Empirical analysis across Claude, ChatGPT, Perplexity and Gemini: which brands get cited, which disappear, and why? Including a reproducible probing setup so you can run the same test on your own brand.
What 400 probes across 4 models and 10 brands consistently show.
The Hidden Champions in this study do not disappear from the models — they show up in over 95% of relevant queries. What does differ strongly is how models answer: how deeply they embed a brand, and which sources they pull from.
Visibility index per brand, averaged across all 4 models. 100 = perfectly cited, high mention density, frequently first in lists.
| # | Brand | Type | Visibility index | Citation rate | Density | Best model |
|---|---|---|---|---|---|---|
| 1 | Knauf | B2B |
95
|
100% | 13.6 | ChatGPT (GPT-5) |
| 2 | Stihl | Consumer |
92
|
100% | 11.8 | ChatGPT (GPT-5) |
| 3 | Miele | Consumer |
91
|
100% | 11.7 | Gemini 2.5 Pro |
| 4 | Festo | B2B |
90
|
100% | 12.9 | ChatGPT (GPT-5) |
| 5 | Kärcher | Mixed |
88
|
100% | 8.8 | Gemini 2.5 Pro |
| 6 | Würth | B2B |
87
|
95% | 9.9 | Gemini 2.5 Pro |
| 7 | Sennheiser | Consumer |
87
|
98% | 11.3 | ChatGPT (GPT-5) |
| 8 | Trumpf | B2B |
86
|
98% | 11.1 | ChatGPT (GPT-5) |
| 9 | Liebherr | B2B |
86
|
95% | 10.6 | Gemini 2.5 Pro |
| 10 | Hilti | B2B |
82
|
90% | 11.3 | ChatGPT (GPT-5) |
Same brands, same queries — four very different answer styles.
Citation rate: 98.0% · Mention density: 17.6 · Avg. words: 577
High-volume profiler. GPT-5 writes long-form (577 words on avg.), repeats brand names often (density 17.6) and tends to produce structured top-N lists.
Citation rate: 99.0% · Mention density: 16.2 · Avg. words: 814
Longest output. Gemini 2.5 Pro produces the most verbose answers (814 words, density 16.2). Reasoning tokens must be budgeted generously.
Citation rate: 100.0% · Mention density: 6.5 · Avg. words: 359
Structured and concise. Claude Sonnet 4.5 answers more compactly (359 words, density 6.5) — high hit-rate, less wordy noise.
Citation rate: 93.0% · Mention density: 4.9 · Avg. words: 263
Source-anchored. Perplexity Sonar Pro stays brief (263 words, density 4.9) and often produces inline citation markers — closest to a classic search experience.
The models do not differ primarily in whether they know a brand, but in how they embed it. GPT-5 and Gemini generate verbose, dense profiles (>500 words, high mention density). Claude and Perplexity answer more concisely with a clear focus on sources. If you want to "rank" a brand, you should not just measure "am I mentioned", but "in what style, with what depth, with which sources".
Hypothesis going in: B2B Hidden Champions disappear more than consumer brands. The data says otherwise.
What this means: The common assumption "B2B has a visibility problem in LLMs" does not hold for established Hidden Champions with strong Wikipedia/press footprint. The brands probed here are global market leaders with decades-long PR trails — and that is exactly the trail the models read.
The flip side: brands without Wikipedia entries, without German business-press coverage, without strong industry-media backlinks — those could in fact disappear from the answers. This study cannot test that, because it deliberately targets established brands. The next iteration should explicitly include "mid-tier" Mittelstand companies.
Concrete brand × model combinations with citation rate below 50%. With few hits, more the exception than the rule.
This sample contains no real disappearers — no brand falls below 50% citation rate on any model. That is itself a finding: established Hidden Champions are reliably visible across all 4 major LLMs.
The interesting pattern is not "disappearance" but list position: a brand that lands at position 6 or 7 in a "top providers" list is essentially invisible to the end user. This study captures position rank for list queries — see the CSV for details.
Distribution of cited source clusters across all 400 probes.
Across all models, Wikipedia and the respective company domains dominate as primary sources. German business and industry press is visibly present — international outlets like Reuters/Bloomberg appear noticeably less often than they would for a US-only probe.
Citation rate per query type. Helps you understand which user searches your brand actually competes in.
The full probing setup is open. Clone the repo, add your brand, run it — and compare your brand to the 10 Hidden Champions here.
.env file with the four keys$ git clone https://github.com/craid/geo-citation-study $ cd geo-citation-study && npm install $ cp .env.example .env <- add your keys $ node probe.js --brand=your_brand --smoke <- 1 brand × 4 models × 10 queries $ node analyze.js <- aggregation + CSV export
We build GEO measurement setups for brand teams that want to know how LLMs describe them today — and how that shifts when new sources, new models or new competitors enter the picture.