GEO and AI Visibility 2026: Why AI Answers Misrepresent Mid-Market Companies, and the Seven Layers That Change It

How reliably do AI assistants describe companies and facts? The largest independent evaluation up to then, the EBU/BBC study „News Integrity in AI Assistants" (October 2025, over 3,000 evaluated answers from 22 media organizations in 18 countries and 14 languages), found at least one significant issue in 45 percent of all AI answers to news questions; roughly one in five contained major accuracy errors such as hallucinated or outdated information. The Tow Center at Columbia University measured over 60 percent incorrect source attributions across eight AI search engines in 1,600 test queries (March 2025). And in our own GEO audits for DACH mid-market companies, we see the same pattern at the company level: invented managing-director names, wrong founding years, missing mentions in industry recommendations — whenever the brand is not maintained as a consistent, machine-readable entity.

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 markedly. According to a G2 survey of B2B software buyers conducted in March 2026, 51 percent now start their software research with an AI chatbot more often than with Google, up from 29 percent in April 2025. 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: in February 2026 Google AI Overviews appeared on almost every second tracked query. According to a BrightEdge analysis published on 12 February 2026, they appeared on about 48 percent of the queries BrightEdge tracks in February 2026, up from about 31 percent in February 2025; the figure covers a keyword set tracked by the vendor, not all Google searches. So within a single year, the playing field has turned almost halfway into an AI-driven answer space. Where an AI Overview appears, clicks fall: Seer Interactive measured an organic click-through rate of 1.76 percent in June 2024 and 0.61 percent in September 2025 on 3,119 informational search terms from 42 clients, nearly two thirds lower. A broader follow-up from April 2026 covering 53 brands saw the rate rising again, from 1.3 percent in December 2025 to 2.4 percent in February 2026. Anyone not cited in the AI answer sees falling click-through rates in their Search Console whose cause slips right past classic optimization.

Second: AI referrer traffic converts markedly better than classic search. For its benchmark report published on 22 February 2026, the marketing agency Opollo analysed data from 312 IT and technology service firms in North America, Australia and the UK, covering January 2025 to January 2026: on average, 14.2 percent of visitors from AI sources became qualified enquiries, against 2.8 percent of visitors from Google organic search, a rate four to five times higher. A Seer Interactive case study from June 2025 based on a single client found 15.9 percent for ChatGPT, 10.5 for Perplexity, 5 for Claude and 3 for Gemini, against 1.76 percent for Google organic search. The volume is still small: at Opollo, AI sources brought an average of 6.4 percent of traffic in January 2026, and none of these studies provides figures specifically for German mid-sized companies. The quality, however, is disproportionate. Clients who do not yet cleanly capture their AI referrals in pipeline attribution systematically underestimate their actual impact.

Third: brand mentions across the web are more closely linked to AI visibility than backlinks. An Ahrefs analysis of 75,000 brands published in May 2025 found that how often a brand is mentioned across the web correlates far more strongly with its visibility in Google AI Overviews (0.664) than its number of backlinks (0.218). The study covers Google AI Overviews only, not ChatGPT. According to Muck Rack's report “What is AI Reading?” from May 2026, which analyzed more than 25 million links cited by ChatGPT, Claude and Gemini, about 84 percent of citations come from earned media. Muck Rack counts not only press coverage but also Wikipedia, research, government sites, social platforms and other companies' content as earned media. For specific factual questions about a product, however, the systems mostly draw on the brand's own channels, as the previous report from December 2025 already noted. For budget allocation this means: where buyers are still discovering vendors, 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. That is the opposite of what classic SEO taught for three decades.

Who actually gets cited by ChatGPT, AIO and Perplexity

If more than four in five 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.

A Profound analysis published in June 2025, covering 680 million AI citations from August 2024 to June 2025, found the following shares of all ChatGPT citations: Wikipedia 7.8 percent, Reddit 1.8 percent, Forbes and G2 1.1 percent each. Wikipedia was thus cited more than four times as often as Reddit and about seven times as often as Forbes or G2. The mix is shifting, though: in Muck Rack's May 2026 report, ChatGPT and Claude barely cite Reddit any more, only Gemini does; all three use Wikipedia. The reason for Wikipedia's position 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.

