Shop chatbot upkeep: keep delivery times and rules current

A shop chatbot answers questions about delivery times, shipping costs, returns and promotions around the clock – but only as well as the knowledge it draws on. When a deadline, a shipping rule or a promotion changes, the assistant keeps answering with the old information until the source is replaced. This article shows how to maintain your shop chatbot's knowledge: with a validity date for each source, seasonal rules for the holiday season, a source citation under every answer and a weekly list of unanswered questions. If you are just getting started, our introduction to AI chatbots for online retail covers the basics.

Why chatbot knowledge goes stale

A language model does not know your shop. What it says about delivery times, shipping costs or return rules comes from the content it is given when it answers: help pages, FAQ, shipping information, product data, internal notes. If one of these sources goes out of date, the assistant turns it into a fluent, readable and still wrong answer. The tricky part: a wrong answer is usually seen only by the one customer who is asking at that moment. The shop team notices it only when a follow-up question, a complaint or a claim arrives. Unlike a faulty product page, there is no screenshot that several people see at the same time and no colleague who stumbles over it while clicking through. Maintaining the knowledge is therefore not a one-off setup task but part of day-to-day operations – just like prices, stock levels and shipping methods.

The need grows with adoption. Among companies in Germany with 20 or more employees, 36 percent used artificial intelligence in 2025, compared with 20 percent in 2024 (Bitkom). Among companies with 20 or more employees that use AI, 88 percent deploy it in customer contact (Bitkom) – which makes customer service an obvious place to use it. Every assistant that answers customer questions, however, creates an upkeep task that has no slot of its own in many project plans: Who checks the sources? When are they checked? Who decides whether a statement still applies? Without a schedule and without owners, upkeep is postponed until an error forces it. During the launch phase this hardly shows, because the content is fresh. The problems typically start a few months later: when the first season is over, a shipping method has changed or a promotion has ended whose conditions are still in the knowledge base.

What customers expect

When there is a problem with an order, 62 percent of online shoppers would like to use a quickly reachable human contact person, 36 percent want help from a chatbot and 28 percent would use an FAQ on the website (Bitkom, multiple answers possible). The shares are not mutually exclusive and cannot be added up: someone who wants to use a chatbot does not therefore want to do without a person. Two things follow for knowledge upkeep. First, the chatbot and the FAQ have to say the same thing, because both channels are used side by side – a customer who reads a different deadline in the FAQ than in the chat no longer knows which statement to trust. Second, the handover to a person is part of the assistant's knowledge: when it hands over, to whom and at what times is a rule like any other and needs a validity date.

Satisfaction also differs by channel. Of the online shoppers who had contact with a human agent in customer service, 86 percent are satisfied with this service; among users of a chatbot the figure is 50 percent, among users of an FAQ on the website 60 percent (Bitkom). Each figure refers only to the people who used the respective channel. The survey does not give reasons for dissatisfaction. Chatbot answers in shops can fail not only because of the wording but because of the content: the source is missing, has expired or contradicts another one. That is exactly where knowledge upkeep comes in.

Chatbots, by contrast, are currently less in demand in customer service: 36 percent of online shoppers want this kind of automated help.

Bitkom, press release on customer service in online shopping (translated)

Trust is not a side issue here. 38 percent of people in Germany aged 16 and over doubt the reliability of results delivered by AI (Bitkom). An assistant that names the source for every answer makes its statement verifiable: anyone who is unsure opens the linked shipping page and reads it there. This does not make the answer automatically correct, but an error becomes visible instead of disappearing in the conversation. For the shop team, the source citation has a second benefit. If someone reports a wrong answer, it shows directly which entry needs correcting – instead of a search through the entire knowledge base.

