How to make sure your services are found in AI answers, correctly understood and named as a relevant solution.
From classic search engine optimisation to a digital presence that ChatGPT, Microsoft Copilot, Google Gemini, Perplexity and Claude can also understand. Written for managing directors, marketing and sales leaders, and digital officers in small and mid-sized enterprises.
This whitepaper covers a field that is changing quickly. Three rules therefore apply to every figure and statement on the following pages.
Metrics on AI usage and AI traffic age within months. We therefore always name the source and the collection period.
No provider guarantees a placement in AI answers. Anyone who promises this is selling a claim, not a service.
Industry averages are useful for orientation. The basis for decisions is measurement within your own company.
There is no disclosed ranking algorithm for ChatGPT, no confirmed weighting of individual signals and no method that forces a mention. What does exist are documented technical prerequisites, sound quality principles and a measurable approach. That is exactly what this document is about.
The way people search for products, service providers and business partners is changing. Prospects no longer type in single search terms at Google alone. They ask complete questions.
Increasingly, AI systems compile the answers to these questions. In Germany, half of all internet users already use an AI chat at least occasionally instead of a classic search engine; among 16- to 29-year-olds it is around two thirds. At the same time, 42 per cent of those who use AI for search report having already received incorrect or fabricated answers, and only 57 per cent check the results before using them further. This is the finding of a representative Bitkom survey from November 2025.
For companies, this creates a new task. A website no longer has to be understandable only for people and classic search engines. It must additionally supply enough unambiguous, current and verifiable information for AI systems to classify the company correctly.
OpenAI explicitly points out that a top placement cannot be guaranteed. A prerequisite for possible inclusion in ChatGPT Search, however, is that the OAI-SearchBot is allowed to access the website. In May 2026, Google also published its first official guide to optimising for generative search features and made it clear there: the fundamentals remain classic SEO fundamentals; no special files or dedicated formats are required.
The goal is not to produce as much AI-generated content as possible. The goal is a robust digital representation of the company that people, search engines and AI systems can all understand equally well.
AI systems change more than the way search results are displayed. They change the entire process of gathering information.
In a classic search, the user receives a list of web pages. They open various results, compare information and draw their own conclusion. In an AI-assisted search, the system takes over a considerable part of this work.
| # | Step in the AI-assisted search process |
|---|---|
| 1 | The system interprets the intent behind the question. |
| 2 | It breaks the question into several sub-questions and forms its own search queries from them. Google calls this "query fan-out"; ChatGPT works on the same principle. |
| 3 | It searches for suitable sources and retrieves the pages. |
| 4 | It compares statements and weighs up contradictions. |
| 5 | It formulates a consolidated answer. |
| 6 | It names, links or recommends selected sources and providers. |
This shifts the competition. A company no longer competes merely for a place in a results list. It competes to be included in a consolidated answer as a suitable piece of information, a reliable source or a fitting solution. An important side effect: the sub-questions formulated by the system are often not what the user typed. Microsoft now makes these internal search queries visible under the term "grounding queries" — more on this on page 24.
Within the AI answer itself, the user already learns which solution approaches exist, which providers seem relevant, what the differences are, which risks to consider, what costs are realistic and which criteria matter for a decision. The click through to the company website often happens only later. In some cases it does not happen at all.
A Pew Research Center study of Google searches in March 2025 showed: when an AI summary appeared, users clicked on a classic search result in only 8 per cent of visits. Without a summary it was 15 per cent. Only one per cent clicked on a source within the summary. And 26 per cent then ended their session entirely, compared with 16 per cent without a summary.
The Pew study is based on the browsing behaviour of 900 US adults in March 2025. Google publicly criticised the methodology, among other things because the query set was not representative and the comparison period was shifted. The results also relate to Google, not ChatGPT. They do not prove an exact click-loss statistic for your company. But they do show the direction: answer systems change the previous click path considerably.
A brand can influence a purchase decision without that influence showing up as a site visit. AI visibility must therefore measure more than traffic. What also matters:
The term AI visibility is often used too loosely. For a sound assessment you should distinguish at least five levels.
Your content can be found and processed technically. Example: a service page is reached by the OAI-SearchBot and becomes eligible as a source for a current search answer.
A page of your company is named as a source or a link to read more. Example: when asked about introducing an enterprise GPT, ChatGPT points to your guide.
Your company is named in the answer without a page of yours necessarily being linked. Example: "Providers in this area include companies A, B and C."
Your company is classified as a fitting option for a concrete requirement. Example: "For SMEs that need EU hosting and structured AI governance, provider C is particularly worth considering."
Offering, target group, location, scope of services and positioning are reproduced correctly. This level is the most important: an incorrect or outdated representation can be more damaging than a missing mention.
| Type of visibility | Typical effect |
|---|---|
| Source | The company is perceived as professionally relevant |
| Brand mention | The brand makes it onto the shortlist |
| Recommendation | The brand is linked to a concrete requirement |
| Link | The user can visit the website directly |
| Correct representation | Trust and expectation match the actual offering |
A company should therefore not only ask: "Are we mentioned in ChatGPT?" The better question is:
This distinction has direct consequences for management. If discoverability is lacking, it is a technical matter for IT. If source presence is lacking, it is a content matter. If recommendation is lacking, it is usually a positioning and evidence matter. And if the representation is wrong, it is first a matter for the leadership, because the company's own description in the market is then simply incorrect.
