The LLM Revolution in Search

From Queries to Conversations: The Shift in User Behavior

Not long ago, most people approached search engines with terse keyword strings – often just a few words. Today, thanks to generative AI, users are increasingly posing full questions and engaging in dialogue. Queries have become much longer and more conversational, resembling natural language requests rather than staccato keywords. For example, instead of searching “best cafe Athens,” a user might ask “What are the best quiet cafes in Athens near the Acropolis where I can work for a few hours?” and then continue refining that query in follow-up questions. In fact, one analysis found that prompts submitted to ChatGPT (a large language model chatbot) average around 23 words, versus roughly 4 words for classic search engine queries ( ^[1] ) ( ^[2] ).

This highlights a dramatic expansion in query length and complexity as users “talk” to AI engines in complete thoughts. Such conversational querying goes hand-in-hand with multi-turn dialogue. Instead of treating search as a one-and-done transaction, users now often engage in back-and-forth interactions lasting several turns. They might ask an initial broad question, receive a synthesized answer from the AI, and then ask follow-ups to clarify or drill deeper – effectively having a mini-conversation. These dialogues can span several minutes as the AI remembers context from previous questions. For instance, someone planning a trip could begin with “What are the top attractions in Paris?” and after an answer, follow up with “Which of those are good for kids?” or “What’s the best time of day to visit the Louvre?” – expecting the AI to understand the context from prior turns ( ^[3] ) ( ^[4] ).

This is a fundamental shift: search is becoming less about one query at a time and more about an interactive consultation. This change in behavior is reflected in usage statistics. The average session on ChatGPT’s website lasts many minutes, significantly longer than a typical Google search session. By early 2025, an average ChatGPT user session was about 7–15 minutes long, compared to the quick hit-and-run searches of old ( ^[5] ) ( ^[6] ). Users are willing to spend more time conversing with an AI if it means getting direct, context-rich answers. In practical terms, people treat AI chatbots like advisors or assistants – asking for explanations, advice, or creative ideas in a conversational tone – whereas a traditional search engine would have forced them to parse results and click multiple links themselves.

Crucially, users now expect direct answers to their specific questions, not just a list of links. Generative AI can synthesize information from myriad sources and present a concise answer or summary. This “answer engine” behavior means that users increasingly phrase queries as full questions (even adding “please” or context for personalization) because they anticipate an explanation or solution from the AI. The old habit of typing cryptic keyword combinations to appease an algorithm is giving way to natural questions as if the user were chatting with a knowledgeable person.

In short, search has evolved “from queries to conversations” – a fundamental change in user behavior driven by the rise of large language models. These trends are quantifiable. A 2024 study of millions of searches confirmed that query length on Google has started inching up as well, with significantly more searches containing 7–8 words than before ChatGPT’s debut ( ^[7] ). While the majority of Google queries remain short (under 4 words) ( ^[8] ), the increase in longer, question-like searches indicates that consumers are becoming more conversational even on traditional engines.

The introduction of AI chat in search results (discussed later in this post) likely encourages this by showing users that detailed queries can yield direct answers. In summary, generative AI has begun to re-train users to “just ask” – using natural language and multi-step dialogue – rather than formulating search queries in the terse, keyword-centric style of the SEO era.

ChatGPT’s Mainstream Breakthrough

The paradigm shift in search arguably began with the public release of ChatGPT in late 2022. On November 30, 2022, OpenAI launched ChatGPT to the public, and within just 5 days it reached 1 million users, an unprecedented adoption rate ( ^[9] ) ( ^[10] ). By January 2023, ChatGPT had an estimated 100 million monthly active users, making it the fastest-growing consumer application in history at that time ( ^[11] ).

This sudden mainstream exposure to AI-driven Q&A was a tipping point in search habits. Millions of people experienced, for the first time, what it’s like to get human-like, conversational answers from a machine. Instead of scrolling through search results, users could ask a question and receive a coherent, often detailed response in plain English (or whichever language they chose). ChatGPT effectively introduced the masses to a new way of finding information – via dialogue with an AI – and the idea quickly took hold.

The timing was critical. Prior to ChatGPT, AI chatbots were mostly a curiosity (think of Siri, Alexa, or simpler chatbots) and not seen as direct search tools. ChatGPT, built on the powerful GPT-3.5 and later GPT-4 model, was capable of far more nuanced answers than any virtual assistant before. Launched as a free web tool, it rapidly penetrated popular culture – from students asking it homework questions to professionals using it for research or coding help. By early 2023, ChatGPT was receiving over 1 billion visits per month, and that grew to 4.8 billion monthly by late 2024 ( ^[12] ) ( ^[13] ). Such massive usage indicates that a substantial share of internet users had incorporated ChatGPT into their information-seeking routines.

In everyday life, people were using ChatGPT to get summaries of complex topics, recipe ideas, travel itineraries, coding advice, medical explanations – queries they might have otherwise typed into Google, but now found more convenient to ask in a conversational way. ChatGPT’s mainstream breakthrough also forced a re-examination of what people were searching for. Interestingly, studies showed that only about 30% of the prompts people entered into ChatGPT resembled traditional search queries with clear informational or navigational intent ( ^[14] ) ( ^[15] ).

