Since Google announced its beta testing of SGE (now AI Overviews) during Google I/O 2023, the internet has been abuzz with talk of a new GenAI era of search. But this “new” era began in November 2022, when users started asking questions to ChatGPT. The release of OpenAI’s chatbot marked the transition from search engines to answer engines, something we are experiencing today. Yet, this transition could have been much smoother.
VP of Engineering at SE Ranking.
Google AI Overviews: official launch or huge public beta?
At Google’s recent I/O, the Google team made a grand announcement about the public launch of AI Overviews, mentioning the AI word 124 times. Despite the search giant’s flashy presentation, the rollout of AI Overviews has been rather conservative, available only to logged-in users in the US. In SE Ranking’s latest research on AI Overviews, we observed a significant drop in the presence of AI Overviews in search results after Google held I/O. Now, only 8.71% of keywords feature AI Overviews, a substantial decrease from our pre-rollout findings, where 64% of keywords had AI Overviews.
Google’s cautious approach might be due to the criticism AI Overviews have faced from the media and users. It could also be because AI Overviews require more computational resources than traditional search results, and the monetization of AI answers is still in development. This is why the official launch of AI Overviews seems more like a large public beta test for Google to gather more data and train its LLM.
It also seems like Google has internal mechanisms and criteria for evaluating the quality of AI Overviews-generated answers. Post-launch, these criteria appear to have become stricter. Google allows Gemini to generate answers in niches where it is more confident in the responses. According to SE Ranking’s latest research about AI Overviews, the Relationships niche leads, triggering AI Overviews 26.62% of the time, followed by Food and Beverage (24.78%) and Technology (18.11%). In contrast, niches more closely related to user well-being, such as Healthcare (0.44%), Legal (0.34%), and News and Politics (0.24%), trigger AI Overviews less than 1% of the time. This is presumably because AI’s hallucinations and biases have greater potential to cause harm in these niches.
Google and OpenAI vs. hallucinations and biases
Harmful answers can still appear in niches where Google frequently triggers AI Overviews. For instance, there was a recent incident where an AI Overviews suggested adding glue to pizza sauce, citing a joke post on Reddit. While integrations with platforms like Reddit, Quora, Stack Overflow, and Wikipedia offer vast amounts of human-generated data, the moderation of such content was not designed to train LLM models. Reddit, in particular, is full of jokes and inappropriate memes.
Interestingly, OpenAI has also struck a deal with Reddit, but ChatGPT does not produce potentially harmful answers as often as AI Overviews do. This discrepancy suggests that Gemini may rely more on citing Reddit directly rather than analyzing the content. On the other hand, OpenAI’s focus on creating an LLM with better “critical thinking” capabilities has allowed it to surpass Google in the AI race. Google had DeepMind at one point, but few people care or even remember.
Google’s major advantage, however, is the enormous volume of data it can use to train its models. Beyond the billions of daily Google searches, the company can access data from 3 billion Android users. This exclusive integration into a vast system like Android is something no other LLM model can achieve. Yes, OpenAI has also integrated its LLM into Apple’s OSes, but the deal is not exclusive, as Apple follows a no-vendor-lock policy. Apple has also created its own LLM and plans to integrate other models, like Claude.
However, considering OpenAI’s partnerships with Reddit, Apple, and Microsoft, they may have the ambition and resources to surpass Google. Will OpenAI launch its own answer engine? It’s entirely possible.
OpenAI is more agile and youthful compared to the massive and cumbersome Google. The company also tends to make bold or rather emotionally driven decisions — like the departure and return of Sam Altman and the recent controversy involving Scarlett Johansson’s voice during the launch of ChatGPT-4o, among others.
Nevertheless, OpenAI is staying noticeably cautious. It seems to be weighing the pros and cons of launching an answer engine that would directly compete with Google. This deliberation stems from concerns about AI hallucinations and biases, which could lead to the spreading of misinformation if not carefully managed. Alternatively, OpenAI may leave the launch of answer engines to other players who plan to use OpenAI’s LLM.
AI Overviews main challenges
Google’s advantage regarding the volume of data available for training LLMs is that it can only be surpassed by Google itself. According to SE Ranking’s latest research on AI Overviews, scrolling to organic results is becoming increasingly difficult. Featured snippets appear next to AI Overviews 45.39% of the time. When featured snippets and AI Overviews appear together, their sources match 61.79% of the time, meaning users see the same information multiple times. Additionally, ads now accompany AI Overviews 87% of the time, up from 73% before the rollout during Google I/O 2024.
The co-occurrence of AI Overviews, featured snippets, and ad blocks may be a step towards training the model to compile an AI Organized Search Results Page. Currently, users might have to skip the first page and go to the second to find organic results that better answer their questions.
Hallucinations in AI Overviews’ answers, duplication of information across different snippets, and the distant placement of genuinely helpful answers could alienate users. This might reduce their loyalty and trust in Google and push them to adopt new search methods, including ChatGPT. Moreover, OpenAI already provides users with more than just search answers. Gemini has only now begun to integrate into G-Suite, and the success of this integration remains uncertain. In the alternative and worst-case scenario, some users might blindly trust AI Overviews’ answers without fact-checking. Google and other LLM models must take this responsibility seriously.
If users prioritize quick and accurate answers, the shift to AI Overviews and an AI Organized Search Results Page could hurt the SEO industry, small and medium-sized businesses, news outlets, and blogs made by independent creators. Once AI satisfies users’ queries, the need to scroll to organic results and dig deeper will disappear. This change in user behavior will make it difficult to drive traffic, and reaching the target audience outside of advertising will become nearly impossible. Consequently, more businesses, media outlets, and creators will likely migrate to social media, email newsletters, or niche platforms, while some may go bankrupt and struggle to recover.
It is unclear whether Google will seek a balance between transitioning to an answer engine and supporting business users (this includes media and independent creators). Perhaps the SERPs as we know them will become a feature like “Google Classic”? If not, we could be heading towards an even more fragmented and dull Internet.
So, who will lead the GenAI era of search?
Both Google and OpenAI have significant advantages. Google’s user base and exclusive integrations are unmatched, while OpenAI’s years of focused development and influential partnerships position it as a strong competitor. The winner of this race will likely be determined by who can best balance technological innovation with user needs and ethical considerations.
Overcoming hallucinations, addressing user adoption and trust challenges, and supporting businesses and media will be crucial to leading the search revolution. Fact-checking remains the primary challenge for all AI models. ChatGPT often declines to answer certain questions, stating it’s “just an AI model,” while Google reduces the number of AI Overviews triggered answers.
Despite its flops this year, Google could still win the GenAI search race. Integrating Gemini into Android will help 3 billion people use AI in their daily lives. For example, Android users will be able to use AI to identify whether a tomato is still good to eat simply by opening the camera and speaking. If nearly half the planet uses Gemini, it will become the most trained AI model in the world. However, to do this at this scale, Google must find a way to obtain these resources without consuming electricity at the rate of a large European country.
Nevertheless, remember that OpenAI can be unpredictable, and this is a rapidly changing field. Current and new advancements that we may not yet be aware of could shift the landscape entirely.
We’ve featured the best AI writer.
This article was produced as part of TechRadarPro’s Expert Insights channel where we feature the best and brightest minds in the technology industry today. The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/news/submit-your-story-to-techradar-pro
https://cdn.mos.cms.futurecdn.net/rNZmVCrdHzszaDCyTrBWmj-1200-80.jpg
Source link
