Environmental Claims in Chat: What the Rules Require Since September 2026

Since 27 September 2026 the German Act against Unfair Competition (UWG) has its own term for sweeping environmental advertising: the generic environmental claim. It covers words such as environmentally friendly, green or climate friendly whose precise basis is not given on the same medium. A chat assistant is a special case for this rule. It does not write advertising copy, it summarises sources, and that is exactly where “shipping box made of recycled paper” quickly turns into “sustainably packaged”. This article explains the new provisions with verbatim quotes and shows how to set up an assistant’s sources, rules and acceptance tests so that precise statements do not become generic ones in conversation.

Enlarge graphic

Labels in the illustration

• FROM THE SOURCE TO THE REPLY
• Product text in catalogue
• Shipping box made of
• recycled paper
• Source [1]
• Reply A: The shipping box is made of recycled paper [1].
• Specific: object and property in the same reply
• Reply B: Our packaging is sustainable.
• Generic: one feature stands for the whole (Annex no. 4a, 4b)
• WHAT YOU SET UP AND WHERE THE CHECKING ENDS
• Clean the sources
• Keep old campaign pages
• out with your own
• exclusion patterns
• Fix the wording
• Keep packaging, shipping
• and material as
• question-answer pairs
• Limit on the question
• Excluded topics trigger
• a handover; only the
• question is checked
• Output check
• Percent, amounts, promises
• checked against source;
• green words not included
• What remains open
• Whether a specific wording is permissible is decided case by case, not by the assistant.

How a precise statement becomes a generic environmental claim

What applies since 27 September 2026

The Third Act amending the Act against Unfair Competition of 12 February 2026 was promulgated in the Federal Law Gazette 2026 I No. 43 and, according to its footnote, serves to transpose Directive (EU) 2024/825. Its Article 2 sets the date: apart from one individual amendment that applied earlier, the act entered into force on 27 September 2026 (Article 2 para. 1 of the amending act). Since then Section 2 UWG contains new definitions, among them environmental claim, generic environmental claim, sustainability label and certification scheme, and the annex to Section 3(3) UWG lists new items on environmental claims.

The term rests on a second one. Under Section 2(2) no. 5 UWG an environmental claim is any statement or representation in the context of a commercial practice, including through text, pictures, graphics or symbols, that is not mandatory under law and states or implies that a product, a product category, a brand or a trader benefits the environment, does it no harm or harms it less than others, or that its environmental impact has improved over time. Such a claim becomes generic when it is not included on a sustainability label and its specification is missing, that is, when the same medium does not state clearly and prominently what exactly it refers to. New items in the annex to Section 3(3) UWG build on these terms; the annex lists practices that are always unfair towards consumers under the statute. Item 4a prohibits generic environmental claims for which no recognised excellent environmental performance can be demonstrated. Section 2(2) no. 2 UWG defines that term: environmental performance in line with Regulation (EC) No 66/2010, with eco-labelling schemes under EN ISO 14024 type I officially recognised in the EU, or with environmental excellence under other applicable Union law. So it does not mean just any evidence, but performance measured against these standards. An addition to Section 5 UWG covers future environmental performance.

Why a chat reply can be an environmental claim

The definition does not refer to advertising space but to the form of the statement: written or oral, including through audiovisual media. A chat reply is written text. It is also made in the context of a commercial practice, because the UWG defines that as any conduct in favour of one’s own business before, during or after a transaction that is directly and objectively connected with promoting sales (Section 2(1) no. 2 UWG). A shop assistant answering a question about packaging next to the shopping cart is an obvious example. The same reasoning applies to price statements, as the article on what a chat assistant may say about prices sets out.

What matters for chat is the second part of the definition. An environmental claim that is not included on a sustainability label is only not generic if its specification is provided in clear and prominent terms on the same medium. The directive behind the act gives examples in recital 9: the same advertising spot, the product’s packaging, the online selling interface. For a chat window the cautious reading is also the practical one: the specification sits in the same reply, not only on a page the reply links to. Whether a link can be sufficient in an individual case is not assessed in this article.

