Haenisch Research studies recommendation behavior across controlled prompts, buyer contexts, providers, and repeated observations. The work distinguishes being mentioned from being recommended and treats an individual AI answer as an observation, not a stable ranking or proof of why a recommendation occurred.

Each published study reports its question, collection conditions, sample, prompts, definitions, exclusions, limitations, data availability, and version history.

Published studies

Study 02 · HRS-2026-002 · Repeated-prompt studyHow Many AI Recommendation Checks Are Enough? We Found No Universal Answer

Across 160 repeated consumer AI responses, more observations produced a clearer recommendation picture. Candidate discovery, inclusion, primary selection, and position stabilized at different rates.

Read the study → Study 01 · HRS-2026-001 · Observational matched-context studyAI Recommendation Sets Shifted Across Buyer Contexts—and One Answer Was Not a Stable Ranking

Across 160 consumer AI responses, recommendation sets and primary selections often differed between matched contexts. Identical prompts also varied substantially, limiting causal interpretation.

Read the study →

Only completed, approved work that meets the publication standard below appears here. Haenisch does not use placeholder publications or present internal measurements as public validation.

What the research measures

  1. Mention or inclusion
  2. Recommendation or selection
  3. Recommendation order
  4. Candidate and vendor distributions
  5. Visible citations and stated evidence
  6. Variation by prompt, provider, buyer context, geography, and collection conditions

How studies are conducted

  1. Freeze questions or prompts when comparability requires it.
  2. Use repeated independent observations when a distribution matters.
  3. Document consumer-surface and environment provenance.
  4. Define outcomes, validity rules, and exclusions explicitly.
  5. Retain complete raw responses or state data availability clearly.
  6. Interpret observational evidence conservatively.
  7. Preserve designs for later remeasurement when comparison is possible.

Haenisch does not infer hidden retrieval, indexing, model routing, training, or reasoning behavior from a consumer answer.

Minimum publication standard

A study enters this index only when collection is complete; provenance and exclusions are preserved; denominators and definitions are explicit; limitations are written; claims are supported; and publication has been separately approved.

Every study reportsWhy it matters
Study ID, version, dates, authors, and canonical URLCreates a durable, citable record.
Exact question, prompts, buyer context, and protocolDefines what was actually tested.
Provider, surface, visible model, environment, and timestampsPreserves collection provenance without inferring hidden behavior.
Planned, valid, and excluded observation countsKeeps denominators and validity decisions visible.
Observed results, supported inferences, and hypotheses as separate layersPrevents observational findings from becoming unsupported causal claims.
Limitations, conflicts, funding, spend, and data availabilityMakes the evidence boundary inspectable.
Corrections, remeasurements, and version historyPreserves the original record as understanding changes.

Versioning rule: Clarifications use a minor version. Material reinterpretations preserve the earlier version and create a major version. Remeasurements receive a dated update page instead of silently overwriting the original study.

Read the full Haenisch methodology and the measurement note Why One ChatGPT Answer Is Not a Ranking.

Research inquiries: ausi@haenischintelligence.com.