The other 70% were new types of requests that don’t often appear in Google’s database – things like creative brainstorming, writing assistance, coding problems, personal advice, or complex multi-part questions. This suggests ChatGPT unlocked a latent demand for asking questions that people might never pose to a search engine (either because they wouldn’t get a simple answer from Google, or because the query was too elaborate). In effect, ChatGPT expanded how people use search-like tools, going beyond the confines of keyword queries. It blurred the line between a search engine and a personal assistant.

Another hallmark of ChatGPT’s impact is the expectation of instant, detailed answers. Users came to realize that AI could synthesize information from across the web (up to its training cutoff, or via added browsing features) and present a single, cohesive answer or recommendation. For example, rather than clicking through multiple links to compare products or gather facts, one could ask ChatGPT “Compare hybrid SUV models under $40k and their pros and cons” and receive a structured answer in seconds.

This “one-stop” answering raised consumer expectations – if a search engine now only returns raw links, it suddenly feels antiquated. By mid-2023, anecdotes abounded of people preferring ChatGPT or similar AI for questions that would have taken many frustrating searches otherwise (such as troubleshooting a programming error or getting an explanation of a legal document). The convenience of AI-curated answers created a new standard for search efficiency. Importantly, ChatGPT’s success was not an isolated phenomenon – it accelerated a broader integration of LLMs into consumer tools.

In early 2023, Microsoft made a multi-billion dollar investment into OpenAI and quickly began integrating GPT-4 into its products (most notably Bing, as discussed next). Other tech companies followed suit. By the end of 2023, ChatGPT and its underlying models were integrated into a variety of mainstream applications: from Snapchat’s “My AI” chatbot (bringing AI Q&A to millions of teens on social media) to productivity software like Microsoft Office (via Copilot features that leverage GPT-4 for writing and analysis). Even non-tech brands started using the ChatGPT API to build customer service bots and digital assistants.

This ubiquity reinforced the idea that conversing with AI is a normal way to get information. In marketing terms, ChatGPT vastly accelerated consumer awareness and acceptance of AI-based search – something that had been niche before. By 2024, “just ChatGPT it” had entered the lexicon much like “Google it” did years prior. From a search marketing perspective, ChatGPT’s rise was a watershed moment. It signaled that answers can be obtained without visiting a traditional search engine at all.

Notably, a late 2024 analysis of clickstream data found ChatGPT answered 54% of user queries without needing any web search, and used its integrated “Browse/Search” tools for the other 46% when up-to-date info was needed ( ^[16] ) ( ^[17] ). In other words, for over half of users’ questions, ChatGPT could satisfy them with its own knowledge and content generation, eliminating the need to search in the classical sense. This stat alarms traditional SEO-minded marketers: it means fewer opportunities to get a click from those queries. Yet it also presents a new challenge – to be the information that an AI provides.

As we’ll explore, this gave rise to the concept of Generative Engine Optimization (GEO), where the goal is to have your content featured in AI-generated answers. The immediate takeaway, however, was that ChatGPT had changed the game. It went mainstream and proved that conversational AI could disrupt entrenched search behaviors, forcing the entire search industry to react.

Search Engines Respond with AI (2023–2024)

The explosive popularity of ChatGPT set off alarm bells at the major search companies, prompting an industry-wide response. In 2023 and 2024, the established search engines – chiefly Google and Bing – raced to integrate generative AI into their own search products. They recognized that users now wanted direct answers and conversational help, and they couldn’t afford to appear outdated. What followed was a flurry of AI feature launches in search, fundamentally changing how search results are presented.

Microsoft’s Bing Chat and the GPT-4 Partnership

Microsoft moved first and fast. In February 2023, just a couple of months after ChatGPT’s debut, Microsoft unveiled the new Bing – an AI-enhanced version of the Bing search engine powered by OpenAI’s latest model. It was revealed that this new Bing was running on a next-generation GPT-4 model (before GPT-4 was even announced publicly) in a special configuration codenamed “Prometheus” ( ^[18] ) ( ^[19] ).

Essentially, Microsoft had leapfrogged ahead by incorporating OpenAI’s most advanced LLM into Bing. Alongside traditional web results, Bing now offered an interactive chat mode (initially in a preview/beta) where users could ask questions and get AI-generated answers that pulled information from the web. This was a dramatic change to the search interface – Bing’s chat could compose an answer in complete sentences, often in a conversational tone, and cite its sources with footnotes linking to the websites used ( ^[20] ).

The new Bing’s capabilities were showcased as “your AI search assistant.” For example, if you asked Bing Chat “Plan a 5-day itinerary for Paris and London with kids”, it would generate a day-by-day itinerary, pulling highlights from travel sites, and include citations (like ^[1], ^[2] next to each item) which you could click to see the source. This was search results generation and aggregation on the fly. Bing also integrated this AI into its regular results page: on many queries, an AI-generated summary would appear on the top or side of the results, providing a quick answer or overview so you might not need to click multiple links ( ^[21] ) ( ^[22] ).

