Bench Talk

How generative AI is tipping the scales in the fight for the internet’s most valuable resource: you

Remember when “surfing the web” meant visiting different websites that each did one thing? You’d read news on news sites, browse text-only hobbyist forums, play games on gaming sites, and watch videos on video sites. Websites were bulletin boards where we all saw the same content, but the challenge was knowing where to look.

Content aggregation sites like Digg and Reddit launched in the mid-2000s, offering users a chance to see trending content across the web in one feed. But Facebook’s 2006 launch of the News Feed set a Rube Goldberg machine of innovation into action,[1] where social media titans experimented with how to keep users from hopping to a different website by giving them everything they visited the web for in one place. All you needed to do was keep scrolling to get your news, updates from friends, online groups, and entertainment.

Fast-forward to 2019, and the viral tweet in Figure 1 perfectly captured the late stage of the algorithmic content treadmills:

Figure 1: Ironically, I first saw this tweet as a screenshot on Reddit. (Source: Author/Reddit)

Search Engines Want Your Attention Too

Search engines also began employing the “keep users from leaving your site” strategy. Normally, you would enter your query in a search engine and click the results to find the information you’re looking for. In 2014, Google launched Featured Snippets,^[2] which pull a quick answer from the top result to save you a click (Figure 2). We call that a “zero-click search” in the ‘biz, and they keep web admins awake at night. Having your content appear in a zero-click search means you have the information people are looking for, but you are not rewarded with their traffic on your site.

Figure 2: I got the answer I was looking for without having to click through to the top result. (Source: Author)

The recent explosive rollout of generative artificial intelligence (AI) left almost no piece of software untouched, and online search was no exception. Now, in addition to the featured snippet, search engines offer a generated answer to your query instead of directly pulling a quote from a result (Figure 3).

Figure 3: Google’s AI Overview gives you a more thorough answer than the featured snippet. (Source: Author)

AI Overviews siphon even more traffic from websites. Last year, research firm Gartner predicted that traffic from search engines will drop by 25 percent by 2026 due to searchers getting the information they need from an AI tool rather than clicking on websites.^[3]

AI Shifts the Balance of Power in Online Search

AI-powered search is a solution to a problem that generative AI created. There’s so much slop online that it makes sense to have AI weed through its own content for us. Even before the generative AI content explosion, the top search results were often lazy keyword-stuffed listicles full of affiliate links. Today, anyone can churn out thousands of words in seconds with a few mouse clicks, creating an ever-expanding library of web pages competing for your attention.

The most important takeaway from all of this is that the equilibrium of search engines and websites has shifted. There was a balance between the three agents involved in seeking information:

• The user wants information.
• The search engine knows where to find this information.
• The web page holds the information.

So when a user made a search query, the search engine would act as an operator and point the user to the web page containing the information they want (Figure 4).

Figure 4: In the days before AI, all three agents got what they wanted (Source: Author)

Today, search engines powered by generative AI algorithms have consumed all the information they can find. Instead of acting as an information operator that connects users to web pages, search engines offer their own answers (Figure 5).

Figure 5: Search engines use generative AI to cut other web pages out of the equation (Source: Author)

The changes to search bring both benefits and new challenges for users. As we’ve demonstrated when looking for a recipe, quick informational searches are much more convenient with generative AI. The tradeoff is that you may need to fact-check your results.

Generative AI is helpful to a fault. It will never tell you it doesn’t know something. Like an older sibling, it’ll just make something up. We call these happy little AI accidents “hallucinations.” Google’s AI infamously advised users to glue cheese to their pizza and gave recommendations on daily rock intake (at least one small rock per day, for those curious).^[4]

The true challenge is in fact-checking AI results with other content also written by AI. It’s possible that even if you scroll past the AI results, you’re still reading content that someone else generated with AI. Did they check it for hallucinations? Maybe, maybe not. Hallucinations have manifested as made-up citations in court documents,^[5] false news headlines,^[6] and an airline chatbot that created a discount for a customer.^[7]

It’s one thing to interact directly with a large language model (LLM); we know that the raw output will likely need some polish. But when using our news apps or a corporate chatbot, we expect guardrails to keep the AI running by the books. The challenge is many of these giant AI models are black boxes running on secret sauce, so it’s hard for developers to guarantee accurate outputs when they can’t see what’s happening under the hood.

Each AI model iteration decreases its chances of hallucination, but this is happening in parallel with an increase in the amount of AI-generated content being used to train AI, either unintentionally due to the increase of AI-generated content on the open web, or intentionally with synthetic data created by AI for the purpose of training itself.^[8] The free-for-all of scraping every word humans have ever written is coming to an end as more web domains put virtual fences around their content to keep the bots out, so increasingly large models will have no choice but to create their own training materials.^[9]

What’s Next for Users?

The internet is in another growth spurt and there’s no knowing how it will emerge from its current awkward phase. AI certainly provides convenient new ways for us to get information, provided we can rely on that information. But it also creates many new ways for social media and tech companies to compete for your attention. OpenAI, Google, and Microsoft now offer agentic AI “deep research” features that deliver an extensive report on a topic of your choice. Setting aside the concerns about adding a hallucination-prone layer of abstraction between users and primary sources, these tech developments point to a future where “surfing the web” becomes something for algorithms, not humans, to do.

AI can only regurgitate existing content, so original research and expert insights will become more valuable. I predict the humble newsletter will have a strong resurgence over the next few years, where people can cherry-pick the authors and publications they want to hear from to filter out the ever-increasing noise. We’ll also see websites walk the tightrope of optimizing themselves for both AI algorithms and human users.

Sources

[1] https://www.cnet.com/pictures/facebook-then-and-now-pictures/
[2] https://blog.google/products/search/reintroduction-googles-featured-snippets/
[3] https://www.gartner.com/en/newsroom/press-releases/2024-02-19-gartner-predicts-search-engine-volume-will-drop-25-percent-by-2026-due-to-ai-chatbots-and-other-virtual-agents
[4] https://www.forbes.com/sites/jackkelly/2024/05/31/google-ai-glue-to-pizza-viral-blunders/
[5] https://www.reuters.com/legal/transactional/another-ny-lawyer-faces-discipline-after-ai-chatbot-invented-case-citation-2024-01-30/
[6] https://www.bbc.com/news/articles/cd0elzk24dno
[7] https://www.cbsnews.com/news/aircanada-chatbot-discount-customer/
[8] https://www.msn.com/en-us/technology/artificial-intelligence/nvidia-google-openai-turn-to-synthetic-data-factories-to-train-ai-models/ar-BB1raFhg
[9] https://www.nytimes.com/2024/07/19/technology/ai-data-restrictions.html