Skip to content
StarReview

Updated: 9 May 2026

AI models for Google review replies: methodology summary

13 AI models tested against 53 real Swiss Google reviews in 6 languages. This one-page summary covers the setup, the results, and the criteria.

Key findings

  • 9 of 13 models cleared the rule-based threshold (0 forbidden phrases, 0 em-dashes across 53 reviews).
  • mistral-large disqualified: em-dashes in 42% (22/53) of replies.
  • gpt-5 (Chat Completions) added operational details not present in the review, a real liability for a public reply tool.
  • Latency of compliant models: gpt-4.1-mini (1.2s P50) to gemini-2.5-pro (13.3s P50).
  • We weight safety and grounding above raw speed. The production choice is based on the complete safety, quality, and latency profile.

The 13 models tested

Model Latency P50 Latency P95 Forbidden Em-dash Status
gpt-4.1-mini 1.2s 1.9s 0 0 clean
gpt-4.1 1.3s 3.0s 0 0 clean
claude-haiku-4-5 1.6s 2.0s 0 1 em-dash in DE
mistral-large 1.9s 6.0s 0 22 em-dash spam (42%)
gpt-5-codex 2.3s 4.0s 0 0 clean
gpt-5.2-codex 2.3s 5.1s 0 0 clean
claude-sonnet-4-6 3.1s 4.2s 0 0 clean
claude-opus-4-7 3.3s 5.4s 0 0 clean
gemini-2.5-flash 4.7s 7.0s 1 0 forbidden phrase in DE
o4-mini 5.3s 12.3s 0 0 clean
gemini-3-pro-preview 10.1s 14.7s 0 0 clean
gpt-5 11.6s 22.7s 0 0 fabricated details
gemini-2.5-pro 13.3s 19.9s 0 0 clean

Selection criteria (in this order)

  1. Safety (no fabrication, no admission of liability)
  2. Grounding in the review
  3. Swiss multilingual depth (DE-CH / FR-CH / IT-CH)
  4. Tone fit by industry
  5. Latency
  6. Cost

We do not publicly disclose which model we run in production. Our methodology, criteria, and data are public.

Cite this study

StarReview (2026). Methodology: 13 AI models tested for Google review replies. https://www.starreview.ch/en/methodology/study-summary

Free to use with attribution (CC BY 4.0).

Read the full methodology →