Voice as an input method is no longer a question of the future but one of implementation. 26% of companies in Germany use artificial intelligence (Federal Statistical Office); among them, 42% rely on speech recognition and 34% on methods for generating natural language (Federal Statistical Office). Both halves of a voice interface – listening and answering – are therefore already present in many businesses, but rarely in the online shop itself. This article sorts out what can be documented about voice control, where it measurably helps a shop, and which obligations around accessibility and data protection come with it. Anyone planning the AI features of a shop decides first about the data paths, not about the voice.
How widespread voice technology really is in commerce
The basis for any consideration of voice control is the reach of online retail itself. 95% of people aged 16 to 74 in Germany used the internet in the past three months, 86% have shopped online at least once and 70% did so in the past three months (Federal Statistical Office). 80% of this age group search online for information about goods and services (Federal Statistical Office). Searching for products is therefore the most common reason to enter a shop at all – and that is exactly where voice control comes in.
On the company side, the use of artificial intelligence has grown sharply within a year: 41% of companies with twenty or more employees use it, compared with 17% a year earlier (Bitkom). Among the companies that use it, 88% apply the technology in customer contact (Bitkom). The survey is based on 604 companies with twenty or more employees interviewed by telephone, so it describes the commercial cross-section rather than retail alone.
- Use differs clearly by company size: 26% of all companies use artificial intelligence, rising to 36% among those with 50 to 249 employees and 57% from 250 employees upwards (Federal Statistical Office).
- Among the companies that use it, text mining leads with 52%, while automating workflows and decisions reaches 27% (Federal Statistical Office) – voice control needs both building blocks.
- The market for the underlying language processing is growing by 25.7% per year in the 2026 to 2031 forecast period (MarketsandMarkets), so the tools are becoming cheaper rather than scarcer.
None of the surveys cited measures how many orders are triggered by voice. Anyone planning voice control should therefore not judge the benefit by checkout alone, but by search quality, accessibility and operating time – figures that can be measured in your own shop instead of being taken from someone else’s market forecast.
What shop search delivers today
Before voice enters the picture, it is worth looking at the existing search. In the Baymard Institute study, 56% of the shops and apps examined do not adequately support their users’ search queries (Baymard Institute). This does not refer to typing errors but to queries that describe a feature, a use case or a product group instead of a product name – precisely the kind of query someone would also speak to a device.
The technical equipment is not the problem. 80% of the shops examined offer an autocomplete list, but only 19% implement all of the tested design patterns correctly (Baymard Institute). When a search returns nothing, half of the shops offer no effective way out: 50% lead neither to an alternative result list nor to a sensible correction (Baymard Institute). A voice interface inherits these weaknesses unchanged, because it uses the same search logic.
| Metric | Value | Source |
|---|---|---|
| Internet use among people aged 16 to 74 | 95% | Federal Statistical Office |
| Online purchase in the past three months | 70% | Federal Statistical Office |
| Companies with 20+ employees using AI | 41% | Bitkom |
| Shops that do not adequately support search queries | 56% | Baymard Institute |
| Shops with no way out on zero results | 50% | Baymard Institute |
| Annual growth of language processing 2026–2031 | 25.7% | MarketsandMarkets |
What a weak search costs was surveyed among more than ten thousand consumers and two hundred website managers in US retail: 76% of the consumers surveyed report a purchase lost to a failed search (Google Cloud). The figure comes from the US market and cannot simply be transferred to Germany; it does describe behavior that every shop operator can verify in their own search log.
Spoken queries are longer than typed ones
Spoken queries are longer than typed ones, and longer queries are answered differently. For searches of one or two words, a machine-generated answer appeared in 8% of cases; for searches of ten words or more, it appeared in 53% of cases (Pew Research Center). People who speak tend to form a sentence rather than a keyword – so the query more often lands in an environment where a machine summarizes instead of returning a result list.
Recording
The browser records a short audio signal. What matters here is not sound quality but whether the recording leaves the device and how long it stays anywhere.
