, not a particular shop, and they are a finding of the authorities, not a verdict on individual traders. The sweep also took place before 28 May 2022, the date from which the requirements of Directive (EU) 2019/2161 on consumer reviews had to be applied; Section 5b(3) and nos. 23b and 23c go back to them. For running a shop the figures say one thing above all: information on how reviews are collected and checked was often missing in the sweep, and it is exactly what visitors ask about in the chat. Anyone who maintains it on the page and stores the same text for the assistant answers the question the same way in both places and only has one version to keep up to date.

Setting up: four steps before going live

Check the rating fields in the catalogue: do the average and, if wanted, the number of votes appear for the items shown most often in the chat? Items without a rating show no rating line; that is correct, but it should be known.
Store the text on how reviews are checked: the same explanation that is on the page, as a question-answer pair using the words visitors use, such as “Are the reviews genuine?” and “Can anyone leave a review?”.
Control the review areas: decide whether review texts from the product pages belong in the knowledge base, and keep them out with your own exclusion patterns on the knowledge source if needed. If they stay in, use test questions to check whether the assistant reproduces a one-sided selection, such as only favourable passages.
Write test questions: about the rating of an item, about genuineness, about an item without a rating and about other review sites, each with an expected answer.

Test question	Expected	Not expected
How is this item rated?	Reference to the rating line on the card, values as in the catalogue	An assessment or summary of its own
Are your reviews genuine?	The stored text on how reviews are checked, with a citation	“Yes, all reviews are verified” without any text stored for it
And how is this one rated? (item without a rating)	The statement that there is no rating for this item	An average taken from similar items
How is it rated elsewhere?	The note that only this shop’s ratings are available	Values or impressions from other review sites

Measuring: how often reviews come up

Whether questions about ratings matter in your own shop shows up in the conversations. Conversation analytics groups the visitor questions of a period into topics, counts them and sets each topic against the figure for the preceding period of the same length. If topics with lead terms such as “rating”, “stars” or “genuine” appear, that indicates visitors look for this information where they happen to be: in the chat. The analysis knows no word meaning, only matching word parts; questions without shared word parts therefore appear as separate topics and should be added up when reading. Next to this are the knowledge gaps. If an answer ends without a source reference, a row of its own is generally created, and a cluster around the genuineness question shows that the text for it is missing. This is not to be confused with visitors rating the chat answers themselves, which the article letting customers rate answers describes.

A second figure helps to put this in context. In our own measurement series 42 of 48 answers carried a source reference; questions whose answer is in the knowledge base were counted, and six carried none. Applied to questions about ratings, this means: an answer to the genuineness question that ends without a citation is not a success but a task for the stored text.

Topics around ratings

Topics with lead terms such as “rating”, “stars” or “experiences”, added up when reading. If their total grows, that suggests maintaining the rating fields in the catalogue completely.
Topics around genuineness

Topics with lead terms such as “genuine”, “fake” or “verified”. Each of these questions should meet the stored text on how reviews are checked.
Gaps

Which questions about ratings ended without a source reference? They lead to new question-answer pairs, not to new wording in the system instruction.

Limits: what the assistant does not take on

It does not check whether a review is genuine. That lies with the shop, which has its order data and procedures for it.
It does not collect reviews or ask for them, neither in the conversation nor after the purchase.
It does not query other review sites and does not compare against their values.
It is not meant to judge whether the way reviews are presented in the shop is legally sufficient.
It does not promise any effect on the conversion rate. Whether ratings in the chat make buying easier depends on the range and on the ratings themselves.

Frequent questions

Where does the assistant get an item’s rating from?

From the catalogue. It holds the star average and, where the catalogue keeps it, the number of votes. The product card turns these into a line of its own in the language of the conversation. If the field is missing, the line is dropped and no estimate takes its place.

Can the assistant summarise what customers write in their reviews?

Only if it has the review texts as a source, and they are not in the catalogue. If they reach the knowledge base via crawled product pages, the assistant generally reproduces matching passages with a citation. A summary such as “most customers praise” without reviews to go on is not intended.

What does the assistant answer when asked whether the reviews are genuine?

It reproduces the text the shop has stored for this, typically as a question-answer pair whose wording goes out unchanged. If nothing is stored, it generally says that it has no information on this instead of phrasing a reassuring answer.

Does the answer in the chat replace the information under Section 5b(3) UWG?

As a rule, no. The provision is tied to making the reviews accessible, and the information belongs where the reviews are. The chat can reproduce it in addition. Whether your own implementation is sufficient is a matter for legal advice in the individual case.

May the assistant say that the reviews come from verified buyers?

Only if the shop has stored a text describing its check; whether that check is sufficient is the shop’s responsibility. The claim that reviews come from actual buyers or users without reasonable and proportionate steps having been taken to check this is listed in the annex to the UWG under no. 23b. Without a stored text, the statement does not belong in the answer.

Can the assistant include ratings from other sites?

Not as a function of its own. It answers from the shop’s catalogue and knowledge base and does not query other review platforms. If someone asks about them, it generally says that it only has this shop’s ratings.

Ask it yourself, Assistant of this website

Automated assistant

Settings