Holiday season: the season with its own rules

According to the German Retail Federation (HDE), retail generates 18.5 percent of its annual turnover in the months of November and December; that is what the press release on the holiday forecast of 5 November 2025 states. For November and December 2025, the federation forecast total turnover of 126.2 billion euros, and for online retail a nominal increase of 3.3 percent to 22.2 billion euros (HDE). These are forecasts from the previous year, not actual figures; an HDE forecast for 2026 was not available at the time of writing. For the shop chatbot this means: in the weeks that account for a large part of annual turnover, an especially high number of time-bound rules apply at the same time – shipping cut-offs, extended returns, promotion conditions, changed service hours. Each of these rules has a start and an end, and each can turn up in an answer.

The season opens with the promotion week around Black Friday, which in 2026, as usual, falls on the Friday after Thanksgiving, at the end of November. For Black Friday and Cyber Monday 2025, the retail federation expected turnover of 5.8 billion euros (HDE). According to a survey by IFH Köln on behalf of the HDE among over 1,000 online shoppers, conducted in October 2025 and thus before Black Friday on 28 November 2025, 54 percent of respondents also use Black Friday for Christmas shopping (HDE). Questions about delivery before the holidays can therefore start earlier than some teams publish their Christmas content. If you are preparing the technical side of these days – for example with a virtual waiting room for Black Friday –, plan the assistant's knowledge in the same work step.

• shipping cut-off for delivery before the holidays, separated by shipping method and destination country
• extended return periods for Christmas purchases and their exact end date
• promotion conditions: duration, excluded items, whether they can be combined with vouchers
• vouchers and gift cards: redemption, remaining balance, delivery as a card or a file
• exchanging gifts, for example because of sizes – see our size advice with size charts
• changed service hours between the holidays and around the turn of the year

Legal framework: what the assistant must not get wrong

There is a legal framework for delivery information. Under Article 246a § 1(1) sentence 1 no. 10 EGBGB, an online shop has to inform consumers about the terms of payment, delivery and performance and about the date by which it undertakes to deliver the goods (Federal Law Gazette). This duty is not new. What is new since 27 September 2026 is that this also includes environmentally friendly delivery options, where available (Federal Law Gazette). If a shop offers such an option, the assistant should know it and describe it in the same words as the shipping page. Conversely, it should not state delivery dates that the shop does not state elsewhere. The assistant reproduces the information on the shipping page; it does not make promises of its own.

Competition law also applies. Under Section 5(2) UWG, a commercial practice is misleading if it contains untrue information or other information suitable for deception about certain circumstances; in the version of Section 5(2) no. 1 in force since 27 September 2026, these include, among other things, availability and the method or date of delivery (Federal Law Gazette, gesetze-im-internet.de). Under Section 5(1) UWG, a misleading practice is unfair if it is also likely to cause consumers to take a transactional decision they would not have taken otherwise. If an assistant states an expired delivery time or a promotion that ended long ago, this creates a legal risk – whether there is an infringement in an individual case depends on the circumstances. This makes it all the more important that expired information does not get into an answer in the first place.

Since 19 June 2026, Section 356a BGB has required online shops to provide an electronic withdrawal function that is clearly labeled “Vertrag widerrufen” (cancel contract) or with an equivalent unambiguous wording (Federal Law Gazette, gesetze-im-internet.de). Under Section 355(2) BGB, the withdrawal period is 14 days (gesetze-im-internet.de). If someone asks in the chat how to withdraw from a purchase, the assistant should name the function by its actual label and say where it can be found in the shop. A voluntarily extended return period during the holiday season has to be kept separate from this: it is a goodwill rule of the shop with its own start and end, not a statutory right of withdrawal. If the assistant mixes the two, contradictory statements result. Under Section 312j(1) BGB, it must be stated at the latest at the beginning of the ordering process whether there are delivery restrictions and which means of payment are accepted – the assistant should also take this information from the same maintained source as the order page.

