AußenBlick GKV · 2026 study
Which public health-insurance texts should expert teams review first?
The study presents a reproducible path from 56,198 public pages across 84 German SHI website entities to passages that merit expert review first. It separates linguistic production patterns from substantive clarification needs; neither signal, nor their combination, proves AI provenance, inaccuracy or missing editorial review. No insured-person data, no ranking, no allegation.
Data status May 2026 · The process counts mark review stages, not defects and not a ranking.
Chapter 01 · Public corpus
Many pages do not yet make a governable corpus.
The first observation is scale. Benefit pages, explainers, FAQ pages, campaign pages and older content states sit next to one another. To the public, they can read like one current information offer.
For perspective: even at only five minutes per page, the work would require hundreds of working days.
For expert teams, the governance question is narrower: which public statements are still supportable, and which should be reviewed first?
- Search space
- 56,198 public pages
- Units
- 84 public web properties and sub-portals
- Scope
- publicly available content, no member portal data and no insured-person data
Chapter 02 · Screening points
Scale becomes passages that can be read.
The pre-review works in stages: methodological screening narrows the candidates, then expert-oriented review confirms or excludes them against fixed selection criteria.
A screening point does not mean that a page is wrong. It means that the passage is worth a closer look.
The following numbers describe two record-level process states: evidence-checked review records in the working set and the records routed from it to case review. Neither is a page or defect count.
- Case review
- 21,452 review records; workload, not a page or defect count
- Working set
- 35,998 evidence-checked review records; not a finding count
- Concentration
- transparency, legal framing, medical content, contradictions and linguistic production patterns
Chapter 03 · Evidence trail
A review point counts only when the trail can be bounded.
The decisive step is not discovery; it is limitation. Each robust review point needs a source URL, an original quotation, a retrieval date, a reference trail and explicit statement boundaries.
This keeps visible what the public page actually says and what only the insurer can decide internally.
- Evidence logic
- Source URL · original quote · retrieval date · reference trail · statement boundary
- Separation
- Relevance and evidentiary support are checked separately
- Boundary
- Unsupported screening points are not passed on as review points
Fictional example, freely invented. It shows the method, not a real insurer.
Health information · unit and dosage
- Source
- Example insurer, public guide "Nutrition during pregnancy" (fictional)
- Original quote
- "During pregnancy, we recommend 400 mg of folic acid per day."
- Recorded
- 18 May 2026, archived with retrieval date and reference trail
- Reference
- DGE reference value: 400 µg of folic acid per day in addition to diet when planning pregnancy and in early pregnancy.
- Review question
- Should the publicly visible statement be confirmed or corrected internally? The unit "mg" exceeds the reference value used for comparison (400 µg, 0.4 mg) per day by a factor of a thousand.
- Ownership
- Medical editorial team, expert department
- Statement boundary
- Only the public wording and the stated unit are assessed. No medical recommendation and no statement about the insurer's actual advisory practice.
Select a node of the evidence chain (01–05) or the checkpoint.
Chapter 04 · Patterns
Patterns matter more than isolated cases.
Isolated cases are rarely the strongest observation. The content-governance question becomes clearer when similar types of ambiguity appear across multiple public web properties: advertising-law sensitive wording, ambiguous benefit promises, outdated benefit states or missing editorial traces.
That shifts the frame from single corrections to planned maintenance: does a pattern appear only occasionally, or across several properties?
- Patterns
- Recurring clarification types across multiple web properties
- Priority
- Evidence strength and substantive relevance come before reach
- Guiding question
- Which patterns should expert teams clarify first?
Chapter 05 · Governance
Review responsibility matters more than origin debates.
Whether a text came from a person, an agency or a model does not settle the governance question. What matters is whether public health information is versioned, referenced, approved and regularly updated.
Editorial and expert responsibility makes visible which public statements need to remain current, supportable and bounded.
There is also a new channel: AI search systems now quote insurers' public pages verbatim in their answers. An outdated benefit statement travels into channels that no editorial team can reach.
- Who is responsible for the statement as expert owner?
- Which reference is authoritative?
- When was it last reviewed?
- What happens when benefits, evidence or legal framing changes?
Next step
Confidential method review.
A confidential method review. In 30 minutes, we can walk through the screening logic, selection criteria and five example review points.
The academic write-up is published as an open-access study, aggregated and anonymized: Read the study on Zenodo ↗
Direct contact: [email protected]
Request a method reviewMethod note: the animated dots are illustrative. The numbers describe the working set of the pre-review; only evidence support and statement boundaries turn a screening point into a robust review point.