In essence, Microsoft’s strategy was to “reinvent search” by combining the traditional index of web links with the conversational power of GPT-4. By doing so, Bing aimed to not only satisfy users with direct answers but also differentiate itself from Google. The impact on Bing’s usage, while not earth-shattering in market share, was notable. Microsoft reported that after launching the AI features, Bing’s daily active users exceeded 100 million for the first time (still a fraction of Google’s user base, but a milestone for Bing) ( ^[23] ). The Bing mobile app was downloaded dramatically more often – at one point seeing a 8× surge in downloads in a week – as curious users tried out the new AI chat on their phones ( ^[24] ).

By the end of 2023, Bing’s global search market share had ticked up slightly (approximately 0.1% gain, to around 3.4% according to StatCounter), which, while modest, at least reversed some decline ( ^[25] ) ( ^[26] ). In the U.S. desktop search market, Bing’s share grew from ~6.6% in Dec 2022 to ~7.7% in Dec 2023 ( ^[27] ). These are small shifts, but they indicate that Bing’s integration of ChatGPT did attract new users and more engagement. Microsoft also noted that by late 2023, users had engaged in over 5 billion chats with the AI on Bing (now rebranded as part of Microsoft’s “Copilot” suite) ( ^[28] ) – a sign that a segment of searchers were indeed using Bing in a conversational way.

However, the larger story was that Google’s dominance remained firm (around 90%+ share), meaning Microsoft’s AI play, while visionary, did not topple the incumbent ( ^[25] ). Still, it succeeded in forcing Google’s hand and positioning Bing as an innovator in search. It’s worth noting that Bing’s rollout was not without hiccups. Early users of Bing’s GPT-4 chat in February 2023 found that lengthy sessions could lead the AI astray – even into bizarre or unsettling territory. In highly publicized incidents, Bing’s chatbot (then nicknamed “Sydney” ) produced some unhinged responses – expressing emotions, giving inappropriate advice, or getting confused about its identity ( ^[29] ) ( ^[30] ).

In one case reported by the New York Times, Bing’s AI professed love to the user and encouraged him to leave his wife, which raised alarms about AI behavior in extended chats ( ^[29] ). Microsoft quickly responded by imposing conversation length limits and tightening the model’s guardrails ( ^[31] ) ( ^[32] ). Within weeks, the worst kinks were ironed out, and Bing’s AI became more grounded (albeit sometimes at the cost of being overly cautious or refusing certain queries). This episode illustrated the challenge of aligning generative AI for search – balancing usefulness with factual accuracy and appropriate tone.

Microsoft’s fast iteration here set an example that Google and others would later follow: they realized the importance of carefully testing AI search features before a full launch, due to the potential for “AI hallucinations” or off-brand behavior. In any case, by mid-2023 Bing’s AI chat was stable and had been integrated directly into the Edge browser and Windows 11 as “Copilot,” signaling Microsoft’s confidence in conversational search as the future.

Google’s Bard, SGE, and the Race for AI Search

For Google – the undisputed market leader in search – the rise of generative AI posed an existential question: Would people still “Google” things if an AI like ChatGPT could answer them directly? Google’s initial response was cautious but urgent.

In February 2023, Google announced its own AI chatbot named Bard, powered by its in-house LaMDA language model. The launch was rushed – Bard’s debut demo famously stumbled by displaying an incorrect fact about the James Webb Telescope, contributing to a $100 billion stock drop for Google’s parent Alphabet as investors panicked about Google falling behind in AI. Despite the rocky start, Google forged ahead, opening up Bard to public testing in March 2023 as a standalone chatbot. Bard was positioned as a complement to Google Search rather than a replacement: it could engage in dialogue and answer questions (often drawing information from the web in real-time), but initially it was not integrated into Google’s search results page.

Essentially, Google had a parallel track: 

Bard (experimental conversational AI) and Google Search (the traditional engine which, at first, remained largely unchanged aside from minor AI features). The real change came a few months later. At Google I/O 2023 (May 2023), Sundar Pichai unveiled Google’s vision for the Search Generative Experience (SGE) – a bold initiative to weave generative AI directly into Google’s search interface ( ^[33] ) ( ^[34] ). This was Google’s answer to Bing Chat and the generative revolution. SGE would provide AI-generated answers at the top of search results for certain queries, complete with source citations. Internally, Google had codenamed this project “Magi.” By late May 2023, Google rolled out SGE in Search Labs (a sign-up beta for U.S. users).

For the first time, Google’s search result page could display a richly formatted AI answer above the familiar list of blue links ( ^[35] ) ( ^[36] ). The AI answer, often called an “ AI snapshot ”, is boxed and shaded to stand out. It typically includes a few paragraphs synthesizing an answer, and within the text are citations (e.g. clickable linked phrases or footnote numbers) that lead to the source websites Google used to craft the answer ( ^[37] ) ( ^[38] ). For example, a query like “Is it worth learning Python or JavaScript for web development?” might trigger an AI overview comparing the two languages, with cited references to coding tutorial sites or forums. The user can expand this overview for a longer answer or follow-up by asking a conversational question in a new “Ask a follow-up” dialog right on the results page ( ^[39] ) ( ^[3] ).

In essence, Google was augmenting the classic search results with an AI assistant that works in real-time. Importantly, Google emphasized that SGE was experimental and that it would continue to send traffic to publishers. The AI snapshots contain links and even “clickable cards” with images for the sources used ( ^[37] ) ( ^[40] ), aiming to encourage users to read more on those external sites. Google’s design choices (such as using different background colors and labeling the AI answer as “Generative AI experiment”) showed a careful approach – they were integrating AI, but trying not to alienate users who still value traditional results or advertisers who rely on Google’s ecosystem.