Validity date: every source gets an expiry date

An effective lever is simple: every knowledge entry gets a date from which it applies, a date until which it applies and a responsible person or role. Before it expires, the owner receives a reminder and decides whether the entry is extended, replaced or ended. This way the team does not have to go through the entire knowledge base regularly, only the entries that are about to expire. For entries without a natural end – such as the description of payment methods or the address for returns – a review date is enough, for example every six or twelve months. What matters is that the date is attached to the entry itself and not kept in a separate spreadsheet that is easily forgotten in day-to-day work. An entry that carries its own expiry date can be filtered out automatically as soon as it no longer applies.

# Sample entry, values chosen freely
entry: shipping-christmas-2026
title: Christmas shipping info
valid_from: 2026-11-16
valid_until: 2026-12-24
owner: Logistics
source: /help/shipping-christmas/
reviewed_on: 2026-10-05
successor: shipping-january-2027
languages: [de, en]

Not everything belongs in the knowledge base as text. Delivery status, availability and prices change too often to maintain them by hand. The assistant queries this information at runtime – from inventory management, the ERP or the product data, for example from a PIM system connected to the shop. The rule behind it: whatever an interface can supply is queried, not copied. A text that states the delivery time of a single item may already be wrong the day after it was written. The knowledge base, by contrast, describes the rules – how delivery times come about, what applies to pre-orders, when an item is shown as unavailable. This way there are no two truths for the same fact. In a wholesale shop with prices behind a login, one more condition applies: the assistant should quote prices only to signed-in customers, and only those of their customer group. How such shops keep their product pages visible in search is covered in our article on login prices in wholesale shops.

A source in every answer

The source citation under every answer does three jobs. For customers, it makes the statement verifiable. For the team, it is the direct route to the entry that needs correcting. And in the evaluation, it shows which sources the assistant actually uses – and which do not appear at all. Entries that are not cited a single time over months are candidates for clean-up or a sign that the search does not find them. Entries that are cited frequently deserve particularly careful upkeep: many answers depend on them, and an error there has a correspondingly frequent effect.

Technically, this requires the assistant to answer only from approved entries. For each request, the system first retrieves the matching entries and hands them to the language model with the instruction to answer exclusively on that basis. The title and address of the entries used are appended to the answer. If the search finds no matching entry, the assistant says so openly and offers a handover to a person instead of formulating an answer from general model knowledge. Expired entries are filtered out before the search, so an entry whose expiry date has passed is not available as a source for an answer.

• Every answer names at least one source with title and link.
• Expired entries are filtered out before the search, not afterwards.
• Without a matching entry, the assistant responds with a notice and offers a handover.
• If two valid entries contradict each other, the newer one wins and the conflict goes on the work list.
• Legal texts are linked verbatim, not reworded by the model.
• Each source carries language and destination country, so that answers for Austria or Switzerland do not come from German rules.

Plan seasonal rules in good time

Seasonal rules have one special feature: their end is known before they begin. A season calendar records for each rule when it is published, from when it applies, when it ends and which entry takes its place afterwards. The successor is easily forgotten. If the Christmas shipping info ends on 24 December, questions about deliveries between the holidays still come in – and without a successor the assistant then either does not answer at all or answers with the last valid information, which no longer fits. Planning the successor together with the entry closes this gap before it opens.

1. Early October: collect seasonal topics – shipping, returns, promotions, vouchers, service hours.
2. Assign owners for each topic and create texts with a validity date.
3. Have legally relevant texts approved by legal counsel before publication.
4. Load entries with a start date in advance so that they apply automatically on the key date.
5. Create a successor for every entry that ends before the turn of the year.
6. Evaluate after the season: which questions remained open, which entries came too late?

Many shops already have a fixed rhythm for updates, extensions and security fixes. It makes sense to tie knowledge upkeep to the same schedule: if you install updates monthly as part of Shopware maintenance, you can go through the list of soon-to-expire entries in the same session. And it is not only seasonal rules that change the knowledge base. Mandatory information outside the season belongs in it as well, such as the information on the ElektroG take-back symbol in the online shop, which customers may ask about when buying electrical devices. If such a duty changes, the assistant's answer changes too – but only if someone updates the entry.

