GEO and AI Visibility May 2026: Why 96 Percent of DACH Mid-Market Companies Are Hallucinated in ChatGPT — and the Stack That Changes It

In spring 2026, maxonline published an empirical sample of 150 mid-market companies from Germany, Austria and Switzerland that has been quoted in nearly every strategy conversation in the GEO community ever since. The study systematically confronted ChatGPT with factual questions about these companies. The result: 96 percent of the managing-director names it gave were invented. 78 percent of the founding years were wrong. 68 percent of the employee counts were inaccurate. 56 percent of the tested companies were not mentioned at all in industry-specific recommendations. For only 3 percent of the firms was the ChatGPT answer fully correct.

These numbers are no longer a fringe topic in the DACH mid-market. They are the real visibility problem of 2026. While in summer 2024 most mid-market managing directors still treated „AI search" as an optional future trend, buyer behavior has shifted fundamentally within 18 months. According to a multi-source evaluation from Q1 2026, 73 percent of all B2B buyers research their vendors primarily in ChatGPT, Perplexity, Google AI Overviews or Gemini before they even start a classic Google search. Anyone who does not appear, or appears incorrectly, in this first research layer is effectively out of the shortlist game.

This article summarizes what has proven robust in Vincency client projects since our last GEO audit in early May 2026. It is the technical answer to the question: „We are either absent or wrong in ChatGPT — what does that concretely cost us, and which measures change it within two quarters?"

What shifted fundamentally between January 2025 and May 2026

Three shifts define the current state of Generative Engine Optimization. Anyone planning their SEO investments in May 2026 without these three points is optimizing a playing field that no longer exists in that form.

First: Google AI Overviews appear on 48 percent of all searches — and on B2B technology queries even on 82 percent. In February 2025, AIO coverage was still at 31 percent. So within a single year, the playing field has turned almost halfway into an AI-driven answer space. At the same time, the organic click-through rate on queries with AIO has collapsed by 61 percent — from an average of 1.76 percent to 0.61 percent. Anyone not cited in the AI answer sees a continuous erosion in their Search Console whose cause slips right past classic optimization.

Second: AI referrer traffic converts dramatically better than classic search. A widely noted Seer Interactive analysis of B2B clients from Q1 2026 shows: ChatGPT converts at 15.9 percent, Perplexity at 10.5 percent, Claude at 5 percent, Gemini at 3 percent — versus Google Organic at 1.76 percent. A consolidated evaluation across 312 B2B firms arrives at an average 4- to 5-fold conversion advantage of AI traffic over Organic. The volume is still small — currently between 4 and 9 percent of total traffic at a typical B2B mid-market company — but the quality is disproportionate. Clients who do not yet cleanly capture their AI referrals in pipeline attribution systematically underestimate their actual impact.

Third: brand mention frequency is by far the strongest citation predictor. An evaluation of the key AIO and ChatGPT citation studies from Q1 2026 shows: the correlation between brand mention frequency on external authoritative sources and AI citation rate is 0.664. Classic backlinks correlate only at 0.218 — roughly three times weaker. 82 percent of all AI citations come from earned media, not from one's own owned content. This has far-reaching consequences for budget allocation: a mid-market company that is well known and cited in its industry, even with a mediocre website, beats an unknown competitor with a perfect website in AI visibility. That is the opposite of what classic SEO taught for three decades.

Who actually gets cited by ChatGPT, AIO and Perplexity

If 82 percent of AI citations come from earned media, the next question is: from which sources exactly. This is where a central lever lies, because many mid-market companies aligned their PR activities in 2024 to the wrong priorities — based on the classic understanding of authority, which only partly holds in the AI era.

The robust citation shares from ChatGPT answers in May 2026 break down as follows: Wikipedia 7.8 percent, Reddit 1.8 percent, Forbes 1.1 percent, G2 1.1 percent. Wikipedia is therefore roughly seven times as important as all other individually measured sources. The reason is structural: Wikipedia entries serve exactly the architecture that retrieval-augmented-generation systems favor — entity anchors, dense internal linking, sameAs linkage to Wikidata, neutral tone, a strong trust signal. A Wikipedia mention acts in the AI citation graph like a double-weighted backlink from 2010.

Reddit climbed to second position between mid-2025 and early 2026. Notably: according to a Semrush evaluation, 80 percent of the Reddit threads cited by AI have fewer than 20 upvotes. This means it is not viral threads that get cited, but threads with high topical relevance and semantic clarity. A subjective first-hand report in an industry-specific subreddit can place a mid-market company more strongly in the AI citation graph than a professionally produced press text in an online magazine.

For the DACH mid-market this means, in practical terms: the classic PR strategy of the 2010s — press release to Handelsblatt, FAZ, industry bodies — remains important for trust signals, but is no longer a primary GEO lever. What truly counts is combined visibility across three layers: a precise, regularly maintained Wikipedia entry for the company and, where applicable, the managing directors; continuous participation in industry-specific Reddit, Quora and Stack Exchange communities with named employees; and structured data on one's own domain that links these entities via schema.org sameAs to the external authority sources.