OpenAI operates several crawlers with different tasks. Treating them as "the one AI bot" regularly leads to wrong decisions.
| User agent | Purpose | Effect of a block |
|---|---|---|
| OAI-SearchBot | Search features in ChatGPT. No training. | According to OpenAI, your content generally does not appear as a content source in ChatGPT search answers. Navigation links may nonetheless appear under certain circumstances. |
| GPTBot | Collection of content that may contribute to training generative base models. | Your content will no longer be collected for model training in future. Models already trained retain the knowledge they hold. |
| ChatGPT-User | User-initiated retrievals: when someone asks ChatGPT to look at a specific page. | ChatGPT cannot retrieve your page live during a user request. |
The key point: the settings are independent of one another. You can allow the OAI-SearchBot and block the GPTBot at the same time. Your website is then in principle accessible for current ChatGPT search results without you allowing your content to be used for model training. For many SMEs, this is precisely the sustainable default setting.
User-agent: OAI-SearchBot
Allow: /
User-agent: ChatGPT-User
Allow: /
User-agent: GPTBot
Disallow: /
Sitemap: https://www.example.com/sitemap.xmlThis configuration is not a general recommendation for every company. It has to fit your data, copyright and publication strategy. For a company with strongly differentiating expertise, a training block can make sense; for a company that primarily wants to become known, the trade-off may come out differently. What matters is that the decision is made deliberately and documented.
OpenAI additionally recommends not blocking access from its published IP ranges via hosting provider, firewall, web application firewall or content delivery network. In practice, this is the most common reason for missing mentions: the robots.txt allows access, but a bot-management rule at Cloudflare, Akamai or AWS answers the requests with an error code. For ChatGPT, the website is then effectively absent — without anyone noticing.
According to OpenAI, changes to robots.txt take roughly 24 hours to be taken into account in the search features. For verification, OpenAI publishes a machine-readable IP list for each bot.
Granting the crawler access does not mean that all pages are indexed, that content appears for every fitting question, that your company is preferred, that a particular position is reached or that the representation always stays identical. OpenAI names relevance and reliability as general factors, but does not publish a complete list or weighting of the signals used. There is no guaranteed top position.
ChatGPT is the largest channel, but not the only one. Optimising for a single provider builds a concentration risk.
For most SMEs in the German-speaking region, four system families are relevant. They differ technically less than the marketing debate suggests — the common denominator is always the same: accessible, clearly structured, evidenced content.
| System | Access | What you should know |
|---|---|---|
| Google AI Overviews, AI Mode google.com |
Googlebot (search), Google-Extended (AI training) |
Google-Extended controls training only, not visibility in search. Since June 2026 there is also a dedicated switch in Search Console to exclude content specifically from generative search features. |
| Microsoft Copilot, Bing bing.com |
bingbot | Microsoft stresses that its grounding infrastructure supplies numerous AI assistants with web content. Microsoft recommends XML sitemaps and IndexNow and, as the only provider, offers detailed citation data. |
| Perplexity perplexity.ai |
PerplexityBot, Perplexity-User |
Smaller, but with high purchase intent in B2B research. Not included in the standard reporting of many analytics tools — see page 24. |
| Anthropic Claude claude.ai |
ClaudeBot (training), Claude-SearchBot, Claude-User |
Present in SMEs mainly in the enterprise context. Same logic: the search crawler and the training crawler can be controlled separately. |
For most companies with commercially used, publicly accessible content, a pragmatic baseline follows from this:
A block that no one knows about is the costliest form of invisibility. Check your server logs: which AI bots actually accessed your site in the last 30 days? What status codes did they receive? Do your systems answer with 200 — or with 403 and 429? This analysis takes half a day and is often the most effective single step in the entire project.
Despite different bots and interfaces, the requirements are remarkably uniform: clean, semantic HTML with headings, lists and tables; content that is present in the source even without JavaScript; unambiguous statements instead of advertising phrases; verifiable evidence; a recognisable date of last update. Anyone who meets these fundamentals is working for all systems at once — and does not need to maintain platform-specific special files.
Robust AI visibility arises from six interconnected layers. If one layer is missing, the other five carry only so far.
The relevant content must be reachable, crawlable and processable: a correct robots.txt, a reachable XML sitemap, working canonical tags, no unintended noindex instructions, stable status codes, short load times, content without a login, understandable HTML.
AI systems must be able to recognise what your company is called, where it operates, which services it offers, which customers it works for, which locations and contacts exist, and which profiles and directory entries belong to it.
The website must answer the questions that potential customers actually ask. Not just "what do we offer?", but also: who is the solution suitable for? What are the prerequisites? What alternatives are there? How does a project run? Where are the risks and limits? Which systems can be integrated? What types of cost should be expected?
Statements need a traceable basis: standards and norms, studies, your own project experience, concrete approaches, comprehensible examples, authors with a professional profile, a date of last update, and transparent limitations.
Company information should also appear consistent and credible outside your own website: industry directories, trade portals, partner pages, associations, trade media, credible customer references, company profiles and public registers.
AI answers are dynamic. Visibility must therefore be checked repeatedly and improved step by step. A one-off optimisation without measurement is a claim, not a result.