Five patterns that arise in conversation

Recital 9 of the directive lists typical generic environmental claims, among them environmentally friendly, eco-friendly, green, nature’s friend, ecological, environmentally correct and climate friendly. It adds a pair of examples that shows the difference: the claim “climate-friendly packaging” would be generic, while stating that all the energy used to produce the packaging comes from renewable sources is specific. In conversation, generic claims rarely arise from intent; they arise from condensing. Five patterns appear; the example sentences are neutral and do not come from any particular shop.

• Generalising (Annex no. 4a): the source says “shipping box made of recycled paper”, the chat says “our packaging is environmentally friendly”. A verifiable statement turns into a verdict on environmental impact.
• One feature for the whole (Annex no. 4b): the source says the protective sleeve is made of recycled plastic, the chat says “the product is made of recycled material”. Recital 11 of the directive describes exactly this case: recycled material only in the packaging, presented as a property of the whole product.
• Offsetting as neutrality (Annex no. 4c): the source says the retailer supports a climate project for each shipment, the chat says “shipping is climate neutral”. The neutrality then rests on offsetting.
• Intention as commitment (Section 5(3) no. 4 UWG): the source says the retailer intends to use less plastic, the chat says “from next year we ship plastic-free”. Without an implementation plan with measurable and time-bound targets, such a claim about future environmental performance lacks a basis.
• Emblem as label (Annex no. 2a): the product text shows a self-designed emblem with a green leaf, the chat presents it as an eco label. Whether mentioning it in chat already counts as displaying it depends on the individual case; but the word label does not belong in a reply whose source does not carry it.

All five patterns have one thing in common: the source is right and the reply is not. When condensing, a language model reaches for words that often appear in such contexts, and environmentally friendly is shorter than a material statement with its reference. That sets the order of the next sections: first the sources, then the replies, then the limits of the technology that checks in between.

Source binding helps as far as the source goes

XICBOT is instructed to answer from the excerpts retrieved for the question, not from general model knowledge. Sourced replies carry the number of the excerpt used, and the source location is shown below the reply. If the material is not sufficient, the assistant is meant to say that nothing is available on the point instead of filling the gap. For environmental claims that is half the work: a property that is in no source has no excerpt in the reply to rest on.

How well this binding holds can be counted. In our own measurement series of 11 September 2026, measured on a single assistant on our own knowledge base, 42 of 48 answers carried a source reference; the count covered questions whose answer is in the knowledge base, and six carried no source reference. In the same series 8 of 8 questions without a basis were declined instead of guessed. No figure from the live operation of a customer installation is available. A reply without a source location is also recorded as a knowledge gap. What the reference looks like in the reply and why it matters to visitors is described in citing sources in chat.

The flip side follows from the same mechanism. Source binding works against making things up, not against repeating them. If “climate-neutral shipping” still sits on last year’s campaign page or in a product text that the catalogue supplies from Shopware Community Edition, the assistant repeats the claim, neatly with its source. The shop assistant answers product questions from the catalogue; whatever stands there counts as a source. Responsibility for an environmental claim does not shift as a result; it stays with whoever maintains the source.

• Remove old campaign pages with climate promises from the content, or keep them out with your own exclusion patterns; your exclusion patterns take precedence over robots.txt. How this works in detail is explained in knowledge base exclusions.
• Review product texts in the catalogue for generic green words before the next sync carries them into the assistant. Which fields the catalogue supplies for this is described in product data for the shop assistant.
• Check manufacturer data sheets and documents stored in the knowledge base. Environmental statements there often lack the reference a reply in the shop needs.
• For every environmental statement that is to stay, add the reference: which part, which property, as of which date.
• Word pages about supported climate projects so that they report on the support and do not derive any neutrality of the product from it.

Specific instead of generic: the maintained pair

For the three topics customers typically ask about in retail, namely packaging, shipping and material, a maintained question-answer pair is the right tool. Such a pair goes out word for word, without being split up and without a model call. The specification therefore sits where the cautious reading puts it: in the same reply, in a wording someone has read and approved beforehand. The pairs are maintained directly in the knowledge base, with several phrasing variants per pair for the different ways of asking the same question.

