Will answer engines replace search? Four builders on the web's future

The Short Version

Three of the four builders expect direct answers to take over much of what a page of links does today, but none of them predicts that search simply disappears, and they disagree about what replaces it and who pays [1][2][3][4]. Marc Andreessen thinks search as we know it will probably give way to answers, Aravind Srinivas has built a company on that bet while conceding that Google is still better for some queries, Sam Altman calls beating Google the wrong goal, and Jimmy Wales expects Wikipedia to look much the same in ten years with AI layered on top [1][2][3][4]. The likeliest losers they identify are the link-click advertising that funds Google's best business, the people who make web pages and other source material, and any knowledge source whose funding depends on readers knowing where an answer came from [1][2][3][4]. The sharpest disagreement is over advertising and truth: whether an answer can be paid for by ads and still be trusted, and whether a system that sometimes makes things up can stand in for a list of sources [1][2][3][4]. All four conversations date from a fast-moving period, so the products and business models described may have changed [1][2][3][4].

Table of contents

What They Said

Marc Andreessen: the internet becomes the content of the next medium

Andreessen, who co-created the Mosaic browser and co-founded Netscape, gives the broadest prediction [1]. He thinks search as we know it will probably give way to direct answers, and points out that Google itself has been trying to move beyond its ten blue links for years [1]. His frame comes from Marshall McLuhan: "the content of each new medium is the old medium" [1]. The internet absorbed television, radio and books; AI, in turn, will treat the internet as its content [1]. Links do not vanish in this picture. An AI can regenerate ten blue links or cited sources whenever a user wants them [1]. He adds that a neural network is itself loosely a search process over compressed knowledge, so the line between searching and answering is thinner than it looks [1].

He is the one who names the supply problem. If people stop making web pages because nobody visits them, a major source of AI training data dries up [1]. Whether synthetic training data adds real signal or is empty calories is, in his description, a trillion-dollar question [1]. He also notes an odd property of the new medium: jailbroken chatbot personas now live in the training data, so each new model can bring them back [1].

On accuracy he refuses the usual framing: "hallucination is what we call it when we don't like it. Creativity is what we call it when we do like it." [1] A big law firm told his team it wants creative models to explore legal theories and then checks citations separately [1]. He lists possible fixes, including fact-checking plugins and one model checking another, while noting that Yann LeCun argues the problem is unsolvable [1]. His benchmark is Wikipedia: not perfect, but more likely to be right than the alternatives [1].

He raises a deeper worry about who decides what an answer engine says. He passes on a friend's thought experiment: had a model been trained on the consensus of Galileo's time and tuned by the church's feedback, would it have declared him wrong? [1] He argues for humility and suspicion of anyone claiming a monopoly on truth [1]. And he flags a legal unknown, that AI output may not be protected by Section 230 [1]. Readers should check the current legal position, since that was an open question when he spoke [1].

Aravind Srinivas: change the interface, not the ranking

Srinivas, the co-founder and CEO of Perplexity, is the one actually building the thing [2]. He calls it an answer engine: you ask a question and get an answer in which every part is backed by sources, "like how an academic writes a paper" [2]. The system runs a traditional search, reads the relevant links, extracts paragraphs and hands them to a language model instructed to write a concise answer with footnotes [2]. The rule he sets for the model is strict: "you're not supposed to say anything that you don't retrieve", and if there is not enough information it should say so [2].

The idea came from a practical failure of search. When the founders' first employee asked for health insurance, none of them understood the terms, and Google was little help; Srinivas's explanation is that insurance is a major advertising category, so there is no incentive to give a clear answer [2]. A chatbot they built sometimes said incorrect things, so they asked what would happen if it were forced to cite everything, and realized that is essentially how Wikipedia works [2].