The six layers build on one another. Excellent content on a website that the crawler cannot reach is ineffective. A perfectly accessible website without an unambiguous company identity leads to confusion. And an unambiguous identity without professional substance produces mentions, but no recommendations. Anyone who starts with layer 3 before layer 1 is secured is investing in content that no one reads.
Before you create new content, you should assess the current status in a structured way. The following 100-point model is a practical decision aid, not an official standard from an AI provider.
| Points | Interpretation and typical next steps |
|---|---|
| 0 – 39 | Essential foundations are missing. Start with technology and company identity, not content. |
| 40 – 59 | Technical and content basis in place. Now build out service pages and evidence. |
| 60 – 79 | A good starting point for targeted optimisation. Set up the prompt universe and measurement. |
| 80 – 100 | Advanced, measurable AI visibility. Focus on representation quality and share of voice. |
Classic SEO projects often begin with a keyword list. For AI visibility, that is not enough.
In ChatGPT, a user rarely writes just "AI consulting Stuttgart". They ask a situation-specific question:
This single question contains seven dimensions: service, company size, location, technology environment, data protection, integration needs and decision situation. It is precisely these dimensions that determine whether your company appears in the answer. A keyword approach captures just one of them.
The most reliable source for your prompt universe is not a keyword tool, but your own sales team. Review the last 50 initial conversations: which questions were actually asked? Which objections came up? Which comparisons were drawn? These questions are already commercially validated — someone asked them just before spending money.
For each core service you should document around 30 to 60 business-relevant questions. This collection is the foundation for both content planning and measurement.
| Criterion | Guiding question |
|---|---|
| Proximity to purchase | How short is the path from this question to an enquiry? |
| Revenue potential | What order value typically hangs on this question? |
| Strategic relevance | Does the question fit the positioning we want to build? |
| Available evidence | Can we back up the answer with our own experience or data? |
| Competitive intensity | How many providers already answer this question well? |
| Regional relevance | Does location or catchment area give us a real advantage? |
A simple table is enough. What matters is that it is maintained and that sales and marketing use the same list. A proven structure:
| Field | Example |
|---|---|
| ID | EGP-014 |
| Question, verbatim | "Which providers introduce an enterprise GPT for a mid-sized company running Microsoft 365?" |
| Question class | Provider question |
| Core service | Enterprise GPT |
| Priority | High |
| Responsible page | /services/enterprise-gpt/ |
| Evidenced by | Manufacturing case example, approach model, author profile |
| Last test | 14 Jul 2026 · mentioned: no · source used: no |
Every important question needs several phrasings, because AI systems are sensitive to context. Create at least four variants for each core question: a short version, a detailed one, one with an industry reference and one with a location reference. Variants with company size and with a budget or other constraint are also useful.
Three core services · 30 prioritised questions each · four variants each · three repetitions per test. That comes to 1,080 individual queries per full measurement run. For a quarter, that is feasible with a structured approach. Anyone who needs more should consider tool support — see page 30.
Many companies build a prompt universe out of questions they would like to answer. But only questions that are actually asked are useful. If your sales team has never heard a question and it does not appear in any support enquiry either, it does not belong on the list — no matter how good your answer to it would be.
An AI visibility strategy does not need an uncontrolled mass of articles. It needs a coherent information architecture in five layers.
The central company page answers: who is the company? What does it offer? Who does it work for? In which regions does it operate? What sets it apart from other providers? What are its professional focus areas? This page is the anchor for every classification — it must be written so that a system can form a correct sentence about your company from it.
Each important service gets its own page. Examples from the Kramberg portfolio: AI employee, enterprise GPT, Company Brain, AI governance, AI telephony, AI visibility. A service page must not contain only advertising statements. It must explain which problem is solved, how the solution works, what the prerequisites are, which systems are involved, where the limits are and how a project runs.
The same solution is described differently in different industries. Industry language improves not only the effect on the reader — it provides unambiguous context signals.
Each prioritised use case gets a concrete description: starting situation, previous process, target process, roles involved, data required, interfaces, controls and expected benefit. These pages are above-average in how citable they are in AI answers, because they answer a question fully instead of merely asserting an offering.
Five service pages with real substance work better than fifty blog articles without. When you have to choose between a new article and deepening an existing service page, almost always choose the deepening. In its own guide to generative search features, Google explicitly warns against scaled content production designed primarily to influence search systems.
A good service page combines sales impact with professional substance. Nine building blocks that have proven themselves in practice.
"An enterprise GPT is a dialogue-based AI application that searches approved company information and provides answers with traceable sources." The definition should be precise enough to be quoted or summarised — a system can adopt this sentence without distorting it.
Not: "For companies of any size." Better: "Suitable for mid-sized companies with recurring knowledge questions, multiple document sources and clearly definable access rights." A target group that includes everyone excludes everyone in AI answers.
Knowledge is spread across file shares, emails and specialist applications. Employees spend a long time looking for current documents. Answers depend on individual experienced people. New employees need long onboarding times.
Select sources · adopt permissions · prepare content · configure search and answer generation · enable source display · test quality · monitor operation.
Available knowledge sources, subject-matter owners, documented permissions, suitable example queries and defined quality criteria.
No automatic guarantee of error-free answers. No release of sensitive content without a permissions concept. No replacement of professional responsibility. No lasting quality without keeping the sources up to date.