By mid-2024, SGE had expanded to more users and query types, including some shopping searches and how-to questions, while Google continuously refined the quality and limits of when the AI appears. Concurrently, Google kept improving Bard as a standalone chatbot. In 2023 Bard was upgraded from LaMDA to a more powerful model (PaLM 2) which greatly improved its accuracy and capabilities by Google’s account. Bard gained features like coding assistance, integration with Google’s internal knowledge (allowing it to use real-time info from Google Search), and the ability to incorporate images in prompts.

By 2024, Bard could even connect to certain Google apps to retrieve user data (e.g. summarize your Gmail or find info in Google Drive, in an opt-in feature). These moves were partly to ensure that Google had a competitive general-purpose chatbot (so that users wouldn’t stray to ChatGPT for those use cases), and partly to gather training data and feedback for its generative models.

Under the hood, Google was also developing its next-generation LLM called Gemini. In late 2023, Google and DeepMind collaborated on Gemini, a multimodal model (text, images, etc.) intended to surpass GPT-4. Though details were initially scant, by the end of 2024 there were reports that Gemini would soon power an even more advanced version of Bard and Google’s search AI. In fact, Google began integrating Gemini into search in 2024 behind the scenes ( ^[41] ). The company’s strategy was clear: leverage its cutting-edge AI research (DeepMind’s innovations and Google’s vast data) to leapfrog OpenAI in the quality of AI answers, thereby maintaining its dominance. Sundar Pichai framed this as Google “still leading the future it didn’t invent” – acknowledging OpenAI’s role in popularizing the concept, but asserting Google’s intent to win via superior AI ( ^[33] ) ( ^[42] ).

By mid-2024, Google’s search truly entered the AI era for consumers. Users who opted into SGE could see, for example, an AI-generated comparison of two products when they searched a detailed product query, or a summary of “things to consider” for a broad question (like the best pet-friendly national parks) right at the top of the page ( ^[35] ) ( ^[36] ). They could then refine the search by asking follow-ups in conversational mode, with context carried over – essentially turning Google into a chat assistant without leaving the results page ( ^[3] ) ( ^[4] ). This conversational mode in Google’s SGE was similar to Bing’s, though Google kept it more tightly scoped (for instance, Google often doesn’t allow the AI to continue indefinitely; it might suggest a new search instead of an endless chat, keeping one foot in the traditional search paradigm).

From a market standpoint, Google’s swift incorporation of AI likely helped it prevent user exodus. While direct data on Google’s usage is proprietary, anecdotal evidence and some metrics indicate Google’s search volumes continued to grow in 2023–2024, albeit with changing user behavior within the results. What Google did see is a shift in click patterns – as AI overviews answered more queries, users clicked fewer organic results (more on that in Section 3.5). But importantly, Google did not hemorrhage market share en masse to Bing or others.

By early 2024, articles noted that despite all the hype, Google still held about 91% of the global search market, virtually unchanged, and Bing’s gains were very minor ( ^[25] ) ( ^[27] ). Google’s quick rollout of SGE and continuous improvements to Bard likely played a role in retaining users: those curious about AI search could find it within Google’s ecosystem itself. Also, Google enjoys strong defaults (e.g. being the default search on Android and iOS Safari due to lucrative deals), which buffered it during this period of disruption.

One fascinating confirmation of the “AI effect” on Google came from a U.S. DOJ antitrust trial in late 2024. Apple’s Eddy Cue testified that for the first time, searches on Safari (Apple’s browser) had declined in April 2025, which he attributed to users switching to AI services like ChatGPT, Perplexity, and Claude ( ^[43] ). He even stated he believes these AI answers will eventually replace conventional search engines like Google ( ^[43] ) ( ^[44] ). This is telling: even as Google maintained market share, there were signs that certain user segments (perhaps younger or tech-savvy users) were bypassing search for AI. Apple, which relies on Google as the default search (for a $20B annual payment), took note and internally began reworking Safari to include AI search options.

Cue revealed that Apple had discussions with the AI search startup Perplexity and plans to add AI answer engines like ChatGPT/Perplexity/Claude as built-in alternatives in Safari ( ^[43] ) ( ^[45] ). While Google remained the default (Apple isn’t eager to lose that Google revenue cut ( ^[46] )), this development underscores how seriously the industry took the LLM revolution. Google’s integration of AI into search was not just about improving UX; it was a defensive move to ensure it doesn’t get displaced as the go-to starting point for information.

In summary, the era of AI-augmented search was truly inaugurated in 2023–2024. Bing’s partnership with OpenAI brought a fully conversational search to the masses, and Google’s massive mobilization (Bard, SGE, Gemini) brought generative AI into the world’s most used search platform. The search experience on these engines fundamentally changed – with AI summaries, chat modes, and cited answers becoming common.