According to a Semrush study published in November 2025 (217,000 prompts, data as of October 2025), Reddit was the most cited domain on Perplexity, second on ChatGPT Search and third on Google AI Mode. Notably: 80 percent of the cited Reddit posts had 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. 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. A Profound analysis published in July 2025, based on roughly 80,000 prompts per platform, found that within one month 40 to 60 percent of cited domains were new, depending on the platform: 54 percent for ChatGPT and 59 percent for Google AI Overviews. A SISTRIX analysis from April 2026 measured as much as 74 percent new domains per week for ChatGPT Search, but found that for 53 percent of prompts Google AI Overviews did not change a single source over 17 weeks. Volatility is high, then, but differs greatly by platform. A central adjustment screw is content freshness: in our assessment, sitemap entries without lastmod and schemas without a dateModified property are deprioritized. We internally call this the 3-month citation cliff: in our experience, 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 seven layers of a GEO optimization

The following seven layers systematically eliminate the mistakes described above. Their order is not arbitrary: each layer assumes the one beneath it is in place. Starting at layer 5 means optimizing something the crawlers cannot read. The technical expression varies with the existing stack, the sequence does not.

Layer 1: pre-rendering or server-side rendering for every route. Vincency's own site has run on Astro with statically generated HTML since the end of August 2026 (previously a React application on Vite with a pre-rendering plugin). 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. On dateModified, a warning from our own experience is worth the space: feeding the value from a build-time constant reports every article as changed today on every deploy, including the two-year-old ones. A freshness signal that always fires carries no information, and genuine maintenance becomes indistinguishable from no change at all. Set dateModified per article to the date of the last substantive edit and let it fall back to datePublished otherwise. 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: in a log file experiment by OtterlyAI (February 2026), only 84 of more than 62,100 AI bot requests to a test site over 90 days went to llms.txt, about 0.1 percent. Ahrefs found an llms.txt on 28 percent of the 137,000 domains in its web analytics, by its own account an upper bound, and 97 percent of those files received not a single request in May 2026. Google states explicitly that Google Search ignores the file. We still implement it for every client, for two reasons. First: for AI agents that search a website on purpose, it is the designated entry file. 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 AI crawlers we consider relevant. Cloudflare has offered a toggle that blocks AI crawlers with a single click since July 2024. Since 1 July 2025, Cloudflare has asked every new domain at sign-up whether to allow AI crawlers, and since 15 September 2026 new domains receive presets based on whether the site carries ads: search crawlers stay allowed, and AI training and agent bots are restricted only on ad-funded sites. Anyone who has such a block in place without knowing it, and does not lift it 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?" mirrors the way buyers phrase their questions to AI assistants; „Pricing structure of our voice-AI solution" does not. That makes it easier for the systems to match a section to a question. We have not found a study that puts a fixed multiplier on the resulting AI citations (as of October 2026). 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 saw in that period.

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, in our experience a clean GEO implementation typically pays for itself within three to five months, even calculated conservatively, with the AI traffic volume of 2026, still relatively small but qualitatively superior.

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 pay-per-crawl as a private beta in July 2025 and moved AI Crawl Control to general availability in August 2025. On 1 July 2026 Cloudflare itself called charging per crawl a crude measure of value and announced that it is starting to shape Pay Per Crawl into Pay Per Use as an experiment, with payment for example when content appears in a partner's search results; since 15 September 2026, the presets for new domains restrict AI training and agent bots on ad-supported pages, while search crawlers stay allowed. Mid-market companies therefore have to decide whether to structure access 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. After Google's March 2026 spam and core updates, 79.5 percent of top 3 URLs changed position according to an SE Ranking analysis (100,000 keywords, location New York), compared with 66.8 percent after the December 2025 core update. 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 the version of its Quality Rater Guidelines dated 11 September 2025, 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: according to G2's survey of March 2026, half of B2B software buyers start their research with an AI chatbot more often than with Google. 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 as long as most competitors do not act: every mid-market company that sets up a clean GEO architecture 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 AI answers reproduce your facts correctly or invent them — 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

Sources: all sources are linked where they are used and were retrieved on 15 and 16 October 2026, the Ahrefs study last on 17 October 2026.