Initial assessment · data and process review · pilot · professional validation · go-live · operation and further development.
Case example, demonstrator, methodology, checklist, technical architecture and a named subject-matter author.
A CTA that fits the decision situation: "Check suitability for your own use case." Not "Contact us now" — that is not an action someone wants to take in the middle of researching.
It is a quality signal. Providers who name limits supply AI systems with exactly the differentiating information needed for comparison and selection questions. Anyone who asserts only advantages is treated as interchangeable — and usually is.
Even excellent content stays invisible if it is not reliably accessible technically. This is the least spectacular and, at the same time, the most effective layer of the entire model.
For Bing and its AI-assisted interfaces, Microsoft recommends XML sitemaps and IndexNow among other things. IndexNow actively informs search systems about newly created, changed or removed URLs instead of waiting for the next crawl. The effort is small; the benefit is noticeable for frequently updated content.
Important information should also be present in the delivered HTML. Fully JavaScript-based pages can in principle be processed, but they increase technical complexity and the risk of failure. A simple test: disable JavaScript in a fresh browser profile and load your most important service page. If the service description, target group and contact are then no longer visible, a crawler without JavaScript execution sees the same nothing.
Mobile presentation, short load times, legible font sizes, descriptive alt text, understandable link text and a clear separation of main content and navigation. These requirements serve your visitors first. That they also make machine processing easier is a welcome side effect, not an end in itself.
Structured data is useful. But it is not a secret ChatGPT code — and anyone who sells it as one should raise your suspicion.
Structured data describes content in a machine-readable format. It helps search systems classify organisations, people, offerings and page types more clearly. Google confirms that structured data provides explicit clues to a page's meaning and, among other things, can make information about companies and people easier to understand.
In its guide to generative search features, Google makes clear that no special machine-readable files and no particular schema.org markup are required for inclusion in these features. There is no dedicated GEO or AI schema. For ChatGPT, too, there is no official statement that particular Schema.org types bring about a preferred placement.
Depending on the website, the following may be relevant: Organization, LocalBusiness, Person, Service, Product, Article, BreadcrumbList, FAQPage and ContactPoint. For advisory services, Service is the fitting type; for software products, SoftwareApplication.
The structured information must match the visible page content.
Invisible services, invented reviews or unverifiable claims must not be
added in the markup alone — that is not only ineffective but can be
treated as a violation.
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Example Ltd",
"url": "https://www.example.com/",
"logo": "https://www.example.com/logo.png",
"description": "Consulting and delivery of digital solutions for mid-sized companies.",
"address": {
"@type": "PostalAddress",
"streetAddress": "1 Example Street",
"postalCode": "70173",
"addressLocality": "Stuttgart",
"addressCountry": "DE"
},
"contactPoint": {
"@type": "ContactPoint",
"contactType": "sales",
"email": "contact@example.com"
},
"sameAs": [
"https://www.linkedin.com/company/example",
"https://www.xing.com/pages/example"
]
}
</script>sameAs field is the most important partIt connects your website with your other profiles and thus makes it machine-traceable which information across the web belongs to the same company. This is exactly where the mapping most often fails in practice. Maintain all approved company profiles here — and only those.
Structured data is a supporting measure. It does not replace understandable service pages, robust company information, professional evidence, consistent external profiles or technical accessibility. If your budget stretches to just one thing, do not spend it here.
AI systems must recognise which information belongs to the same company. Different company names, old addresses, contradictory service descriptions or unmaintained profiles make this mapping considerably harder.
Each of these contradictions is harmless on its own. In sum they produce a blurred picture — and a blurred picture reliably leads, on selection questions, to a system preferring the provider it can classify unambiguously.
Maintain a binding master-data overview. It is the basis for the website, directories, partner pages, press profiles and social-media presences — and it ends the debate about how the company actually describes itself.
| Field | Binding entry |
|---|---|
| Brand name | Consistent spelling, including capitalisation |
| Company name | Designation as per the commercial register |
| Short description | One or two approved sentences |
| Long description | 100 to 200 words |
| Core services | A maximum of five to seven focus areas |
| Target customers | Company size, roles and industries |
| Region | Area of operation and locations, shown separately |
| Contacts | Name, function and profile |
| Profiles | Complete list of approved company profiles |
| Update | Owner and review date |
A clean master-data record is the one measure in this whitepaper that works immediately, costs nothing and serves every system at once. It is also the prerequisite for marketing, sales and leadership describing the same company — which, in our experience, is not a given.
A common claim is that AI systems predominantly prefer external sources. Other studies reach the opposite conclusion. Both are true — and that is precisely the insight.
A Yext analysis of 6.8 million source citations from 1.6 million answers across ChatGPT, Gemini and Perplexity found that 86 per cent of the sources came from areas that companies own or can substantially influence. Company websites accounted for 44 per cent, directory listings for 42 per cent. The analysis took user intent and location into account and focused on certain industries and query types.
Another study from 2026 analysed around 168,000 URL-based source citations for 128 brands in 13 languages. There, 85.7 per cent went to external websites and only 14.3 per cent to brand-owned domains.
The sources used depend, among other things, on: the type of question, the user's location, the industry, the language, brand awareness, the current news situation, the AI platform in question, the search mode and the time of the query. For a company with a strong brand, the distribution looks different than for a regional specialist. For a comparison question, different than for a provider question.