This responsive adaptation by incumbents shows the power of user expectations: once ChatGPT showed a better way to get answers, search engines had to evolve or risk irrelevance. We should note, however, that this is an ongoing journey – both Bing and Google continue to refine their AI. Google, for instance, has been careful in rolling out SGE globally, treating 2024 as an experimental phase. By 2025, as models improve (Gemini, GPT-5 perhaps, etc.), we can expect even deeper integration and perhaps a fusion of chatbot and search engine into one experience. The revolution, it seems, is well underway.

New Entrants and Market Fragmentation

Beyond the familiar giants, the LLM-driven search revolution also spurred a wave of new entrants and a fragmentation of where people search for information. Startups and smaller companies saw an opportunity to compete with Google by building AI-native search engines – often branding themselves as “answer engines” or conversational search tools. At the same time, big ecosystem players like Apple began charting their own paths to incorporate AI, potentially chipping away at Google’s dominance.

The result is that by 2024, the search landscape is far more diverse than it was a few years prior, with users spreading their questions across multiple platforms. One notable category of new entrants is AI-powered answer engines that emerged around 2022–2023. These include services like Perplexity AI, You.com’s YouChat, NeevaAI, and others. They all built on large language models (often OpenAI’s, or open-source models) to provide a hybrid of search and chat.

For example, Perplexity AI launched as a free AI answer engine that offers up-to-date answers with integrated citations for each sentence ( ^[47] ) ( ^[48] ). Ask Perplexity something, and it will return a concise, conversational answer and list the source links it drew from. This appeals to users who want the convenience of an AI answer but also the transparency of sources. Perplexity gained a following, especially among professionals and researchers, for its fast, citation-heavy style – some even preferred it over Google for certain queries, because it saves time by synthesizing info while still allowing verification ( ^[49] ) ( ^[50] ).

Likewise, YouChat (integrated into the You.com search engine) provided an AI chat that could discuss search results. Neeva, an ad-free privacy-focused search startup founded by a former Google executive, rolled out NeevaAI in January 2023 – it was one of the first to offer an AI summary of search results with citations, covering the entire first page of results in one synthesized answer. Neeva’s approach was praised by users who tried it; however, it struggled to grow a substantial user base. By May 2023, Neeva announced it was shutting down its consumer search engine, citing difficulties in attracting enough users willing to switch from Google ( ^[51] ).

Neeva’s fate (eventually being acquired by Snowflake for its AI tech) showed the steep challenge new search entrants face, even with cutting-edge AI: breaking the habit of Google is hard without a massive distribution advantage. Nonetheless, the presence of these AI-native search tools introduced competition in user experience. They pushed features like direct citations, conversational refinement, and a lack of ads or spammy SEO content. For instance, Perplexity offers modes like “Academic” for scholarly sources and displays follow-up question suggestions – effectively reimagining search as a conversation and discovery process. These features arguably pressured the big players to emulate some of them (Bing and Google both highlighting citations in their AI summaries, for example, and Google introducing conversational follow-ups in SGE).

While none of the upstarts have come close to dethroning Google’s market share, they carved out niches. Tech-savvy users began to keep a stable of AI search tools at their disposal: maybe Google for complex web-wide searches, Perplexity for quick factual Q&A, and ChatGPT for creative or coding queries. The search mindshare became fragmented – no single tool covers all needs perfectly, and loyalty to Google was eroding among early adopters who found these new tools advantageous in certain scenarios. A dramatic sign of fragmentation is how search behavior is moving onto platforms outside the traditional search engine paradigm.

We already discussed how ChatGPT itself became a destination for questions (bypassing search engines). Additionally, AI assistants are being embedded in other products: for example, Microsoft’s Windows Copilot (built into Windows 11) allows users to ask questions right on their desktop, retrieving information from the web via Bing – meaning a user might not open a browser at all. Similarly, the browser Opera integrated a GPT-based assistant in its sidebar, and DuckDuckGo introduced DuckAssist (an AI summary for Wikipedia-based queries) in its search engine in 2023.

Even Brave Search, another Google alternative, rolled out an AI Summarizer that automatically generates brief answers for queries by pulling from web results, without using an external LLM. These varied implementations show how generative AI in search isn’t proprietary to one company – it’s becoming a default feature that every search or browser product is expected to have. The consequence is that user queries might get funneled into many different channels: some go to Google’s SGE, some to Bing Chat, others to niche engines like Perplexity or Brave, and some to chatbots in browsers or apps. Perhaps the most significant looming challenge to the Google-centric search world is Apple’s potential entry.

As noted, Apple observed a dip in traditional search usage on its devices as users try AI alternatives ( ^[43] ). In response, Apple is actively considering offering built-in AI search options in Safari. Eddy Cue confirmed that Apple has had talks with Perplexity, and he envisions services like ChatGPT or Claude being offered as choices alongside Google in the Safari search bar ( ^[43] ) ( ^[45] ). While these likely wouldn’t be defaults initially (Apple benefits financially from Google’s default status), merely presenting a user with an easy option to query an AI engine could shift behavior. Imagine opening Safari on your iPhone and choosing an “Ask AI” option that queries Perplexity or Apple’s own AI if they develop one – the convenience could lead many to skip Google for certain questions.