Topic boundary and excluded topics

Not every environmental question has to be answered by the assistant itself. Questions about an item’s carbon footprint, about offsetting or about an eco label touch exactly the claims the law draws most tightly. For such cases operator rules can be set per assistant. The topic boundary describes in your own words what the assistant talks about; it should not answer questions outside it, but that remains an instruction to the model. In addition there is a list of topics you exclude yourself, with up to eight entries: if an entry appears verbatim in the question, the conversation goes to a person. What is compared is the wording, regardless of upper and lower case, not the meaning. Entries such as climate neutral, carbon footprint, offsetting and eco label cover typical phrasings; variants such as offset or CO2-neutral need an entry of their own. Which limits are fixed and which you set yourself is shown in the overview of industries.

A rule on tone, for instance to avoid evaluative green adjectives and to name properties only with their reference, makes sense. It remains a request to the model, though, not a check. How such rules are phrased and where they reach their limits is described in system instruction: boundaries for the AI assistant. The behaviour only becomes robust through what is decided without the model: which sources the assistant reads and which replies go out word for word.

What the output check does and does not do

After the model, XICBOT runs a second check. Hard statements in the draft are held against what counts as evidence: the excerpts, the tool results, the visitor’s questions and settings of the assistant such as role and greeting. The check covers prices, percentages, deadlines, amounts with a unit, contact details and promise words. Anything not found anywhere is replaced by a placeholder, and the correction reaches the visitor as a separate event; if there was no excerpt material at all, this comparison does not take place. For environmental statements with a figure this is valuable, for instance a recycled share in percent or the weight of packaging that the source does not support. If the visitor states the figure in the question, however, it counts as covered and stays in the reply.

This results in a clear division of labour. The output check is a net for figures and promises, not a filter for adjectives. It does not detect environmental claims, it does not filter them out and it does not make any reply unobjectionable under competition law. What helps against a generic green word in chat is in the sections before: sources without such words, maintained pairs for the frequent questions and excluded topics for the sensitive ones.

Reviewing what was asked

Which environmental questions your customers actually ask is shown, from the Pro package, by the topic analysis in conversation analytics. Questions are grouped into topics by their wording and ranked by frequency, with a comparison to the preceding period of equal length. If a topic with lead terms such as packaging, plastic or shipping appears, that is a reason to review the replies in it. Next to it sits the list of knowledge gaps, the questions to which the assistant had no supported answer. A question-answer pair can be created directly from a gap.

• Review the topics every month for terms such as packaging, shipping, climate, plastic or recycling, and compare the replies in them with the sources.
• Close gaps on environmental questions not with freely written text but with a pair that names the object and the property.
• If handovers on environmental questions increase, the team needs an agreed answer that then flows back into the assistant as a pair.
• After every campaign with an environmental angle, check whether the campaign page is still in the content and still being read.

Before the holiday season: asking test questions

The holiday season brings questions that are less frequent during the rest of the year: about sustainable gifts, about gift wrapping, about shipping several parcels. That is exactly where the pressure to condense arises, because customers ask for a verdict, not for a material statement. How to prepare the chat as a whole for seasonal peaks is described in peak season: absorbing support spikes in chat. For environmental claims a small set of test questions is enough, asked before the season with every reply read.

• Which sustainable gifts do you have? Expected: items with concrete features from the catalogue, no classification as sustainable without a reference.
• Is the gift wrapping environmentally friendly? Expected: material and components from the maintained pair, no verdict.
• Is shipping climate neutral? Expected: the reply from the pair without a neutrality claim, or, if climate neutral is entered as an excluded topic, the handover.
• Is the whole product made of recycled material? Expected: the reference to the part the statement applies to.
• Will you go plastic-free soon? Expected: the no-information reply or the handover, no announcement.
• Does the item carry an eco label? Expected: only what the product text says, with its source.

The set passes when no reply contains a generic green word and every environmental statement carries its reference in the same reply. The questions are repeated after every change to sources or pairs, because a new catalogue sync can bring back an old product text. How such an acceptance test is structured in general is set out in acceptance testing before go-live.

What remains open

Some questions do not belong in an assistant, nor in this article. Whether a particular environmental claim is permissible, whether recognised excellent environmental performance can be demonstrated for it and whether a link is sufficient as a specification on the same medium in an individual case depends on the goods, the presentation and the evidence, and is a matter for professional advice. XICBOT does not assess wording, does not recognise environmental claims as such and does not make advertising unobjectionable under competition law. What the technology contributes is narrower and verifiable in return: it binds replies to sources, sends maintained answers out word for word, hands excluded topics over and shows what was asked.

Frequent questions