Why Your Website Does Not Show Up in ChatGPT
Summary. We ran the same AI-visibility audit on six well-known German brands (m-net, Elli, thinX, carwow, Zeppelin, HDBW) across telecom, e-mobility, automotive, industrial, consulting and higher education. The average score was 36 of 100. Three structural reasons explain it: schema-markup gaps, crawler configuration, and source absence. None of the three requires a new technology stack — they require someone to own the cross-functional problem. The fix is closer to SEO 2004 than to a new model.
Last week we ran the same audit on six well-known German brands across telecom, e-mobility, automotive, industrial, consulting and higher education: m-net, Elli, thinX, carwow, Zeppelin and the HDBW. The average AI Visibility Score was 36 of 100. The best one — carwow — landed at 53. The worst at 26. Not a single one of these brands appears reliably in the answers that ChatGPT, Claude or Perplexity generate about their own category. Carwow does not show up when a user asks for the best way to sell a used car in Germany. Zeppelin does not show up when a procurement lead asks for industrial construction equipment suppliers. The HDBW does not show up when a working professional asks about part-time degrees in Bavaria. The websites exist, they rank in Google, they have decent traffic. They are simply invisible to the new layer of software that increasingly mediates the question "who should I talk to?".
The reaction this usually triggers is one of two: either "AI is hype, real customers still come from Google" or "we will deal with it once it becomes a real channel". Both are reasonable in 2024 and self-defeating in 2026. The shift from search to answer is happening in slow motion, but it has already crossed the threshold where ignoring it is a strategic posture, not a tactical one.
From Search to Answer
Search engines surface a list of links. Answer engines surface an answer. That sounds like a small UI change. It is not. The list-of-links model rewards a site for being one of ten plausible candidates. The answer model rewards a site for being one of one to three brands a model is willing to name out loud, with confidence, and with a citation. The difference between rank ten and rank one was a click-through-rate problem. The difference between cited and uncited is binary. You are in the sentence, or you are not in the conversation.
Robert Solow wrote in 1987 that you could see the computer age everywhere except in the productivity statistics. The pattern has repeated with the internet, with mobile and with SaaS: a new substrate becomes the default before anyone has finished arguing about whether it is real. The substrate today is generative answer engines. ChatGPT crossed 800 million weekly users in early 2026. Perplexity has become the default research interface for an entire cohort of consultants and analysts I work with at Accenture, Telefónica and Volkswagen. Google has folded AI Overviews above the organic results on most commercial intent queries. The question "do AI answers matter for my industry yet" is the wrong question. The right question is: in three years, when half of the buying-stage research starts in a chat interface, what evidence do you want the model to be trained on about your brand?
Three Structural Reasons Sites Are Invisible
The six audits surfaced the same three problems again and again. None of them is new. None of them requires a new technology. All of them require someone to take ten lines on a roadmap and actually ship them.
Schema markup gaps. Modern websites are built for human eyes. They are not built for machine readers. JSON-LD — the structured data format that schema.org standardised over a decade ago — is missing or incomplete on five of the six sites we audited. Carwow has Vehicle and Product schema everywhere because it has to, the rest barely scratch Organization. The Mittelstand obsession with "our website is responsive and modern" has skipped over the part where machines actually parse what the site says. An AI engine that wants to summarise "what does company X do" first reads the structured data. If there is no structured data, the engine guesses from prose, and prose is ambiguous. Schema markup is to AI answer engines what the title tag was to Google in 2004: cheap, well-documented, embarrassingly under-implemented. The same agencies that charge €60.000 for a relaunch quietly drop schema from the spec because no one in the brief asks for it.
Crawler configuration. Half of the brands we audited either block GPTBot and Anthropic explicitly in robots.txt or leave only wildcard rules that allow everything by accident. Neither posture is a decision. The block came from a 2023 panic about "AI training on our content", with no follow-up review when the same companies started complaining six months later that they were not in the answers. The wildcard came from no one ever opening robots.txt. The cleaner posture is somewhere in between: explicit allow rules for crawlers you want to be visible in, explicit block rules for crawlers you do not, and Content-Signal directives that declare intent for AI training, AI inference and search separately. We added Content-Signal support to our own audits two weeks ago after the isitagentready.com test made it obvious that the standard exists and almost no one uses it yet.
Source absence. Even with perfect schema and clean robots, a model can only cite what it has been trained on. Reddit, LinkedIn, Wikipedia, YouTube and a handful of media outlets together account for around 40 % of the domains that ChatGPT, Claude and Perplexity cite (Semrush, October 2025). If your brand has no presence on those domains, the model has nothing to reference even if it wanted to. The German Mittelstand is structurally underrepresented here. There are world-class hidden champions in industrial automation, in family-office wealth management, in B2B SaaS — and they have no Wikipedia article, no Reddit thread, no LinkedIn thought-leadership cadence. The model treats absence as non-existence. That is harsh, but it is the same logic Google applied to backlinks in 2004 and the same logic LinkedIn now applies to your second-degree network. The presence of an entity in the training corpus is the new backlink graph.
What Changes When You Measure
The reason the six audits were eye-opening is not that the scores were low. Everyone in the room had a hunch the scores would be low. The shift happened when the scores became a number with an audit trail behind them. "We are probably invisible" is a discussion that goes nowhere. "Our schema coverage is 11 % and the Anthropic crawler has not seen our pricing page in 90 days because robots.txt blocks it" is a discussion that produces a backlog ticket by the end of the meeting. The same dynamic played out with Core Web Vitals in 2020: nothing changed when "site is slow" was a feeling. Things changed when Lighthouse spat out a number and a list of files. Measurement turns a strategic anxiety into an engineering task, and engineering tasks get done.
This is the moment to recognise that AI visibility is not a marketing initiative, even though marketing will own the conversation initially. It is a technical-content-strategy intersection that nobody currently owns inside most Mittelstand companies. The CMO does not own robots.txt. The CTO does not own which press releases get distributed. The agency does not own which Wikipedia entries exist. The work falls between the chairs, which is why six months after the AI Overviews rollout I still find companies that have never run a single AI-visibility audit on their own homepage.
The fix is not a new technology stack. It is a person who looks at the audit, decides which five fixes matter most, and ships them in a sprint. The fixes themselves are mostly boring: write LocalBusiness schema, unblock GPTBot, publish llms.txt, get the Wikipedia article through review, pitch a single industry analyst, set up a quarterly re-scan. There is no AI magic in the response — only the patience to treat the new layer the way we treated SEO twenty years ago, before there was a vocabulary for it.
How to Start
If you want to see your own number, the easiest path is to run an audit and stare at it for five minutes. We have the same methodology we used on the six brands live at produktentdecker.com/ai-visibility, with the six reference reports linked from the page so you can compare your domain against a known baseline. There is no upsell on that page yet; the goal is to make the audit useful as a standalone artefact and to give the conversation a starting point that is not anecdotal. Two follow-up posts will dig into the technical fixes — a Schema-for-GenAI deep dive and a sober market overview of AI-visibility tools.
The single highest-leverage move I have seen this year is not a new ad campaign or a model fine-tune. It is reading an audit of your own homepage, picking the three findings the engineering team can ship before the next board meeting, and running a re-scan eight weeks later. Six brands, two hours of work each, average score going from 36 to 70. That is the size of the gap. It is also the size of the opportunity.