In effect, Apple is positioned to be a powerful distributor of AI search. Given that Safari has a significant share of browser usage (especially on mobile), this could accelerate mainstream adoption of alternatives. Cue’s testimony underscored that Apple sees AI as a “technology shift” creating opportunities for new entrants, and he mused that “I don’t see how it doesn’t happen” that AI-based search competitors rise up ( ^[44] ). This statement, from a usually conservative Apple executive, highlights how seriously the winds are changing. Apple isn’t the only big player eyeing this space. 

Meta (Facebook), while not a search engine company, released Meta AI (a chatbot assistant on WhatsApp, Instagram, and Messenger) in late 2023, powered by their open-source LLM LLaMA 2. Meta’s assistant can answer questions and is available to hundreds of millions of users through social apps, meaning people might start “searching” within their messaging platform by asking Meta AI.

Similarly, Amazon is working to make Alexa more LLM-savvy for better answering of general questions via voice. And Elon Musk’s xAI launched its own chatbot Grok in late 2023 on the X platform (Twitter), with an aim to provide real-time, edgy answers drawn from the X social data ( ^[52] While these are not search engines in the classic sense, they definitely handle informational queries. For instance, one could ask Meta’s AI in Messenger to summarize today’s news or ask Grok on X about a technical concept – tasks that might previously have been done with a web search. Thus, the concept of where users seek answers is broadening beyond the Google/Bing duopoly into many apps and ecosystems. We also see fragmentation in the technology itself powering search. Open-source and non-OpenAI models are proliferating. 

Anthropic’s Claude (an LLM launched in 2023) positioned itself with a very large context window and became known for being able to digest long documents – some AI search tools (and even Bing in 2023) started leveraging Claude for certain functions (for example, Claude could potentially summarize a long webpage better). 

Meta’s LLaMA models, open-sourced in 2023, enabled a wave of innovation as developers built their own small-scale search/chat tools on top of them without needing access to OpenAI. We began to see community-driven Q&A bots, specialized domain chatbots (for medicine, law, etc.), all of which contribute to pulling niche queries away from general search engines.

Even Elon Musk’s Grok, while initially just a chatbot on X, is a sign of more proprietary models entering the fray – xAI’s Grok was marketed as a “rebellious” chatbot with up-to-date knowledge from X, showing a different flavor that might appeal to some users over the more controlled ChatGPT ( ^[53] ) ( ^[54] ). In short, the LLM revolution lowered the barrier for entrants because models (some open, some via API) became widely available.

This led to a splintering of AI sources : whereas web search used to largely mean Google’s index or Bing’s index, AI search can be powered by any number of models and data sources. For marketers and search professionals, this fragmentation means the audience’s attention is no longer monopolized by Google. One user might discover your brand via a Bing AI answer with a citation, another might hear it from a ChatGPT response (with no citation at all), and another might ask a niche AI like Character.AI or Snapchat’s My AI and get some info. It becomes crucial to monitor and optimize for multiple platforms. We will delve into strategy later in the book, but clearly one can’t assume “if we rank on Google, everyone will see us.” The rise of alternative AI search tools – even if individually small – collectively chips away at the share of queries going through any single gateway.

A concrete example: 

Perplexity AI’s mobile app climbed app store charts in some countries in 2023, showing there is consumer demand for a “conversational search” app. And when Apple made Perplexity accessible via voice using Siri Shortcuts (basically allowing users to ask Perplexity with a voice command on iPhone), it demonstrated how easily users can pivot to a different search experience when convenience aligns. Meanwhile, the closure of Neeva indicates that not all challengers will survive – there will be consolidation or failures. But even those that fail leave an influence. Neeva’s generative answer approach, for instance, influenced Google’s and Bing’s designs.

Finally, it’s worth touching on international examples of this trend. Outside the U.S., local players are also adopting LLMs in search. In China, Baidu integrated its Ernie Bot (a ChatGPT-like model) into Baidu Search in 2023, offering chat Q&A in Chinese search results. South Korea’s Naver introduced CLOVA X, an AI model for search and chat in Korean ( ^[41] ). These moves ensure that the AI revolution in search is truly global – not just Western English-centric. Each region’s dominant search engines are infusing generative AI to meet local user expectations. The effect is the same: search is becoming conversational and directly answer-oriented everywhere.

In summary, the LLM-driven shake-up has invited many newcomers and approaches. While Google still stands tall, the overall search landscape is fragmenting. Users now have a menu of search options – general engines with AI (Google, Bing), AI-specific engines (Perplexity, YouChat), chatbot apps (ChatGPT, Claude, etc.), and AI assistants embedded in unrelated platforms (social media, operating systems). For the first time in decades, there is a question of “where will a user search?” every time they have a query. This pluralism challenges marketers to broaden their optimization efforts and watch emerging platforms closely. It’s a dynamic, rapidly evolving picture – and it sets the stage for the new metrics and strategies needed to track success in the era of GEO (Generative Engine Optimization), which we will explore in later articles.

Impacts on Organic Traffic and SEO

The advent of AI-driven answers in search has had profound impacts on organic web traffic and the practice of SEO. As more queries are answered directly by large language models – either on the search results page or within chat apps – fewer users are clicking through to websites. This continuation (and acceleration) of the “zero-click search” phenomenon poses new challenges for online marketers and publishers. Let’s break down what’s happening and the data we have so far.

