Listen: Gemini 4 Hydrogen keeps 45% of its brand recommendations when Google Search is switched on
10,701 brand-recommendation prompts run on Gemini 4 Hydrogen from memory and with Google Search: 45% brand overlap, double the response time, and thought summaries that call live searches "simulated".
Transcript
At DEJAN, we wanted to see exactly how Google Search grounding changes the way artificial intelligence recommends brands. We ran over ten thousand prompts through Gemini 4 Hydrogen twice: once relying purely on its memory, and once with live search turned on.
The results show a massive shift. Only forty-five percent of the recommended brands overlapped between the two methods, and fewer than two percent of the final answers were identical.
Grounding the AI in live search results clearly cleans up outdated information. For example, the AI frequently recommended the brand ESPN BET from memory. But with search turned on, it recognized the brand had been renamed to theScore Bet and updated its recommendation. In fact, out of seventy out-of-date brands the AI recalled, only one actually made it into the final grounded answers.
Of course, there is a trade-off. Search-grounded answers took twice as long to generate, with a median time of nearly seventeen seconds. We also discovered an interesting quirk in the model’s reasoning. In over two thousand cases, the AI wrote about performing a "mental check" or a "simulated search," sometimes adding "if this were a live system," even while it was actively running live searches.
Still, those search results successfully overruled outdated memories, showing just how vital live grounding is for accurate recommendations.
