AI search: better answers or expensive hype?

Type a question, read a finished answer. That is how Google presents its AI search. It shows up in about 43 percent of all Google searches (Source: Similarweb, 2026), out of roughly 14 billion searches worldwide per day (Source: Google, March 3, 2025). And by Google’s own account, more than one billion people now use its extension, AI Mode, every month (Source: Alphabet, Q2 2026 Earnings Call, July 22, 2026). The question for brands: is this the better search, or an expensive rebuild that misses what searchers actually need?

8% vs. 15%

click rate on search results with and without an AI Overview

Pew Research Center, 2025

1%

of searches with an AI Overview led to a click on a cited source

Pew Research Center, 2025

>60%

incorrect answers from eight AI search tools across 1,600 test queries

Tow Center, Columbia University, 2025

Contents

• The starting point: is newer really better?
• The informational bias: when every search becomes a question
• The monetization hypothesis
• The edge-case dilemma: wrong with full confidence
• The other side: where language models shine
• Ready for the new search reality?
• Conclusion: measure instead of waiting
• Frequently asked questions about AI search
• Sources

The starting point: is newer really better?

For twenty years, Google search ran on a simple contract. People search, Google shows ten links, websites get visitors. That contract is being rewritten.

Since 2024, AI Overviews have sat above the search results: automatically generated summaries with a handful of source links. AI Mode, a chat interface for search, followed in 2025. In summer 2026, Google merged both into a single experience (Source: Alphabet, Q2 2026 Earnings Call, July 22, 2026). The answer now mostly comes first, and the links slide down.

Google calls this progress and backs it with reach: more than one billion monthly active users in AI Mode since the global rollout in October 2025, plus “billions of clicks every week” sent to websites from AI features (Source: Alphabet, Q2 2026 Earnings Call, July 22, 2026).

Independent data paints a different picture. Pew Research Center analyzed the search behavior of 900 US adults in March 2025, a total of 68,879 Google searches. When an AI Overview appeared, users clicked on a search result in 8 percent of visits. Without a summary, the figure was 15 percent (Source: Pew Research Center, 2025). The click rate halves.

A second effect comes on top. After a search with an AI Overview, 26 percent of users ended their browsing session entirely. Without a summary, it was 16 percent (Source: Pew Research Center, 2025). The answer makes users end their search, either because it is enough or because it makes them give up.

For presentations:SVGPNG

The informational bias: when every search becomes a question

“The future of search is conversational.” That is the tech giant’s creed. Behind it sits an assumption: people want answers in full sentences. That holds for some searches. For others, it does not.

The SEO industry has distinguished by search intent, the purpose behind a query, for years. Google’s own guidelines for quality raters use categories such as Know, Do, Website, and Visit-in-Person. A special case is the Know Simple query: a factual question with a short, generally accepted answer (Source: Google, Search Quality Rater Guidelines).

• Navigational intent: Someone types “Chase login” and wants exactly one page. Any wall of text in front of it is an obstacle.
• Transactional intent: Someone searches “running shoes size 10 buy” and wants to see an offer, not an essay on cushioning systems.
• Know Simple: Someone looks up a politician’s date of birth. A number is enough. Three paragraphs of biography are not a service. They are ballast.

This is the informational bias of AI search: it tends to treat queries as if a complex question sat behind them. The Pew data also shows the limits of this criticism, though. In March 2025, 18 percent of all searches produced an AI Overview. For one- and two-word searches it was 8 percent, for searches with ten or more words 53 percent. When a search began with a question word, the share was 60 percent (Source: Pew Research Center, 2025).

So Google does make a selection. The problem lies elsewhere: where the summary appears, it displaces the links. And with AI Mode as the default interface, the surface on which that happens keeps growing. The search intent question comes up more often, not less.

The monetization hypothesis

Why is Google rebuilding search despite halved click rates? One explanation is close at hand, and it has to be labeled as a hypothesis. Public data cannot prove it, but it is plausible.

The hypothesis: AI answers create new digital real estate. Where ten blue links used to sit, there is now a block of text that Google controls. Ad formats that look like part of the answer can be embedded in that block. The longer users stay inside Google’s interface, the more surface there is to sell.

Two observations support this reading. First, Google’s track record: every major redesign of the results page in the past 15 years has increased the share of paid space. Second, the Q2 2026 numbers: revenue from “Search and Other” grew 17 percent, while Google cut the cost per AI Mode response to its lowest level since launch (Source: Alphabet, Q2 2026 Earnings Call, July 22, 2026). AI search is getting cheaper to run and brings in more revenue. That is not proof of intent, but it is a strong incentive.

AI search is getting cheaper to run and brings Google more revenue. Anyone who sees pure user-friendliness behind it underestimates the incentive.

The edge-case dilemma: wrong with full confidence

Language models calculate probabilities for the next word. They know nothing. For common questions with many good sources, that goes unnoticed. For rare questions, niche topics, or thin data, it becomes glaring. These are the edge cases.

The Tow Center for Digital Journalism at Columbia University measured how often AI search tools fail there. The researchers gave eight AI search tools excerpts from news articles and asked for headline, publisher, date, and URL. A task a traditional Google search solves within the first three results. The AI search tools answered more than 60 percent of the 1,600 queries incorrectly (Source: Tow Center, Columbia University, 2025).

The range is wide. Perplexity had a 37 percent error rate, Grok 3 94 percent. More than half of the responses from Gemini and Grok 3 pointed to fabricated or broken links (Source: Tow Center, Columbia University, 2025).

The real finding lies in the tone. ChatGPT gave 134 wrong answers and signaled uncertainty only 15 times. Not once did it decline to answer (Source: Tow Center, Columbia University, 2025). That is the edge-case dilemma: a wrong answer sounds exactly as confident as a right one. Users get no signal about when to be skeptical.