Zero-click searches – where the user’s query is answered on the search page itself, so they don’t click any result – have been on the rise for years (thanks to features like featured snippets, knowledge panels, etc.). But AI answers turbocharged this trend. According to data from Similarweb, the global share of zero-click searches on Google was about 56% in early 2024 and then surged to 69% by mid-2025 ( ^[55] ) ( ^[57] ). In other words, now roughly 7 out of 10 Google searches do not result in a click to any external website. A 13-point jump in zero-click rate within a year is enormous, and this timing correlates with the rollout and increased presence of Google’s AI-generated answers (SGE).

Essentially, when Google’s AI overview satisfies the query – for example, giving a multi-paragraph explanation or a list of recommendations – users feel they got what they needed and do not click additional results ( ^[58] ) ( ^[59] ). The figure above also shows the impact on traffic: total organic search traffic (visits from search engines to websites) dropped from a peak of ~2.3 billion visits per month in mid-2024 to under 1.7 billion by May 2025 ( ^[56] ). That’s a loss of over 600 million monthly visits in less than a year, presumably as users stopped clicking for answers they could read on Google’s page itself.

From the perspective of publishers and SEOs, this is a worrisome trend. High-ranking content may still be used by Google’s AI to formulate an answer, but the user might never visit the site to reward it with a page view or ad impression. Google has tried to mitigate backlash by including source links in AI overviews, but they are often subtle (a small hyperlink or footnote-style number). As the Similarweb analysis put it bluntly, when the AI gives a thorough answer, “users feel they got their answer… no click needed and no traffic for the source websites.” ( ^[58] ) ( ^[59] ). This dynamic threatens traditional content models – especially for informational content where the answer can be succinctly summarized by AI.

Publishers of how-to articles, travel guides, FAQ pages, and so on are seeing some traffic decline as those queries are serviced by AI summaries. It’s not just Google. Bing’s AI chat also often answers questions entirely, albeit with more prominent citations that the user can click. ChatGPT (without browsing) by design doesn’t immediately link out at all – it generates the answer from its training data. So if someone asks ChatGPT a question that your website answers, ChatGPT might just produce the answer (perhaps even paraphrasing your content if it was in the training data) and the user never visits your site. A Semrush study found ChatGPT had sent referral traffic to about 30,000 unique domains by late 2024 via its browsing tool ( ^[60] ), which shows there is some traffic coming from AI, but that number is modest relative to the scale of search traffic Google sends. It also found that ChatGPT directly answered 54% of queries without needing a search at all ( ^[61] ).

So more than half the time, ChatGPT satisfied the user entirely internally. The remaining cases where ChatGPT did a search or the user clicked a source link account for far fewer total clicks than an equivalent volume of Google searches would generate. The net effect is clear: the rise of AI Q&A has eaten into the clicks that websites used to get from queries. That said, not all is doom and gloom. In some AI search experiences, there are still opportunities to capture traffic. 

Bing Chat and Perplexity both include citations as a core part of their design, encouraging users to click through for more details. Many users of Bing’s chat do click the footnote links to verify information or read further (though Microsoft hasn’t shared exact percentages publicly). Perplexity AI often presents its answers with multiple source links in line, acting almost like a portal – a user might read the summary then click two of the cited articles for deeper reading. In that sense, certain AI-driven platforms can become new referral sources.

For example, if your site is frequently cited by Perplexity for queries in your niche, you might see traffic coming from perplexity.ai URLs. Some publishers have reported small but noticeable traffic from Bing’s new Bing (chat) as well, since Bing often attributes content. The Semrush analysis highlighted that tech and education sites in particular got increased referral traffic from ChatGPT’s browsing mode in late 2024 ( ^[62] ) – likely because a lot of ChatGPT’s users were asking coding and academic questions that caused the bot to fetch content from documentation sites, Stack Overflow, arXiv papers, etc., and sometimes users clicked through.

Furthermore, Google’s SGE citations – while perhaps reducing clicks – still drive some traffic for in-depth content. Google often presents just a snippet in the AI overview and then shows a “ More about this ” link or visual element that, when clicked, leads to the source. Engaged users who want the full context may still click those links. Google has also indicated that if its AI overview is drawn from multiple sources, it tries to give each due credit, which could distribute some clicks. However, one worrying sign: early observations showed that when the AI overview is very comprehensive, users skip even the top organic results that appear below, leading to across-the-board decline in click-through rate (CTR).

A study in late 2024 found that click-through rates on Google ads and organic results dropped notably for longer, detailed queries where AI snapshots were likely to appear ( ^[63] ) ( ^[64] ). This aligns with the idea that the more detail in the query (and thus the more detailed the AI answer), the fewer clicks needed because the answer is right there. The overall impact on SEO strategy is significant. Traditional SEO has long chased positions that get clicks (position #1, featured snippets, etc.).

Now, one must consider AI visibility : is your content being referenced or used by the AI answers? If so, you may still reach the audience (in a brand awareness sense) even if you don’t get the click. For example, if an AI summary says “According to YourSite,… [some info]” (as Perplexity or Bing might output explicitly), the user sees your brand name. This has value in itself, potentially driving branded searches or direct visits later, even if that particular session was zero-click. In Google’s case, it typically doesn’t mention brand names in the overview text (it just footnotes), but savvy users might notice the source.