Control: full
Control: high, but maintenance-intensive
Control: low — but the most credible
The best strategy combines all three classes instead of relying solely on your own website or on artificially created mentions. Class 2 is the most underrated lever here: directory and partner profiles are quick to update, cost no editorial budget and at the same time clean up the contradictions from Chapter 14.
Mass-generated directory entries, bought mentions and artificial forum posts. In its guide, Google explicitly names "inauthentic mentions" as a tactic to ignore. Beyond the lack of benefit, the legal position in Germany is clear: disguised advertising and fabricated references are open to challenge under competition law.
AI systems need content from which an answer can be justified. General advertising statements are poorly suited to this.
"We offer innovative, holistic and future-proof AI solutions."
Interchangeable. Contains no verifiable information. No answer to a selection question can be formed from this sentence.
"We introduce enterprise GPT solutions for mid-sized companies. The scope covers the selection of approved knowledge sources, role and permission concepts, source display, quality tests and handover into managed operation."
Contains service, target group, approach, concrete components and a boundary.
The foundational study on Generative Engine Optimization by Aggarwal and colleagues, produced in part at Princeton University, examined 10,000 queries on a purpose-built test corpus. Adding sources, citations and statistical figures was able to increase measured visibility in certain tests by up to 40 per cent. The second part of the result is notable: the effectiveness was strongly dependent on the subject area. The 40 per cent is a maximum value under laboratory conditions, not a guarantee of success and not a figure with which to justify a budget.
A specialist article without a named author is an anonymous claim. A specialist article with an author profile, professional background and a traceable history is an attributable statement. For the question of whether a system uses your statement as evidence, this difference matters more than any formatting.
AI visibility becomes commercially relevant when a company is linked to concrete requirements. That needs the language of the industry, not the language of the brochure.
A generic text talks about "efficiency gains through AI". An industry-appropriate text talks about the things the people in that industry actually work with.
The right industry language shows practical experience, increases relevance for concrete user questions, sets your services apart from generalist providers, makes industry classification easier, creates citable statements and improves the quality of quotation and comparison questions. A system that has to decide which of eight providers fits a question about RSA 21-compliant site protection can do nothing with a text about "innovative traffic solutions".
Technical terms must not be inserted as a list. They must stand in a professionally correct context. A list of twenty terms without explanation is the keyword-stuffing reflex in new clothing — it works as little today as it did in 2010, and it damages your credibility with the very readers you want to reach.
Have an experienced person from the relevant department read your service page. If they say "no one here would put it that way", the text is not finished. If they start talking about concrete projects, you are on the right track.
For regionally active companies, a nationally generic service description is not enough. ChatGPT can use location information to tailor results and recommendations — via an approximate IP-based assignment or a voluntarily enabled, more precise location share.
"We are also here for you in Stuttgart."
"From our Stuttgart location we support mid-sized companies across Baden-Württemberg with introducing AI employees, enterprise GPT solutions and AI governance. Projects can be delivered on site or in a hybrid model."
After that, a good location page adds: typical industries, scope of services, regional contacts, approach, a real regional project situation and how to reach you. Only these components turn a mention of a place into information.
Automatically generated pages in which only the place name is swapped provide no independent value. They have been a familiar pattern to search engines for years and are worthless to AI systems, because they contain no distinguishable information. Regional pages make sense only if genuine regional information is available: a location, a contact, a delivery area, a response time, a reference, a regional service offering or an industry-specific focus.
Regional queries are almost always the questions with the best hit rate for SMEs. Competition is manageable, purchase intent is high, and the necessary evidence exists — it just often is not on the website. In our measurements, the mention rate for regional questions is regularly well above that for national questions. Anyone with limited resources should start here.
A provider with 90 employees offers maintenance, repair and modernisation of industrial plant. The example is illustrative and does not describe a single real company.
The website has a generic home page, a "Our services" page, several short news items, no industry-specific use cases and no detailed information about the service process. For a question such as "Which service providers support mid-sized production plants with manufacturer-independent maintenance and retrofit?", the company is not mentioned.
The website does name "maintenance" and "modernisation", but provides no answers to the questions that actually count for a selection decision:
The company no longer delivers merely a general claim, but concrete information for research, selection and decision questions. The essential difference lies not in the amount of text, but in the fact that every question a prospect would ask now finds an evidenced answer on the website.
We deliberately give no percentages for the success of these measures. An illustrative example can show which gaps are typical and which measures follow from them. It cannot prove an effect that will occur in your market, in your region and at your point in time. That is exactly why every project starts with a baseline measurement.
A regional traffic-safety company offers site protection, mobile traffic signals and access protection. Online, the company is found mainly by its name and address.
The website no longer only describes which products are available. It explains how the company delivers a concrete project. This allows the company to be classified as a fitting solution for questions about planning, permits, provision, inspection and documentation — even when the original user question did not contain the company name at all.
Traffic safety is a regulated, highly technical and regionally organised market. It is precisely in such markets that the gap between what a company can do and what can be found online about it is largest. The operational competence sits in the heads of the crew leaders and in the folders in the office — not on the website. The effort therefore lies less in the writing than in extracting that knowledge.
