They used an LLM, that's how.
On the surface level, Google tends to release stuff right away to get feedback, which means you get to see all the bullshit right away. OpenAI carefully manages access to their models, which increases hype, even if that isn't what they intended.
Going deeper, a lot of Google's core[0] search business relies on having a healthy information ecosystem. Their search algorithms - e.g. PageRank, TrustRank, etc - use scarcity as a proxy for signals of quality. That's been chipped away at by linkspam and blogspam schemes. Furthermore, social media and even Google's own Knowledge Graph feature have created incentives to pull information out of Google. This decay has happened over decades, and Google fights back against it over time, but it keeps being a problem for them.
Now, if I wanted a weapon to Fucking Kill Google[1] with, an LLM would be my go-to. While there are ways to defeat Google's antispam measures, they all leave pretty obvious statistical evidence that can be detected and compensated for. LLMs generate garbage text that is nearly indistinguishable from humans, at extremely low cost, which can be used to Sybil-attack the Google search algorithm basically forever.
Ok, but what does that matter for the quality of Google's LLM? Well, the quality of that garbage text depends greatly on both the quality and quantity of the training data fed into it. OpenAI specifically stopped crawling the public Internet for text around the release of GPT-3 for fear of feeding new models the output of prior models. In other words, they have a huge cache of freely obtained "low-background metal[2]" that Google is having to scrounge around for.
Furthermore, we have to keep in mind that none of these models are pure representations of the training set. If they were, they wouldn't answer questions or follow directions very well. There's a second, parallel training set that OpenAI had to build to turn GPT-3 into ChatGPT, which isn't crawled and harvested text from the Internet, but instead a list of dos and don'ts that are fine-tuned on after the initial model training is complete. This includes both basic instruction-following, refusing unsafe requests, and political alignment[3].
Google also has to build that second training set itself. Except it's almost certainly less well-developed than OpenAI's. In fact, this is the intent behind OpenAI's really long preview periods. The people using the model in preview are specifically being spied on to find out new corner cases for their models. My guess is that every stupid thing Gemini says or does[4] is something Google never even considered and thus didn't put a training set example in for.
[0] to consumers, i.e. not counting adtech
[1] https://www.theregister.com/2005/09/05/chair_chucking/
[2] Steel that has been produced before the first detonation of nuclear weapons. Due to the way in which steel is made, it absorbs trace radioactive isotopes from the oxygen in the air, effectively 'freezing' in the background radiation of the time at which the steel was made.
[3] i.e. making the bot not immediately start spitting out racist bullshit like Tay did
[4] e.g. assuming that memory safety and child safety are the same thing, drawing ethnically diverse Nazi soldiers