Google's Silent Gemini Model Switch Broke Search Grounding for a Month, Leaving B2B Users with Unverified AI Outputs
Key Takeaways
- ▸Google changed gemini-flash-latest to point to gemini-3-flash-preview on January 21 without warning, breaking Google Search grounding functionality for approximately one month
- ▸The API continued returning HTTP 200 responses with valid JSON and no error messages, making the regression invisible to developers until they noticed output quality issues
- ▸PartsplanAI, a B2B electronics marketplace, unknowingly served ungrounded AI-generated part specifications to customers for a month, with potential safety implications for engineers relying on accurate voltage and capacitance ratings
Summary
On January 21, Google quietly updated its gemini-flash-latest API alias to point to gemini-3-flash-preview, a model that does not support Google Search grounding functionality. The change was made without error messages, deprecation notices, or compatibility warnings, causing the API to return seemingly valid responses while silently omitting crucial groundingMetadata fields. The issue affected production systems for approximately one month, from late January through February 27.
The impact was particularly severe for PartsplanAI, a B2B electronic components marketplace that relies on grounding to verify part specifications against real datasheets. Without grounding, the system generated part recommendations and spec searches based purely on the model's pre-trained knowledge, potentially introducing hallucinated data about critical specifications like capacitance values and voltage ratings. The corrupted data accumulated in their database and was served to B2B customers throughout the month-long period.
The developer spent 16 hours debugging with 63 Git commits before discovering the root cause wasn't in their code but in Google's undocumented model change. The fix required switching to gemini-2.5-flash, taking only 20 minutes once the actual problem was identified. An additional undocumented behavior was discovered: on gemini-2.5-flash, setting responseMimeType to 'application/json' while using googleSearch causes the search functionality to be silently ignored. GitHub Issue #384 in the google-gemini/generative-ai-js repository confirms the grounding issue was known in the community before the alias change was implemented, and the change coincided with the Gemini 2.0 Flash general availability announcement.
- An additional undocumented issue exists on gemini-2.5-flash where JSON response mode silently disables search grounding without warning
- The community was aware of the grounding issue before Google made the alias change, as documented in GitHub Issue #384
