Chrome’s address bar has become a powerful tool that isn’t just a place to type in website URLs. Google calls it the ‘omnibox’ because it’s also a search field, and it lets you perform a bunch of other tasks as well. to become a lot smarter thanks to machine learning and to better understand what you are looking for.
Chrome’s omnibox will be equipped to provide more accurate and relevant suggestions when you use Chrome, and as you use Chrome over time, the AI models behind it should improve your search suggestions thanks to the improved ‘relevance score’.
Announcing the new opportunity in a post on the Google Chromium blogJustin Donnelly, head of Chrome omnibox engineering, said he surveyed colleagues asking for ways to improve the omnibox, and “The number one answer I heard was ‘improve the scoring system.'” According to XDA Developersthis scoring system is how the omnibox interprets what the user is searching for based on the typed input.
The post also explains that this improved capability will apply to Chrome on Windows, macOS, and ChromeOS.
From static to adaptive scoring models
Donnelly added that the omnibox’s scoring system already worked quite well, but was apparently quite inflexible and static, as it was governed by “a series of hand-built and hand-tuned formulas.” These worked well for a wide range of inputs, but were not easy to improve or adapt in new scenarios.
He said the engineering team responsible for the innovation had been working for some time on a machine learning-powered scoring model that would be more sensitive to various metrics (such as the last time you visited a website), a process that took some time, partly due to the sheer number of searches that occur every day. Now it looks like the improved models are ready to be rolled out.
The team found that the less you visit a particular website, the less often the omnibox returns that site as a suggestion when processing your searches. It also discovered something even more interesting: When a user spent a short amount of time navigating a specific web page, the new model also lowered that page’s relevance score.
The model’s training data revealed a pattern of user behavior where they opened a page, realized it wasn’t what they were looking for, and went back to the omnibox to look for something else. Donnelly said the team wanted to incorporate this finding into their model to lower the relevance score of the first result, and if it weren’t for the model’s new machine learning capabilities, this feature could have been missed as a useful addition.
Onwards and upwards towards more personalized and responsive browsing
It seems that Chromium’s engineering team took their mission quite seriously, with Donnelly claiming that his team was “driven by a genuine belief in the impact of getting this right for our users.” The results of their efforts appear to have encouraged the team to continue working in this direction and explore more specialized search model versions for specific environments.
Donnelly concludes by saying that as part of this ongoing process, the team will observe how users’ interactions with Chrome’s omnibox change over time so it can get a better idea of how it can maintain its relevance score improve. The new model will also enable the team to collect more time-sensitive signals of user activity and then retrain, reevaluate and deploy improved models in the future.
Overall, this sounds like a positive and exciting development, which could deliver a more intuitive and efficient browsing experience. Chrome’s omnibox will get better at knowing your habits and understanding what you want, and it will also get better at knowing what you don’t want. This new functionality is expected to arrive with Chrome update M124.
That said, you’ll probably have to live with handing over even more data to Google regarding your moment-to-moment online habits. If you can, and if you trust Google to handle it responsibly, then you can look forward to what sounds like a well-thought-out and innovative feature.