What you need to know
- Google is now using machine learning in Chrome’s address bar to provide better, more personalized results.
- The company says it is replacing hand-tuned formulas that were difficult to change with malleable ML models that can be quickly adjusted and refined over time.
- The change is available in the latest version of Chrome for Desktop (version M124).
Google is making a hidden change in the latest version of Chrome that aims to improve the suggested web page results displayed in the address bar, also known as the omnibox.
In Chrome 124, these suggestions are now made using machine learning models, which the company says replace “handcrafted and tuned formulas.” Since the address bar is now based on ML models, the results should be more accurate and personalized for each user.
Justin Donnelly, a Chrome engineering lead who works on the Omnibox, explains in a blog post that the old rating system could not be adjusted or changed over time. The engineer described it as “inflexible” and due to the lack of flexibility, “the points system remained largely untouched for a long time.” When looking for ways to improve the address bar and its suggestions, the Chrome team saw machine learning as the obvious solution.
ML models can often identify trends and insights that go beyond the human eye, and that was the case with Omnibox’s models. One noticeable change in the behavior of the address bar due to the move to ML is a shift in the way the “time since navigation” signal is perceived. Previously, the manual formula gave recently accessed URLs a higher relevance score. However, the ML models found that this was actually not what users were looking for.
“It turned out that the training data reflected a pattern where users sometimes navigate to a URL that wasn’t what they really wanted and then immediately go back to the Chrome Omnibox and try again,” explains Donnelly. “In this case, it is almost certainly the URL they just navigated to not what they want, so it should get a low relevance score on this second attempt.
In addition to changing the way it scores results for relevance, Google will use ML models in the address bar to make webpage suggestions “more precise and relevant to you.” Your surfing habits and other data collected by Google are probably used to optimally adapt the behavior of the Omnibox to your needs. In other words, the way people use the Chrome address bar can be used over time to retrain the ML models that power it.
The new address bar is included in Chrome 124 for desktop, but visually you won’t notice any differences. In the future, Google would like to add other signals that are included in the relevance rating, such as the time of day and the environment.