In Chrome 124 on Mac, Windows, and ChromeOS, Google updated the address bar (Omnibox) with ML models to provide better suggestions.
Previously, Chrome leveraged a “hand-built and tuned set of formulas” that were “difficult to improve or adapt to new scenarios.” For example, one of the signals is “Time since last navigation.”
The expected value of this signal is that the smaller the signal (the more recently you went to a particular URL), the greater the contribution it will make towards a higher relevance score.
Google says its scoring system, which is responsible for displaying/ranking URLs and suggested queries, “has been largely untouched for a long time.”
In most cases, ML-trained scoring models were the obvious path forward. But it took many false starts to finally get here. The reason we haven't been able to tackle this challenge for so long is that it's difficult to replace core mechanisms for functionality that is used literally billions of times every day.
This new ML system will allow the Chrome address bar to return “more accurate and relevant” page suggestions. This will allow Google to “gather newer signals and periodically retrain, evaluate, and deploy new models over time.” One of the improvements the model has made over time since the last navigation is:
…if the time since navigation was very short (seconds instead of hours, days, or weeks), the model was decreasing the relevance score. We found that this training data reflected a pattern where users would visit a URL that wasn't what they actually wanted, then quickly return to the Chrome Omnibox and try again. In this case, the destination URL is almost certainly not what you want, so you should receive a lower relevancy score on your second attempt.
Looking ahead, Google is considering incorporating new signals “such as differentiating time of day to increase relevancy.”
The team is also looking into “training specialized versions of the model for specific environments.” This new approach is currently running on desktop, but could be targeted for mobile, enterprise, and educational use in the future.
More about Chrome:
FTC: We use automated affiliate links that generate income. more.