Sit down for two hours with an experienced site manager and have them describe the course of a typical project — from enquiry to dismantling. The notes from that conversation are already a better basis for a service page than any agency briefing template.
AI visibility cannot be captured by a single metric. Eight measures that together form a robust picture.
How many prioritised questions are sufficiently answered by your own content?
For how many repeatedly tested questions is your company mentioned?
For how many questions is a page of yours used as a source?
For how many selection questions does your company make the shortlist of suitable providers?
Are core service, target group, location, size, industry focus, hosting and data-protection details, contacts and service limits correct?
How often are competitors mentioned relative to your company?
OpenAI tags referrals from ChatGPT Search with the parameter utm_source=chatgpt.com. This lets you evaluate incoming visits separately.
Contact request, appointment booking, whitepaper download, initial check, newsletter sign-up, request for a quote.
Different studies reach markedly different conclusions on the size of the AI referral channel. An SE Ranking analysis put the share of AI platforms in 2026 at 0.32 per cent of all analysed website traffic; ChatGPT accounted for 74.78 per cent of AI referrals within that. Conductor found an average of 1.08 per cent AI referral traffic across ten industries and a ChatGPT share of 87.4 per cent.
The studies differ in period, website sample, industry mix and measurement method. What matters is a methodological problem that affects all of them: a considerable share of AI traffic carries no referrer and ends up in the analysis as direct traffic. Market averages for AI traffic are therefore systematically lower bounds. For your company, your own web analytics is more meaningful than any market figure — and even that is a lower bound.
The most important limitation of all traffic metrics: when an AI system answers a question with your information and the user does not click, no session and no measurement arises. The influence on the shortlist was nonetheless real. That is why metrics 1 to 6 on this page are not the soft add-ons to hard traffic. They are the part of measurement that captures the actual impact.
Until early 2026, measuring AI visibility was largely manual work. Within a few months, the three big providers made their own data available for the first time. The state of play at the time of writing.
In February 2026, Microsoft launched the public preview of the "AI Performance" report. It shows how often a website's content is used as a source in generative answers in Copilot, Bing and selected partner surfaces. It displays the number of source citations, the cited pages and the so-called grounding queries — those internal search queries the system forms from a user question. In June 2026, four more analyses were added: intents, topics, citation share and a period comparison.
For SMEs, this report is currently the most useful free tool — not because Bing has the greatest reach, but because the grounding queries show the phrasings with which AI systems actually search for your content. These phrasings often differ considerably from what is in your keyword concept.
On 3 June 2026, Google announced its own performance reports for generative search features. They show impressions from AI Overviews, AI Mode and generative features in Discover, broken down by page, country, device and date. The rollout began with a subset of websites in the United Kingdom; data is available from 18 May 2026, with no retroactive history.
No clicks, no click-through rate, no average position and no search queries. You see that you appeared, but not what it was worth. In addition, AI Overviews, AI Mode and Discover are combined into a single view and cannot be separated. For DACH websites, the report was not yet generally available at the time of writing.
At the same time, Google introduced a switch that lets site owners exclude their content from the generative search features. Google has honoured this setting since 17 June 2026 and makes clear that it is not a ranking signal for classic search. Anyone who activates it receives neither impressions nor traffic from the generative features. This decision belongs on the leadership's desk — and for the vast majority of SME providers the answer is: do not exclude.
On 13 May 2026, Google Analytics 4 added a dedicated standard
channel for AI assistants. Sessions from recognised sources
automatically receive the medium ai-assistant and appear
alongside Organic Search and Paid Search in the standard reports. No
configuration is needed; broad availability was reached in early June
2026.
| Limitation | What it means for your reporting |
|---|---|
| Not retroactive | The channel only counts from 13 May 2026. Your earlier AI traffic stays in "Referral". Keep a custom channel group — only that provides the history. |
| Referrer required | Sessions without a referrer land in "Direct". This affects app access, in-app browsers and copied links — a considerable share of AI traffic according to various analyses. |
| Google's own AI missing | Clicks from AI Overviews and AI Mode are counted as Organic Search, not as AI. For many websites this is the largest share — and invisible in the channel. |
| Source list incomplete | Perplexity is not listed in Google's documentation and still falls under "Referral". Check the current list before you report figures. |
A workable dashboard needs five areas. No more — otherwise it will not be maintained.
| Metric | Target direction |
|---|---|
| brand mention rate | rising |
| source rate | rising |
| recommendation rate | rising |
| share in provider comparisons | rising |
| number of correctly represented services | rising |
| Metric | Target direction |
|---|---|
| factually correct answers | as high as possible |
| outdated information | as low as possible |
| incorrect service attributions | zero |
| negative or misleading representation | falling |
| sources with outdated information | falling |
The metrics from Bing Webmaster Tools show source activity, not automatically the position or authority of a page. A high citation value means your content is being used — not that your brand appears in the visible answer.
For context: Similarweb found a conversion rate of 7.1 per cent for ChatGPT referrals in one analysed group of websites — just below Paid Search at 7.8 per cent. This value, too, is dataset-dependent and not a general expected value for your company. It merely shows that the channel is not a gimmick.
The best dashboard is the one that is actually read. Five areas, three to five figures each, a short comment on the change and a list of open items. Anything beyond that is reporting that keeps itself busy.