For companies, this has a direct consequence. A brand misrepresented in AI answers, with outdated prices, invented features, or wrong contacts, never hears about it. There is no results page left to monitor. Why inaccuracy is now considered the top AI risk is covered in Inaccuracy becomes the top risk.

The other side: where language models shine

The criticism targets one specific application: the automatic, unrequested insertion of AI answers into web search. It does not target the technology as such.

Language models can do something no traditional search can: they understand meaning instead of character strings. A question phrased with different words than the source document still finds its answer. Variants, paraphrases, and context are captured. That is real progress.

Google itself demonstrates this most clearly, with a different product. Gemini Notebook, known as NotebookLM until July 2026 (Source: Google, July 16, 2026), works only with documents users upload themselves. The model answers from a bounded corpus and points to the passage in the document. The difference to web search is fundamental: the data basis is known, limited, and reviewed. Edge cases become rarer because the model does not have to guess from the entire web.

This is the approach we use in AI Operations for companies: language models work on curated, proprietary data, with clear boundaries and traceable sources. The technology’s strength lies in understanding. Its weakness lies in guessing. Good systems keep the two apart.

Ready for the new search reality?

Google’s approach follows a familiar pattern of the US tech industry: ship, observe, fix later. AI Overviews arrived without a choice for users and website owners. Anyone who did not want to appear in AI answers had to leave search altogether.

Only regulatory pressure changed that. On June 3, 2026, the UK Competition and Markets Authority (CMA) required Google to give publishers and website owners control over the use of their content in AI search features (Source: CMA, June 3, 2026). Since August 31, 2026, that control has been available worldwide in Google Search Console (Source: Google Search Console Help, 2026). Based on everything publicly known, it would not have arrived without regulation.

Whether AI search is the better model for web search calls for a nuanced answer. So far, click rates and error rates argue against it. The decision is not made in conference rooms but in the search behavior of billions of people, and the data on that exists only in fragments so far. On top of that, the technology keeps evolving.

For brands, this uncertainty is no reason to wait. It is a reason to measure. Anyone who does not know whether and how their brand appears in AI answers is flying blind.

To make exactly this visibility measurable, we at netzstrategen have built the AI Search Index. It shows in which answers of which AI search tools a brand appears, how it is portrayed, and which sources the models cite instead. The black box becomes a metric. What the index can deliver for a specific brand is something we clarify with you in a free diagnosis call.

Conclusion: measure instead of waiting

AI search is neither the promised savior nor pure hype. It is a rebuild of the internet’s most important artery that does two things at once: it makes search better for some questions and worse for others.

Three things are certain. Where AI Overviews appear, clicks to websites halve (Source: Pew Research Center, 2025). AI search tools are frequently wrong and sound confident while they are (Source: Tow Center, Columbia University, 2025). And Google keeps expanding the reach of its AI search, with more than one billion AI Mode users (Source: Alphabet, Q2 2026 Earnings Call, July 22, 2026).

No decision for or against AI search follows from this. A task does: know and measure the brand’s own visibility in AI answers, and derive measures from it. Everything else is hope.

Frequently asked questions about AI search

What is the difference between AI Overviews and AI Mode?

AI Overviews are automatically generated summaries above the traditional search results. AI Mode is a chat interface in which the entire search runs as a dialogue. Since summer 2026, Google has merged both into one search experience (Source: Alphabet, Q2 2026 Earnings Call, July 22, 2026).

Do websites really lose traffic to AI search?

The independent data says yes. According to Pew Research Center, users clicked on a result in 8 percent of searches with an AI Overview and in 15 percent without one (Source: Pew Research Center, 2025). Google counters that AI features generate “billions of clicks every week” (Source: Alphabet, Q2 2026 Earnings Call, July 22, 2026). Both can be true: more searches overall, fewer clicks per search.

How reliable are AI search tools?

Mostly good for common questions, poor for rare ones. The Tow Center at Columbia University found more than 60 percent incorrect answers across 1,600 test queries on eight AI search tools, usually without any signal of uncertainty (Source: Tow Center, Columbia University, 2025).

Can companies keep their content out of AI answers?

For Google search, yes, worldwide since August 31, 2026: Search Console offers a control for AI Overviews, AI Mode, and AI features in Discover, with no effect on traditional ranking (Source: Google Search Console Help, 2026). Whether opting out makes sense depends on the business model. For most brands, visibility in AI answers is worth more than protection from it.

How do we measure whether our brand appears in AI answers?

With systematic test queries across several AI search tools, evaluated by mention, portrayal, and cited sources. That is what the netzstrategen AI Search Index does. We clarify the starting point in a free diagnosis call.

Sources

• Pew Research Center, 2025: Google users are less likely to click on links when an AI summary appears in the results, July 22, 2025
• Tow Center for Digital Journalism, Columbia University, 2025: AI Search Has a Citation Problem, Columbia Journalism Review, March 6, 2025
• Alphabet, 2026: Q2 2026 earnings call: Remarks from our CEO, July 22, 2026
• Similarweb, 2026: The 2026 Generative AI Landscape
• Google, 2025: AI, personalization and the future of shopping, March 3, 2025 (“more than 5 trillion searches on Google annually”, roughly 14 billion per day)
• Google, 2023: Search Quality Rater Guidelines, update note on the Google Search Central Blog
• Google, 2026: NotebookLM is now Gemini Notebook, July 16, 2026
• Competition and Markets Authority (CMA), 2026: CMA secures fairer deal for publishers and improves Google search services in UK, June 3, 2026
• Google Search Console Help, 2026: Search generative AI control