The three structural mistakes in almost every mid-market site in May 2026

Over the last six months we have carried out 47 GEO audits for German mid-market companies — private practices, premium real-estate agents, M&A law firms, industrial special-purpose machine builders. In more than 90 percent of these audits, exactly the same three structural mistakes showed up, measurably slowing down AI visibility.

Mistake 1: empty or incorrect sameAs arrays in Person and Organization schemas. When an LLM is asked, „Which implantologist in Stuttgart is reputable?", it first attempts entity disambiguation. „Dr. Markus Müller, implantologist" exists multiple times in Germany. Without a clear sameAs linkage to LinkedIn, Wikipedia, Wikidata, one's own organization domain, and possibly PubMed or professional registers, the model cannot uniquely assign the person and, in case of doubt, favors those with a clear entity signal. On 80 percent of the audited sites we see either empty sameAs arrays or hard-coded placeholders like „#" in the UI links. The consequence is binary: either the entity is identifiable in the knowledge graph, or it is not. A half match costs the full citation.

Mistake 2: FAQ and speakable content only in the JSON-LD, not in the visible DOM. A whole generation of React, Next.js and similar JavaScript frameworks renders FAQ sections only after client-side hydration from React state. In the pre-rendered HTML, often only the questions are visible as buttons, with the answers appearing only after JS execution. Crawlers like Bytespider, or Perplexity users that parse raw HTML instead of the JavaScript-rendered DOM, therefore see questions without answers. Google AI Overviews, too, shows a test bias toward answer paragraphs that sit directly beneath the questions in the SSR HTML, over pure schema content. The fix is trivial — render answers as visible p elements directly beneath each question, with schema remaining as an additional layer. But the leverage is considerable: in our client projects we have seen AIO coverage increases of 30 to 50 percent from this single correction alone.

Mistake 3: missing dateModified at the schema level and in the sitemap. An evaluation in April 2026 showed that between 40 and 60 percent of the sources ChatGPT and AIO cite change every month. That is very high volatility. A central adjustment screw is content freshness: sitemap entries without lastmod and schemas without a dateModified property are consistently deprioritized. We internally call this the 3-month citation cliff: content that has been updated neither in its visible date nor in its machine-readable metadata for twelve weeks measurably drops in AI citation rate. Most mid-market sites do not maintain their content on the substance — not because that would be hard, but because no one has brought the date-maintenance detail in the build process up to the ISO level.

The Vincency GEO stack May 2026

In a Vincency implementation we currently work with a seven-layer stack that systematically eliminates the mistakes mentioned above. The layering is constant — the respective technical expression varies by client.

Layer 1: pre-rendering or server-side rendering for every route. Vincency's own site runs on Vite with build-time pre-rendering. For clients we use, depending on the stack, Next.js with the App Router or Astro with an islands architecture. The decisive point is: every single route must already deliver the full content in the HTML source before JavaScript is executed.

Layer 2: a complete schema layer with dateModified, sameAs and nested entities. Organization, LocalBusiness, WebSite, Person (for managing directors and key employees), Service, FAQPage, HowTo, Article, BreadcrumbList — all as JSON-LD, cleanly nested via @id references. dateModified on every entity, injected from a build-time constant, ISO-8601, not hard-coded. sameAs to LinkedIn, Wikipedia (where it exists), Wikidata, industry registers, and XING where relevant.

Layer 3: llms.txt and llms-full.txt as a curated Markdown bridge. Here one has to be honest: a Search Engine Land analysis from April 2026 shows that only about 0.1 percent of AI crawler requests fetch llms.txt at all. We still implement the file for every client — for two reasons. First: the adoption rate is currently below 4 percent of all domains, so whoever has it is explicitly ahead of the curve, which with emerging crawlers like OAI-SearchBot or Claude-Web creates visibility priority. Second: the file forces a disciplined content audit across the entire domain that has strategic value independent of crawler consumption.

Layer 4: robots.txt with an explicit allow for the 18 relevant AI crawlers. By default, many modern hosting setups — especially anything running over Cloudflare — block AI crawlers following the one-click toggle that has been rolling out since mid-2024. As of May 2026, „Block on all pages" is the default for newly created Cloudflare domains. Anyone who does not actively switch this off via an allow directive for GPTBot, ClaudeBot, PerplexityBot, OAI-SearchBot, Google-Extended, Bytespider, Meta-ExternalAgent, Amazonbot, ChatGPT-User, Perplexity-User and roughly ten others becomes invisible — even if their own content is perfectly structured.

Layer 5: entity linkage to external authority sources. Wikipedia and Wikidata are the central levers here. For every client whose market position justifies it, we check the Wikipedia notability criteria and support — where permissible — building a well-founded Wikipedia and Wikidata entry. In parallel, structured LinkedIn profiles with consistent data points to one's own domain, industry directories, and where applicable field-specific databases (for physicians, for example, the German Medical Association's search; for lawyers, the official lawyer directory).