AI answers are not fully deterministic. The same question can, at different times, use different sources, choose a different order, name a provider or not, be phrased differently and respond to a different location context.
A 2026 academic study therefore explicitly warns against treating visibility as a fixed measurement based on single queries. Repeated measurements showed considerable variation in sources and rankings. Anyone who takes a single answer as proof is measuring noise.
Fixed, documented and left unchanged across quarters. Anyone who adjusts the questions each time cannot measure a trend.
Short question · detailed question · question with industry · question with location · question with company size · question with a constraint or budget.
Five runs for strategically important topics. Same region, same language, documented time, same test conditions. Where possible, use a fresh session with no conversation history and no personalisation enabled — otherwise you are also measuring your own usage behaviour.
Question · platform · date · location context · companies named · sources used · your own mention · type of mention · factual errors · anomalies.
"We were number two yesterday."
"Our company was mentioned in 38 per cent of the tested selection questions. In 22 per cent, one of our pages was used as a source. For regional queries, the mention rate was higher than for national queries."
Do not enter confidential information into public AI systems for visibility tests. This applies to customer data, internal quotes, unpublished prices, personal information and trade secrets. A visibility test works exclusively with questions a prospect would also ask — it needs no inside view.
The plan is deliberately scoped to run alongside day-to-day business. It does not replace a strategy, but it does prevent a year from passing without anything happening.
Tasks
Deliverables
Tasks
Deliverables
Tasks
sameAsDeliverables
Tasks
Deliverables
Tasks
Deliverables
Not a lack of budget and not a lack of know-how, but a lack of ownership. If, after week 8, no one is named as responsible for having a service page professionally reviewed and approved, it stays in draft. Clarify responsibilities in week 1, not in week 9. The next chapter describes how.
For a mid-sized company with three to five core services, the internal effort over 13 weeks is, in our experience, around 15 to 25 person-days — spread across IT, marketing, sales and the specialist departments. The largest single item is not the writing, but extracting expert knowledge from employees' heads. Anyone who underestimates this item produces interchangeable texts.
AI visibility is not a one-off SEO measure. It needs clear responsibilities — otherwise it drifts back and forth between marketing and IT and ends up with no one.
| Role | Contribution |
|---|---|
| Leadership | Strategic goal-setting · approval of positioning · prioritisation of services and target industries · decisions on resources · the fundamental decision on training and search crawlers |
| Marketing | Steering the content architecture · company representation · publication planning · monitoring brand mentions · maintaining the master-data record |
| Sales | Providing real customer questions · objections and selection criteria · competitor information · feedback on the quality of incoming enquiries |
| Specialist departments | Professional review · industry-specific terminology · examples · limits and prerequisites · keeping content up to date |
| IT / web | Crawl access · indexability · technical performance · structured data · web analytics · redirects and canonicals · evaluating server logs |
| Data protection and legal | Involvement for personal references · reviews · customer names · copyrighted content · health-related or legal statements · regulated products · public performance promises |
In almost every company we audit, no one is responsible for the question: "Is what is said about us online still correct?" This role needs no dedicated position. It needs a name, an hour a month and the mandate to initiate corrections.
Not every measure is worth doing straight away. These criteria help direct a limited budget to where it has an effect.
| Business relevance | Current visibility | Priority |
|---|---|---|
| high | low | address immediately |
| high | medium | improve in a targeted way |
| high | high | secure and measure |
| medium | low | after the core services |
| low | low | defer |
There are situations in which any content investment evaporates. Check honestly whether one of these points applies to your company:
If two or more points apply, the right next step is not a content plan but a positioning and clean-up project. That is uncomfortable, but considerably cheaper than twelve months of content on an unclear foundation.
| Trigger | Assessment |
|---|---|
| more than 100 questions checked regularly | Manual tests become uneconomical |
| several countries and languages | Location-dependent answers can hardly be captured cleanly by hand |
| many locations | Regional differences are only comparable when automated |
| several brands | Cannibalisation would otherwise stay invisible |
| monthly competitor data | Historisation requires constant conditions |
| automated historisation desired | Reports become reproducible |
For small and mid-sized companies, a structured prompt set, a documented baseline and a monthly manual or semi-automated test are almost always enough to begin with. The free initial data from Bing Webmaster Tools and GA4 covers a considerable part of the need. Software does not solve a positioning problem — it only makes it more visible and more expensive.
A market for claims has grown up around AI visibility in a short time. Seven of them come up especially often.
There is no commitment from OpenAI regarding a special effect of llms.txt
on ChatGPT Search; the crawler documentation does not mention the file.
In June 2026, Google explicitly clarified its documentation: no
machine-readable special files are needed to appear in Google Search
including the generative features, because Google Search does not use
them. Anyone who wants to maintain the file does themselves no harm — it
has become a convention for coding assistants and browser agents. As a
visibility measure it belongs at the end of the list, not the start.
Mass-generated standard articles create no additional professional substance. In its guide, Google warns against scaled content production designed primarily to influence search systems — and names "chunking" and inauthentic mentions as tactics that can be ignored.
Classic SEO fundamentals help with accessibility, relevance and authority — Google confirms this explicitly for its own generative features. But ChatGPT forms its own search queries, uses partly different sources and synthesises the answer. An identical order is not to be expected.
