Google’s Gemini‑Flash silently breaks Search grounding for a month, sources say.

The impact was felt most acutely by PartsplanAI, a B2B electronic components marketplace that relies on grounding to cross‑check part specifications against official datasheets. According to the platform’s operator, the lack of grounding turned every recommendation and spec search into pure hallucination for roughly a month, from late January through Feb. 27. “Wrong capacitance values or voltage ratings from a language model aren't just embarrassing; they cause real problems for engineers downstream,” the developer wrote, noting that corrupted data accumulated in their database and was served to customers without detection.

When the anomaly surfaced on Feb. 27, the team embarked on a 16‑hour debugging marathon that produced 63 Git commits and 13 distinct attempts to fix the issue. They rewrote prompts, rebuilt the search pipeline, tweaked timeouts, and experimented with parallel versus sequential calls, yet none of these changes resolved the problem because the root cause lay in the model alias itself, not in their code. The eventual fix—switching the API call to gemini‑2.5‑flash—restored grounding within 20 minutes, confirming that the regression was isolated to the gemini‑3‑flash‑preview version.

The changelog entry for Jan. 21 merely states, “gemini‑flash‑latest alias now points to gemini‑3‑flash‑preview,” without any mention of the grounding regression or a compatibility warning. A second undocumented behavior compounds the risk: on gemini‑2.5‑flash, setting `responseMimeType: 'application/json'` while also requesting `googleSearch` silently disables the search component, again without error or documentation. This issue was flagged in GitHub Issue #384 on the google‑gemini/generative‑ai‑js repository, indicating that the community had identified the problem before Google’s alias change was made.

The timing of the alias switch is noteworthy. The Jan. 21 changelog was published the same week Google announced the general availability of Gemini 2.0 Flash, a high‑profile rollout that attracted considerable attention from developers and enterprise users. Yet the grounding regression slipped through the public communication channels, leaving anyone who depended on gemini‑flash‑latest with groundingMetadata vulnerable to silent failure. As the developer cautions, “If you’re using gemini‑flash‑latest with grounding, verify that groundingMetadata is actually present in your responses. You may have been affected since January 21.”

The episode underscores a broader reliability concern for AI services that are increasingly embedded in mission‑critical workflows. While Google’s Gemini family promises speed and cost efficiency, the lack of explicit deprecation signals and the silent omission of essential metadata can erode trust among enterprise adopters. Companies that have built verification layers around grounding now face the prospect of retroactively auditing a month’s worth of data for potential inaccuracies—a costly exercise that could have been avoided with clearer versioning and documentation practices.