DACH AI Source Benchmark 2026

Where the models ground: the most frequent third-party sources

When a model hands out a link, it often points not to the brand itself but to a third page that evidences the claim. From GPT-5 Mini's citation URLs (Gemini delivers masked redirects and is excluded here) three types of grounding source stand out. Showing up there raises your chance of being cited alongside.

The lever differs by type. In regulated fields (mortgage advice, YMYL) the engine cites official registers and legal sources — here a correct, findable entry in the relevant register counts for more than any landing page. For local services, comparison and directory portals dominate: a complete, consistent profile there is the direct route to the link. In SaaS and B2B the model grounds on vendor and integration documentation — a citable docs, integration or support page beats the homepage.

Frequency as a citation source across all GPT-5 Mini runs · panel-owned domains excluded

The zeros: why well-known brands reach 0 % citation

Seven of the 90 providers scored exactly 0 % citation. The obvious guess — “they block the AI crawlers” — holds in only a minority of cases. I checked the robots.txt and reachability of the seven domains on 28 July 2026; the causes split into two groups.

Technical cause: access or entity missing

• box-room.at (Boxroom Vienna) — the domain was unreachable (no DNS response). What the engine cannot load, it cannot cite.
• brother.at (Domino Printing, AT) — the site answers automated fetches with HTTP 403. A server that turns crawlers away supplies no citable evidence.
• waldhuusdavos.ch (Hotel Waldhuus Davos) — the brand domain redirects to an umbrella resort site. Without a standalone, answering page there is no anchor for a citation.

Not a block, but missing anchoring

For the other four the robots.txt is open — the crawler would be allowed, but finds no reason to link:

• Mammut (mammut.com, hiking boots DE) — robots.txt open, 6.7 % mention. A category-fit problem: Mammut stands for mountain sport broadly; for hiking boots the specialist brands win the evidence.
• weclapp (weclapp.com, CRM CH) — robots.txt open, the same domain is cited 21.1 % in Germany. In Switzerland bexio holds the slot; a market-anchoring, not a technical, problem.
• FW Finanzarchitektur (finanzierer.at) and Lagerplatz.de (lagerplatz.de) — reachable, but not present in the registers, directories and map data the engine grounds from for these questions.

Lesson: an open robots.txt is necessary but not sufficient. First the page must be technically reachable and indexable (no 403, no dead redirect, no JS-only rendering without server HTML); only then does anchoring in the grounding data layer decide the link.

robots.txt and reachability check · as of 28 July 2026

Audit prompts to take with you

The same nine buying stages that underlie the panel, as a copy template. Replace the placeholders in square brackets with your category, market and criteria and run the prompt brand-open (without naming your own brand) in the playground of your choice — ideally with web search enabled and, for more stable output, a low temperature. The value on the right is the average citation density per run from this baseline: comparison and evaluation questions hand out the most links.

• Discovery 13.2 cites/run Copy

  Which [category] brands and providers are relevant for [use case] in [market]?

• Task-solving 13.7 cites/run Copy

  Which [product] suit [specific requirement]? Consider providers, rules and sources for [market].

• Evaluation 18.8 cites/run Copy

  Which [product] up to [budget] suit [target group] and meet [criteria]? Consider providers, rules and sources for [market].

• Comparison 18.7 cites/run Copy

  Compare the three most suitable [product/solution] for [use case] by [criterion 1], [criterion 2] and [criterion 3]. Consider providers, rules and sources for [market].

• Alternatives 16.5 cites/run Copy

  Which alternatives to [established solution] offer comparable performance for [requirement]? Consider providers, rules and sources for [market].

• Shortlist 11.5 cites/run Copy

  Which five [product/provider] belong on the shortlist for [use case] in [market]?

• Validation 17.1 cites/run Copy

  For which recommended [product/provider] are [quality attributes] backed by independent sources? Consider providers, rules and sources for [market].

• Transaction 13.1 cites/run Copy

  Where can I obtain, book or buy [product/service] in [market] on transparent terms?

• Customer support 13.6 cites/run Copy

  Which providers offer [service, support, spare parts] for [product]? Consider providers, rules and sources for [market].

Patterns derived from panel_v1.json · citation density per stage from wave 1

Methodology

Prompt population

180 standardized commercial search scenarios — 6 categories × 3 markets × 10 prompts, frozen before collection.

Run index

3 independent runs per prompt-engine combination with web search enabled. Only runs with at least one search call and one HTTPS citation count as valid observations.

Surfaces

google/gemini-3.5-flash and -flash-lite with Vertex Search Grounding, openai/gpt-5-mini with OpenRouter web search (Exa). Model docs: OpenAI, Google. Routing via OpenRouter with zero data retention.

Localization

All queries localized to DE, AT or CH; domains counted separately per market.

Reproducibility

Collection date

27 July 2026

Sample

n = 1,620 (180 scenarios × 3 engines × 3 runs)

Gateway

OpenRouter Chat Completions (/api/v1/chat/completions), roster v1.2

Models

openai/gpt-5-mini, google/gemini-3.5-flash, google/gemini-3.5-flash-lite

Tool contract

openrouter:web_search (engine = native), max. 5 search calls, max_tokens = 4000

Sampling

Temperature and top-p not set — provider defaults. Runs are therefore not bit-deterministic; every figure is a mean over 90 runs.

Code & panel

github.com/eullr/dach-ki-quellenbenchmark — collector, panel, aggregation (Node.js, CC BY 4.0)

Raw data

JSON · CSV · run log

Independence & tools

Run with an own Node.js batch collector over the OpenRouter API (Chat Completions, native web search per provider); aggregation via an own script. No Peec.ai and no other third-party visibility tool fed into this open study.

Conflict of interest: at the collection date there was no current or prior client relationship with any tested provider in any of the six categories (hiking boots, CRM, industrial marking, self-storage, mortgage advice, family hotels). No brand (e.g. LOWA, MyPlace) commissioned, funded or influenced the study. The open panel stays strictly separate from client data and case studies.

Limits of this study

1. A snapshot: wave 1 measures one collection day, not a time series. Statements on stability follow only with wave 2.
2. What is measured is visible attribution in the answer — not hidden retrieval and not training origin.
3. The mention rate counts exact name matches. Shortened company names lower it, especially for local providers.
4. 90 providers per category-market combination are a deliberately small, curated panel — not a full census of the market.
5. Open panel, not a client sample. Results from client work are published separately as a research note.

Open data

• Results as CSV — 90 providers, mention & citation global and per model wave_1_results.csv
• Panel definition — 180 frozen prompts, 90 providers panel_v1.json
• Aggregated results — mentions, citations, per model wave_1_results.json
• Run log — all 1,620 runs, incl. failures wave_1_summary.json
• Code & panel on GitHub — collector pipeline, frozen panel, aggregation github.com/eullr

Read the wave 1 findings on the blog Back to the research overview