Transcription
The audio signal becomes text. For shop search, a transcription without speaker identification is enough; voice profiles are not required for ordering and only raise the protection needed.
Intent
The sentence is broken down into attributes: product group, feature, price range, quantity. This breakdown is the part a language model takes over – and the part that should be logged traceably.
Matching
The attributes are matched against the catalog. Whatever is missing here is missing from typed search too: clean product attributes, maintained synonyms and reliable availability.
Answer
The shop answers with a selection, not with a single item. A voice interface that adds to the cart immediately takes away the decision the customer wanted to make.
Handover
Checkout continues along the normal order path, with the same mandatory information and the same confirmation. A second, voice-only order path would mean double effort and double risk.
That machine-generated answers are no longer a marginal phenomenon is clear from their reach: according to the provider, the AI overviews in search have reached more than 1.5 billion users in 200 countries and regions (Google). For a shop this means two things: its own search has to understand longer queries, and its product pages have to be structured so a machine can reproduce them correctly – both prerequisites for a spoken question to reach your own range at all.
Voice control is first a question of access
The most solid reason for voice control is not revenue but access. At the end of 2025, a good 7.8 million people with severe disabilities lived in Germany, which corresponds to 9.4% of the population (Federal Statistical Office). Worldwide, around 16% of people live with a significant disability, about 1.3 billion in total (World Health Organization). For part of this group, voice is the only comfortable way to fill a shopping cart.
Economically the group is larger than it usually appears in planning: the extended market covers 2.3 billion people controlling 6.9 trillion US dollars in annual disposable income (W3C Web Accessibility Initiative). The state of websites does not match that. WCAG 2 failures were detected on 95.9% of the home pages examined, and 33.1% of form inputs were not properly labeled (WebAIM). In a test by Aktion Mensch, 20 out of 65 shop websites examined could be operated by keyboard alone – voice control starts exactly where operation without a mouse ends today.
69% of respondents with access needs simply leave a website with barriers rather than working their way through it, and 83% prefer to shop where they already know the site is accessible (Click-Away Pound). 91% of respondents in an industry survey say digital accessibility improves the overall user experience (Level Access). The benefit is therefore not limited to one group.
Since June 2025, electronic commerce with consumers has been covered by the German Accessibility Strengthening Act. Depending on the offense, the fine ranges up to 100,000 euros, and up to 10,000 euros in the remaining cases (BFSG). Microenterprises are exempt: they employ fewer than ten people and either generate at most 2 million euros in annual turnover or have an annual balance sheet total that does not exceed this figure (BFSG). Anyone wanting to check their own status will find the procedure in our article on the accessibility audit and on our accessibility page.
Language in the markup: what a shop needs
Speech output and voice control require the page to declare its language at all. Success criterion 3.1.1 of the WCAG requires the default human language of each web page to be programmatically determinable (W3C). The European standard takes this up in clause 10.3.1.1 for the document language and in clause 10.3.1.2 for the language of individual passages (ETSI). For a shop with German descriptions and English manufacturer details this is no formality: without the declaration, a screen reader reads the English paragraph with German pronunciation.
How often this declaration is missing is shown by the annual evaluation of one million home pages: the document language was missing on 13.5% of them (WebAIM). The fix costs nothing – a single attribute on the root element – and is still one of the most common findings of all.
Multilingualism is not a marginal export topic either. 76% of respondents prefer products with information in their own language, and 40% will not buy from sites in other languages (CSA Research). Within the European Union, 59% of respondents can hold a conversation in at least one additional language (Eurobarometer) – which does not mean they want to order in it. A voice interface therefore has to carry the same language logic as the rest of the shop.
A spoken query hits the same catalog as a typed one. If color, size, material or intended use are missing as maintained attributes, no language model can derive a selection from them – it will guess, and guessing is more expensive in retail than an empty result list. The effort for voice control therefore lies mostly in the product data and not in speech recognition.
