Google’s AI search hands-on: It’s still not clear why this exists
Google’s Search Generative Experience (SGE), is still a Google Labs feature, meaning parent company Alphabet, Inc. isn’t yet claiming that this is definitely the future of its flagship product.
That’s probably wise, because it would herald a shaky future indeed for humanity’s dominant source of online information. It’s fun, and frequently intriguing, but as I was testing it these past two days, a bemused little voice kept reminding me that Google processes billions of queries per day. Given that we’re talking about a product used by almost everyone, often in life-or-death situations, this new version is much too quirky for prime time.
But it’s also not without its charms! Here’s what using it is like:
SGE competes with Featured Snippets for space on your screen
When performing open-ended, informational queries, Google’s Search Generative Experience feels like a slightly longer version of Google’s Featured Snippets — those chunks of relevant text Google occasionally extracts from a high-ranked result, and places in the results page to speedily answer certain questions. In the example above, the search concerned the afterlife in Tibetan Buddhism, and Google’s AI-generated answer is, in all fairness, more useful than the Featured Snippet.
The Snippet is a general answer about Buddhism — a sort of bare minimum answer to the question that isn’t wrong, but also isn’t great. The AI’s answer includes details about the recently deceased spending 49 days on a plane of existence called the Bardo, however, and those are probably the key details a searcher who typed this would have been hoping to see.
So Google’s Snippet and its AI result are in direct competition with one another, right there on your screen. In my tests, the AI was usually just a rephrasing of the snippet, but was occasionally more relevant to my search, and easier to parse.
Here’s another example: in answer to the question “How far did the Clippers get in the playoffs?” the Snippet seemed to be answering a different question — perhaps “When was the last time the Clippers were in the playoffs?” Both the Snippet and the AI answer gave me what I needed eventually, but the AI felt like it was answering my question.
That’s the good news. The first bit of bad news is that it’s a mess.
The SGE visual experience is redundant and irksome
With SGE turned on, the above queries produced AI-generated results automatically, and the experience of getting an AI result in this way is a little off-putting when you aren’t expecting it.
Google made the strange choice to allow a full results page to load, but then have a little animated AI module materialize with the word “Generating” in it, and then fill in the space with an AI answer — a process that takes about three seconds in total. If one were kind, they might say it’s like watching a dried chrysanthemum flower bloom in their cup of tea. If one were unkind, they might say it’s like watching a shark emerge from the depths, surface, and then spread its monstrous jaws open right in the user’s face.
What’s more, the results page itself gets complicated if you have AI turned on. Google results pages could already be a little messy, particularly if Google senses you’re doing some shopping. Online shoppers are presented with a horizontally-arrayed selection of sponsored product images, perhaps followed by a Maps result with several “Places,” aka. retail locations, and then an actual, traditional search result.
Add one of SGE’s new AI-generated shopping modules to that list, and the results page is now the search engine equivalent of a stacked-to-the-ceiling Scooby Doo sandwich. Exactly how the new AI-generated module is meant to be any sort of improvement upon the old Places module (which was already generated by an algorithm, and therefore arguably a piece of AI tech) is unclear at best.
It’s also always different. When I searched “florist near me” again the day after I captured the screenshot above, I got a similar result, except now it came with a little summary of its findings, “There are many florists near Los Angeles, California, including those that specialize in weddings, offer gift baskets, or have a bohemian-chic vibe,” it told me. Which brings me to my next point…
It’s all a bit…why?
Seemingly as often as not, SGE doesn’t automatically pop up. It’s fortunate that this is the case with many controversial questions about which Google would clearly prefer that its large language model not shoot its mouth off. For instance, SGE didn’t come up automatically when I searched “which religion is right?”
And while it did present me with the option to try and generate an AI answer to this question, when I clicked it, my request was denied.
It’s just as well. The discourse is moving past the phase where it’s fun to throw life’s tough questions at chatbots and watch them squirm. An actual google search for “which religion is right?” is the kind of thing someone might search at 3 a.m., but they wouldn’t be looking for an answer. Whether they know it or not, they’d be looking to see other human beings wrestle with the topic on sites like Reddit and Quora.
And SGE makes no bones about the fact that it mines these types of subjective discussions for answers. Here, when I asked it what people in Balkan countries think about Yugoslavian reunification, Google says part of its answer is “according to a Quora post.”
I think I’d rather just read the Quora post.
The way it churns human writing into algorithmic outputs is, thankfully, disclosed. The sources are linked off to the right side, and you’re welcome to click them. But if you just came for a straight answer to an open-ended question, you probably won’t click.
If Google can get the public to rethink what search is for, there might be something here.
Having said that, the strongest case for the actual usefulness of SGE comes from a type of search that can’t possibly be common: extremely specific — but open ended — queries. For example, while I was writing this, I got a push notification on my phone saying the Supreme Court case Haaland v. Brackeen had been decided, and that the decision upheld the Indian Child Welfare Act. I had already read an article about the case, so I knew some of the broad strokes of what had just happened. I just wanted to know why.
Any reasonable person could — and probably should — just click a news article and scan it for the details they need. But if really pressed for time, one could perhaps try googling the following phrase: “what was the legal reasoning behind today’s supreme court decision upholding the Native American Adoption Law.” I suspect no one really uses Google in this way. Nonetheless, the AI-generated result was, well, decent:
The AI says the relevant law “does not discriminate on the basis of race and does not impose a federal mandate on traditionally state-regulated areas of power,” according to the majority opinion. After reading the AI’s blurb, I found some shades of significance in the accompanying NPR article that the AI could maybe have been clearer on, but for the most part the AI gave me what I needed.
But this is getting a bit Rube-Goldbergian. Given the kilojoules already expended every second answering Google queries, attaching an immense new source of computational power to those Google queries — whose job is to go into the pages being queried and summarize them — is a bit like attaching a sophisticated and expensive robot to the joystick of a claw machine at a video arcade in the hopes of slightly improving my chances of getting a stuffed donkey.
I already know how to buy a stuffed donkey: I’ll just Google it.