There is no confirmation of this. For its own AI features, Google even explicitly states that no particular schema.org markup is needed. Structured data supports machine-readable representation — it does not replace content.
Single answers can be random, time-dependent or location-dependent. Only repeated tests across different phrasings yield a robust tendency. See Chapter 24.
A brand can be named as an unsuitable solution, with outdated information, in connection with complaints, with the wrong target group, or as an expensive and limited alternative. Representation quality matters more than the sheer number of mentions.
AI visibility builds on SEO fundamentals and extends them with source eligibility, brand representation, recommendation context and answer quality. Google puts it soberly in its guide: answer and generative search optimisation are, at their core, still search engine optimisation.
AI visibility touches on matters that go beyond marketing. Six areas to keep in view.
AI systems can use older sources. Important pages therefore need a traceable date of last update and a named subject-matter owner. Different profiles lead to incorrect company representations — and these representations remain effective for months, even after you have corrected the source.
Terms such as "leading", "market-leading", "guaranteed" or "completely error-free" should only be used if you can substantiate them. This is not merely a matter of credibility: under German competition law, misleading commercial practices are subject to warning letters. Claims of unique or top status without a robust basis are a real cost risk in Germany — and invented metrics on a service page are exactly that.
These measures do not work — and they are also legally open to challenge. The only reliable effect is reputational damage when they come to light.
For visibility tests, no customer data, internal quotes, unpublished prices, personal information or trade secrets may be entered into public AI systems. Clarify in advance which systems your employees may use for tests and under what conditions. For companies with an AI policy, this point belongs there, not in a separate marketing rule.
References with customer names, personal photos, reviews and quotes require documented consent. The same applies to author profiles. If you use AI systems yourself to create content, the usual requirements for labelling, copyright and editorial responsibility apply — an AI-generated text relieves no one of liability for its content. Data protection under the EU GDPR is not a compliance add-on for us, but an argument: it is one of the questions that prospects actually put to AI systems.
Check regularly: which points of criticism are named? Which sources generate this criticism? Is it factually correct? Has the underlying problem already been fixed? Is current information published? Do customer service, product or communication need improvement?
This whitepaper does not replace legal advice. For assessing specific statements, references and advertising claims, please consult your legal advisers.
To tick off in your next leadership meeting. If you cannot confidently answer more than a third of the points, you know your first task.
If you had to reduce this checklist to three points: does the OAI-SearchBot actually reach our website? Does every one of our profiles online describe the same company? And is there, for each of our five most important services, a page that fully answers a customer question? Anyone who answers these three questions with yes has achieved more than most competitors.
AI visibility is not a short-term marketing technique. It is the result of an understandable, verifiable and consistent digital representation of the company.
Companies do not become visible in ChatGPT because they use a particular phrase especially often. They become visible when enough robust signals come together:
Referral traffic from AI systems is currently much smaller than search-engine traffic in many industries. Even so, the business impact can be considerable, because research, comparison and shortlisting already happen inside the AI system. What you see in your analytics tool is not the impact. It is the small part of the impact that left a click behind.
In retail, current data shows the economic potential especially clearly. For May 2026, Adobe Analytics found that visitors who arrived at US retail sites via AI assistants generated 53 per cent more revenue per visit, a 54 per cent higher conversion rate and 53 per cent longer time on site than other visitors. A year earlier, the same group had still converted at roughly half the rate of other visitors.
The data comes from Adobe Analytics and is published by the vendor alongside its own product for AI visibility. It has not been independently verified. It relates to US online retail and cannot be transferred directly to B2B services in the DACH region. What it shows is not your expected value, but a mechanism: the high degree of pre-qualification that arises from AI-assisted recommendations. Anyone arriving via an assistant already has the comparison behind them.
For SMEs, the decisive question is therefore not:
The decisive question is:
The good news: the work that leads to this answer is not a new discipline. It consists of describing your own company cleanly, honestly and traceably — and making sure that this description is technically accessible and says the same thing across every channel. That is unspectacular. It is the reason so many companies have not yet done it.
How visible is your company in ChatGPT? A structured AI visibility check answers this question with data instead of guesswork.
We start with a documented baseline measurement — not with a proposal. Only once you know where you stand can you decide whether, and to what extent, a measure makes sense.
krambergai.comKrambergAI GmbH, based in Leinfelden-Echterdingen, develops and supports AI solutions for small and mid-sized companies in the German-speaking region: from AI employees and enterprise GPT through Company Brain and AI governance to visibility in AI systems. Data protection under the EU GDPR and "Made in Germany" are not add-on options for us, but the basis of our work.
We build digital products that make work calmer. Relief, control, security and sovereignty matter more to us than speed and hype — on this topic too.
KrambergAI GmbH · https://krambergai.com
All figures as of July 2026. Metrics on AI usage and AI traffic age quickly — please check the respective current version before reusing them.
All information was checked to the best of our knowledge at the editorial deadline in July 2026. Data from tool and software vendors is marked as vendor data, because it is regularly published alongside their own products. We cannot guarantee completeness or lasting accuracy.
Legal notice and contact: KrambergAI GmbH · Leinfelden-Echterdingen · Local Court of Stuttgart HRB 804171 · https://krambergai.com · Full details pursuant to Section 5 of the German Telemedia Act (TMG) and information on data processing under the EU GDPR are available on our website.