Layer 6: question-oriented H2 headlines in the visible content. An H2 that reads „What does an AI phone agent cost for a private practice?" gets cited in ten times more AI answers than „Pricing structure of our voice-AI solution" — with identical follow-on content. This is one of the highest-ROI single measures of 2026: phrase every H2 as a conversational question that a real buyer would type into ChatGPT or Perplexity. On a typical B2B page with 25 H2 headings, this single change produces measurable citation lifts in our client projects within eight weeks.

Layer 7: a monthly citation-monitoring loop. Clients who do not measure what AI platforms say about them are flying blind. For every DOMINATE and CROWN client we set up a weekly test set of 20 to 40 industry-typical queries that is queried automatically via the respective platform APIs (ChatGPT, Perplexity, Gemini, Claude). The output is a dashboard with citation frequency, fact-check rate, average position rank and the share of hallucinated answers. The data feeds back into a quarterly GEO review.

What a GEO optimization realistically delivers in the first year

Two concrete client cases from the first five months of 2026, rounded and anonymized, but methodologically representative of what we currently see.

Client A is an M&A boutique law firm from Frankfurt with 14 employees. Starting position in January 2026: against a baseline defined by Vincency of 30 industry-typical queries („best M&A firm for mid-market sale", „business law firm in Frankfurt with a tech focus" and the like), the firm appeared in 2 of 30 ChatGPT answers and in 0 of 30 AIO answers. Seven hallucinated answers (client examples that never existed, partner names that do not exist). Investment in a 4-month GEO implementation: 14,000 euros setup, 1,400 euros monthly retainer. As of May 2026: 19 of 30 ChatGPT answers contain the firm, 11 of 30 AIO answers. The hallucination rate dropped to 2 of 30. In the same period, the AI referrer share of website traffic rose from 1.8 to 7.4 percent, and 4 of the 9 new mandates in Q1/Q2 explicitly named an AI platform as the initial recommendation source in the first-contact question.

Client B is an aesthetic-medical clinic from Munich focused on non-invasive procedures. Starting position in February 2026: AIO coverage on local queries („best clinic for hyaluronic acid in Munich" and comparable) at 0 of 25 queries. No Wikipedia entry, the managing-director person not uniquely identifiable in ChatGPT (name identical to three other physicians). Investment: 11,000 euros setup, 1,100 euros retainer. As of May 2026: AIO coverage on 9 of 25 queries, ChatGPT mention on 17 of 25. A Wikipedia entry for the clinic built (notability demonstrated via public study participation), Person schema with sameAs to LinkedIn, the German Medical Association directory and Wikidata. First inquiries via AI referrer in April 2026: 23, versus 4 in January.

For both clients, the monthly first-inquiry value lies well above the retainer. In industries with mandate or treatment values from four to five figures per conversion, a clean GEO implementation typically pays for itself within three to five months — even calculated conservatively, with today's still relatively small but qualitatively superior AI traffic volume.

What becomes foreseeably relevant in the second half of 2026

We consider three developments very likely by the end of 2026. Anyone building their GEO infrastructure now should account for these three movements in the architecture.

First: pay-per-crawl models will go mainstream. Cloudflare launched its pay-per-crawl system as a preview in early 2025, with general availability announced for Q3 2026. Within 12 months, mid-market companies will have to decide whether to structure the cost per crawl for Anthropic, OpenAI and Google as a revenue source or accept it as a cost item. Anyone who does not think this through will, in case of doubt, switch to block and disappear entirely from the AI index.

Second: the volatility of AI citations will become more structural. March 2026 was, with 79.5 percent of top-3 positions shifted, the most volatile Google core update of all time. We assume this volatility will remain, because RAG systems are inherently less stable than deterministic ranking algorithms. The consequence for mid-market companies: a one-off GEO investment without a continuous monitoring loop will inevitably erode. GEO is not a project with an endpoint, but an ongoing operational process comparable to conversion optimization in the late 2010s.

Third: E-E-A-T will act more strongly as a citation filter. In its updated Quality Rater Guidelines 2026, Google reaffirmed once again that trust is the most important of the four pillars — more important than experience, expertise or authority alone. For AI search engines that base their answers on trust sources, this means: content with clearly named authors backed by verifiable credentials is increasingly favored. Mid-market companies that published their content up to 2025 without author bylines, without Person schema and without documented author expertise will progressively fall out of the citation pool in 2026/27. This is no longer an optional best practice, but a necessary hygiene factor.

Conclusion

As of May 2026, GEO is no longer a niche discipline for DACH mid-market companies, nor is it an „SEO trend". It is the operational answer to a measurable shift in buyer behavior — 73 percent of B2B buyers begin their vendor research in an AI interface, not in classic search. The structural levers are known: pre-rendering, a complete schema layer, sameAs linkage to Wikipedia and LinkedIn, question-oriented headlines, continuous citation monitoring. The leverage of these measures is large, because most competitors will not yet act over the next twelve months — and so every mid-market company that sets up a clean GEO architecture now builds a measurable lead in the citation graph of the relevant AI platforms.

If you want to know how your company currently appears in ChatGPT, Perplexity and Google AI Overviews — and whether you are among the 96 percent whose managing-director names are invented — we are happy to run a compact baseline analysis. 30 industry-typical queries, an honest result, with no sales pressure.

Frequently asked questions about GEO and AI visibility