Data protection: what applies to voice recordings
A voice recording is personal data, and access to the end device is regulated separately. Under section 25 of the German Telecommunications Digital Services Data Protection Act, storing information on the end device and accessing information already stored there is only permitted with consent unless it is strictly necessary (Federal Ministry of Justice). The processing itself is governed by the principles of Article 5 GDPR, the legal basis by Article 6 GDPR and the design by Article 25 GDPR – data protection therefore belongs in the architecture of the voice feature and not in a notice underneath it.
- The microphone only becomes active after a deliberate action, never in the background and never permanently.
- The recording is discarded after transcription; at most the text is stored, and only for as long as it is needed for the order.
- Voice profiles, speaker identification and emotion analysis stay out of scope: they are not needed for search and raise the protection required considerably.
- The data flows between shop, transcription and language model belong in the record of processing activities under Article 30 GDPR.
- Transmission and storage need technical measures appropriate to the risk under Article 32 GDPR; a voice recording over an unsecured connection is not a defensible state.
- If a high risk to the rights and freedoms of data subjects is likely, a data protection impact assessment is required under Article 35(1) GDPR.
- Consent must be revocable, and the revocation has to switch the voice path off in fact, not merely hide it.
- Children and guests on the same device are the normal case, not the exception: a voice order without a second confirmation is therefore a poor default.
Where the language model runs
Where the language model computes determines how much protection is needed. The European Data Protection Board adopted its opinion on AI models on 17 December 2024 and states that a model can only be considered anonymous under two cumulative conditions (EDPB). For externally operated models, the German Federal Office for Information Security explicitly describes the possibility that the operating company accesses inputs and outputs and uses them for further training (BSI). For a voice feature in a shop, this is the decisive fork in the road.
- No sensitive data may be embedded in system prompts (BSI) – a voice feature that writes customer data into the prompt breaches this requirement.
- Local operation means running an AI model under your own responsibility on your own hardware (BSI); that is the benchmark against which any outsourcing has to be measured.
- For any planned fine-tuning, local processing on the device in use is to be preferred (DSK).
- Transcription in the browser keeps audio data on the device and makes consent for transmission unnecessary, as long as nothing else is transferred.
- If processing is outsourced after all, the data processing agreement belongs before the first recording and not after the first incident.
- A failure of the voice path must not stop the shop: typed search remains the main path, spoken input is the addition.
Technically, running in the browser has become the realistic option. The interface for computations on the graphics unit is fully supported in 85.72% of measured browser sessions and partially supported in 1.63% (caniuse.com). A language model in the browser retains up to 80% of the native speed of the same device (WebLLM). That is enough for transcription and intent recognition, but not for a full catalog match – which stays on the server. How this can be divided up in a shop is shown in the article on AI features without data transfer.
Visibility: what voice answers quote from a shop
The second reason to deal with spoken queries lies outside your own shop. Some visitors now arrive through an assistant answer rather than a result list. That traffic behaves differently from the rest: anyone arriving this way has the pre-selection behind them and is checking rather than searching. How large the share is in your own shop is shown by the referral sources in your visitor report.
53% of the pages cited by AI systems use structured data (Ahrefs). Adding the markup retroactively to 1,885 pages, however, produced no measurable increase in citations in the same study (Ahrefs). Structured data is therefore the prerequisite for being reproduced correctly, not a lever that can be pulled afterwards.
Adoption is lower than the debate suggests. The Product schema appears on 0.77% of all pages measured (HTTP Archive), while JSON-LD overall is present on 41% of the pages recorded (Web Almanac). For the AI features of search, the provider explicitly requires no special schema and no additional machine-readable files (Google). Anyone marking up product data cleanly therefore does so for the accuracy of the reproduction, not for a promised ranking.
Implementation: from voice input to checkout
At the end comes checkout, and that is the toughest spot in the shop. The cart abandonment rate in online retail, averaged across more than fifty studies, is 70.22% (Baymard Institute). Large shops can raise their conversion rate by an average of 35.26% through better checkout design (Baymard Institute). A voice interface affects these figures only if it saves steps – not if it places a second path alongside the existing one.