Unanswered questions as a work list

A productive source for upkeep is the questions to which the assistant had no good answer: questions without a matching entry, conversations handed over to a person, answers that someone marked as not helpful. Once a week, the team reviews these questions, grouped by topic, and decides on one of four measures for each group: create a new entry, correct an existing entry, improve the search so that an existing entry is found, or deliberately refer the question to customer service because it needs an individual review. The effort stays manageable as long as the list is worked through weekly. If you leave it for months, you face a pile that can hardly be sorted any more.

One rule applies to the review: chat transcripts can contain personal data, such as order numbers, names or addresses. For the evaluation, the question and the topic are usually enough; order numbers and contact details are removed beforehand, and transcripts are kept only for a limited time. How we handle the conversations of our own chat assistant is described in our privacy policy. Upkeep also includes looking at unusual inputs: if someone tries to push the assistant into other statements with crafted instructions, this should stand out in the evaluation. Which safeguards make sense for this is described in our article on prompt injection in AI features in the shop.

Test run before the key date

Before a seasonal rule applies, a short test run with real phrasings is worthwhile. Customers rarely ask about the “shipping cut-off”; they ask: “Will it still arrive before Christmas Eve?”, “By when do I have to order?” or “Can you make it by Friday?”. Five to ten such questions per rule are enough, ideally in the variants that actually reach customer service. The test checks whether the answer names the right source, whether it reproduces the rule correctly and whether it hands over to a person at the right point. A second run with a date after expiry shows whether the assistant really no longer uses the expired rule and cites the successor instead.

Good candidates for test questions can be found in the customer service inbox: the requests from the previous season show how people really phrase things. If you already bundle requests in an AI ticket system in customer service, you can filter them by topic and put together a test collection. Over time, this collection becomes a regression test: after every change of language model, every change to the search and every adjustment of the instructions, it runs again. This reveals when a change in one place makes an answer elsewhere worse – before customers notice it.

Context: where companies stand

How widespread is artificial intelligence in German companies in the first place? The Federal Statistical Office counts companies with 10 or more employees. According to its figures, a total of 26 percent of companies used artificial intelligence in 2025; among small companies with 10 to 49 employees the share was 23 percent, among medium-sized companies with 50 to 249 employees 36 percent (Federal Statistical Office). Companies that considered AI but did not use it mainly cited a lack of knowledge at 72 percent and uncertainty about the legal consequences at 62 percent as reasons (Federal Statistical Office). Both hurdles relate directly to knowledge upkeep: anyone who knows where the assistant takes its answers from and who is responsible for them can also judge which statements are legally sensitive and where approval is needed.

In a European comparison, a total of 19.95 percent of companies in the EU with at least 10 employees used at least one AI technology in 2025 (Eurostat); for Germany, the Federal Statistical Office reports 26 percent (Federal Statistical Office). Both figures come from the same harmonized survey, published by two different bodies. Compared with the previous year, the EU share rose by 6.47 percentage points (Eurostat). In EU retail, 48.18 percent of companies that use AI used it for marketing or sales (Eurostat). Across all sectors, 8.76 percent of companies in the EU used AI technologies that generate written or spoken language or program code (Eurostat) – a share of all companies, not of AI users, and not a pure chatbot figure. The Federal Statistical Office plans to publish new official figures for Germany in November/December 2026 (Federal Statistical Office).

A shop chatbot is only as current as the sources it answers from. The technology behind it is solvable; what matters is the organization: every statement needs a leading source, a validity date and a person who is responsible for it. Start with the topics that are time-bound in the holiday season – shipping, returns, promotions –, give each entry an expiry date and a successor, and review the open questions once a week. This helps keep the assistant's answers up to date even when deadlines, rules and promotions change. If you would like support with this or want to set up a new assistant, talk to us.