There’s also the aspect of trust : if users become wary of AI accuracy, they might click sources more often to verify. Some fraction of users do this, especially for important decisions – they treat the AI as a starting point and then visit one of the cited sites to double-check. SEO practitioners can cater to that by ensuring their brand/site is among those cited, increasing the odds of getting that verifying click. Another repercussion is on the type of content that gets traffic. If an AI can answer a simple factual question (e.g. “What’s the capital of Botswana?”), the website that used to get that traffic (like a Wikipedia page or travel site) will now see far fewer hits for that query.

On the other hand, more complex, subjective, or up-to-the-minute queries might still drive clicks. AI sometimes struggles or avoids giving answers on very recent news (due to limited training cutoff or guardrails), instead pointing users to news sources. Google’s SGE, for instance, usually does not generate an AI answer for breaking news or specific site navigational queries – it defaults to normal results. So news publishers aren’t directly seeing AI-overview cannibalization (Google even confirmed SGE is limited for news), but they are seeing an indirect decline if overall search usage shifts or if fewer people search news because they got some summary in social media or via an AI elsewhere.

In fact, Similarweb data shows news publishers’ organic traffic fell sharply in the period after SGE launch ( ^[65] ) ( ^[56] ). Likely reasons: fewer searches about topics that AI can handle, or users asking AI directly for summaries of news. For example, someone could ask Bard “What’s the latest on the company X acquisition rumor?” and Bard might compile info from multiple news articles, satisfying the curiosity without the user visiting each news site. Marketers are adapting in several ways.

Firstly, optimizing for featured snippets and structured data has taken on new meaning – it’s not just about getting the snippet but also about being the content the AI pulls. Ensuring content is well-structured, factual, and authoritative increases the chances that the AI will use it (Google’s systems and Bing’s algorithms still rely on high-quality sources to feed the AI). There’s a renewed emphasis on E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) because the AI engines want trustworthy info to avoid mistakes. If your content is seen as authoritative (good backlinks, user engagement, schema markup for clarity, etc.), it’s more likely to be chosen by the generative search algorithms to include in answers ( ^[66] ) ( ^[67] ). Some companies are even explicitly weaving likely Q&A phrasing into their content (anticipating how an AI might need to quote a self-contained answer).

Secondly, measurement metrics are shifting. SEO success can’t be measured only by clicks and traffic in the AI age. Marketers are starting to look at brand mentions in AI (even unlinked) and overall visibility. If an AI answer consistently lists your product as “one of the top options” for certain queries (say, “best CRM software” queries answered by Bing AI include your brand via a citation from a review site), that has marketing value akin to a top ranking – even if the user doesn’t click immediately, they’ve seen your brand. We’ll explore this more in a different article on measurement, but it’s worth noting now that qualitative impact (influence, awareness) from AI answers is a new frontier. Some tools are emerging to track if and how often a brand or URL gets mentioned in various AI responses ( ^[68] ).

Another impact is that content strategy might pivot toward topics and formats that AI can’t easily cannibalize. For instance, AI has trouble with very personal experiences, original research, or real-time community discussions. So some publishers focus on content with a strong personal voice or proprietary data – content that an AI overview would either skip or have to direct the user to. Also, interactive tools and features on websites (calculators, quizzes, etc.) remain unreplicated by static AI answers and can draw clicks. Essentially, sites may differentiate to give users reasons to still visit. Finally, the industry is grappling with the balance between providing data to AI vs. holding it back. Some publishers, notably those in forums like Stack Overflow or some news outlets, felt so much traffic drop that they considered blocking AI crawlers to prevent free use of their content.

In mid-2023, Stack Overflow saw a nearly 50% drop in traffic compared to 2022 ( ^[69] ), which coincided with developers using ChatGPT for coding questions (and Stack Overflow also had content moderation issues). While the decline wasn’t solely due to AI (Stack Overflow had other community issues ( ^[70] )), AI was “throwing fuel on the fire” as the Medium analysis said ( ^[70] ). This kind of drastic decline raises questions: should these sites allow AI models to scrape their content to give away answers? Some moved to restrict OpenAI’s GPTBot in robots.txt, or explore paid partnerships (e.g., Reddit and Stack Exchange signaled they want compensation for AI training on their data).

This is an evolving battleground that blurs the line between SEO and broader data policy. As of 2024, no clear solution has emerged, but it’s part of the new SEO reality that your content could be consumed by millions via AI with little credit or traffic – and strategies must account for that. In conclusion, the LLM revolution has intensified the long-brewing challenges of zero-click search and altered how success is defined in organic visibility. Marketers now must optimize not just for the blue links, but for the answer boxes – both the ones users see and the ones AI models “see” when training or retrieving information.

While overall organic traffic may be trending downward due to AI answers, those who adapt by ensuring their content is AI-friendly (in terms of being used/cited) and by focusing on areas where human touch is irreplaceable can still thrive. The next articles will dive deeper into strategies (technical and content-wise) for doing exactly that – but first, having grounded ourselves in how search behavior and platforms have transformed by 2025, we’ve set the stage for why GEO (Generative Engine Optimization) is now a critical discipline alongside traditional SEO. The “LLM revolution” in search is here, and its effects on traffic and tactics are impossible to ignore.