- Search first: before voice is added, synonyms, attributes and the handling of zero-result queries are put in order in the typed search.
- Voice as input, not as a separate channel: the transcription result lands in the existing search field so that logging, filters and analysis keep working unchanged.
- Shorten checkout: an average checkout flow shows 23.48 form elements, of which 14.88 are actual input fields, while 12 to 14 elements count as a short flow (Baymard Institute).
- In most checkout flows the number of form elements shown can be reduced by 20 to 60% (Baymard Institute) – this step works on every input method, spoken input included.
- Measure instead of estimating: zero-result queries, drop-offs per step and operating time per input method belong in a measurement before and after launch, otherwise the benefit remains a claim.
Speed is no side issue here. A mobile load time faster by one tenth of a second was associated with 8.4% higher conversion and 9.2% higher average order value in retail (Deloitte). A voice feature that pulls in a model while the page is loading can consume that gain on its own – it therefore belongs behind a deliberate action and not in the first page load.
Implementing voice control with XICTRON
XICTRON builds voice control as part of the shop, not as an add-on. We first examine the existing search and the product data, then clarify which processing can stay on the device, and connect voice input to the existing order path – with logging, revocation and a typed main path that works completely without the voice feature. Our work on AI features and on accessibility goes hand in hand here, because the same places in the shop are affected.
The usage figures come from the Federal Statistical Office and Bitkom, the market forecast from MarketsandMarkets. The shop search figures come from the Baymard Institute and from a Google Cloud survey, the figures on query length and reach from the Pew Research Center and from Google. The accessibility figures come from the Federal Statistical Office, the World Health Organization, the W3C Web Accessibility Initiative, WebAIM, Aktion Mensch, Click-Away Pound and Level Access. The visibility figures come from Ahrefs, the HTTP Archive and the Web Almanac, the figures on technology and checkout from caniuse.com, WebLLM, Deloitte and the Baymard Institute. The legal requirements follow from the German Accessibility Strengthening Act, the WCAG success criteria of the W3C, the ETSI standard EN 301 549, section 25 of the German Telecommunications Digital Services Data Protection Act and Articles 5, 6, 25, 30, 32 and 35 GDPR; publications by the European Data Protection Board, the BSI and the DSK are added to these. The language figures come from CSA Research and from the European Commission’s Eurobarometer. Every figure in this article is backed by a source, a deep link and a verbatim quotation.
26% of companies use artificial intelligence technologies; among those with 50 to 249 employees it is 36%, and from 250 employees upwards 57% (Federal Statistical Office). Among the companies that use it, 34% work with methods for generating natural language (Federal Statistical Office).
56% of the shops and apps examined do not adequately support their users’ search queries (Baymard Institute). Many do offer an autocomplete list, but only 19% implement all of the tested design patterns correctly, and 50% offer no effective way out after a search without results (Baymard Institute).
People with severe disabilities make up 9.4% of the population in Germany (Federal Statistical Office), and worldwide around 16% of people live with a significant disability (World Health Organization). WCAG 2 failures were detected on 95.9% of the home pages examined (WebAIM) – for part of this group, spoken input is the most comfortable way in.
Success criterion 3.1.1 of the WCAG requires the default human language of each web page to be programmatically determinable (W3C); the European standard takes this up for the document and for individual passages (ETSI). The document language was missing on 13.5% of the home pages examined (WebAIM) – without it, a screen reader reads foreign-language passages incorrectly.
Under section 25 of the German Telecommunications Digital Services Data Protection Act, storing information on the end device and accessing it is only permitted with consent unless it is strictly necessary (Federal Ministry of Justice). A legal basis, purpose limitation, a record of processing activities and appropriate technical measures under the GDPR are added to this. The recording should be discarded after transcription.
The averaged cart abandonment rate is 70.22%, and better checkout design can raise the conversion rate of large shops by an average of 35.26% (Baymard Institute). A mobile load time faster by one tenth of a second was associated with 8.4% higher conversion in retail (Deloitte). Voice only contributes to this if it saves steps.