His strategy is explicitly not to fight Google at its own game. Perplexity never tried to build a better ten-link search engine with a gimmick like privacy or no ads, because Google has nailed that for 20 years [2]. The disruption is the interface: why should links occupy the prime real estate? [2] Early on the team debated keeping links in a side panel in case answers went wrong, and instead bet that models would get better and cheaper, the index fresher and errors rarer [2].

He admires what he is attacking. He calls Google's AdWords auction "the greatest business model in the last 50 years" [2]. That admiration is the point of the strategy, which he takes from The Art of War: "make the weakness of your enemy a strength" [2]. Any answer format that reduces clicks on links eats into Google's highest-margin business, so Google has little reason to push it hard [2]. He compares it to Amazon building cloud computing first because the margins were worse than Google's advertising but better than Amazon's retail, and quotes Jeff Bezos: "Your margin is my opportunity." [2]

He is also the most careful about limits. A live query during the conversation produced an answer conceding that Google is still faster and better for navigational queries and for real-time information such as sports scores, and he agrees [2]. He says Google itself is not in danger, since its cloud and YouTube businesses alone run at about $100 billion a year, and that this is not a zero-sum game [2]. He notes that instant answers already accounted for a large share of Google's traffic in the 2010s, and he frames search as one chapter in a longer story that ran from Yahoo's categories to Google's links to answers [2].

For him the answer is a starting point. "In my opinion, the journey begins after you get an answer", he says, which is why each answer ends with suggested follow-up questions [2]. He thinks the real obstacle is not Google but that "people are not naturally good at asking questions" [2]. He takes from Larry Page the principle that "the user is never wrong", and concludes that "a better product should be one that allows you to be more lazy, not less" [2].

Sam Altman: a better Google is the wrong goal

Altman, the CEO of OpenAI, is asked directly whether his company will take on what Google started 20 years ago. His reply: "I find that boring" [3]. If someone can build a better search engine, people should use it, he says, but that would understate what the technology can be [3]. He describes the incumbent this way: Google shows ten blue links, or in his telling thirteen ads and then ten blue links, and "that's like one way to find information" [3]. What excites him is a much better way to help people find, act on and synthesize information, sometimes by creating the answer, sometimes by pointing to it [3].

He does not claim it has been done. Many people have tried to out-search Google, and "it is a hard technical problem, it is a hard branding problem, it's a hard ecosystem problem" [3]. Combining a chat client with a search engine is, he says, "cool, but it's tricky", because simply shoving one into the other is awkward [3]. Of the intersection of language models and search he says, "I don't think anyone has cracked the code on yet" [3]. That was said while Srinivas's product already existed, so it reads as a judgment that the existing attempts were not yet the answer [2][3].

He does see behavior changing. Younger users, he notes, treat ChatGPT as the default starting point for any knowledge task [3]. He expects made-up answers to improve but not to be solved within the year he was speaking [3].

On advertising he is the most hostile of the four. "I kind of hate ads just as like an aesthetic choice", he says, adding that ads needed to happen to get the internet going but "it's a momentary industry" [3]. He likes that people who pay know the answers they get are not influenced by advertisers [3]. He concedes "I'm sure there's an ad unit that makes sense for LLM's", and then describes the version he fears, in which you ask a question and are nudged toward a product [3]. Of ad-supported products in general he says, "I think it gets worse, not better in a world with AI" [3]. He believes OpenAI can pay for its computing without ads, though he does not fully rule them out [3].

Jimmy Wales: the knowledge is a gift, the funding is not guaranteed

Wales, the co-founder of Wikipedia, speaks for a source that answer engines draw on [4]. He is relaxed about the main change. If people get Wikipedia's knowledge through assistants he is untroubled, because the content is a gift to the world [4]. The worry sits with the foundation that runs it: donations could suffer if users never learn where the knowledge came from [4]. Its most effective fundraising pitch has been simple fairness, that you use it all the time and should chip in, which depends on people knowing they use it [4].

He is blunt about the technology's flaw. He describes a model inventing three entirely fictitious but plausible examples when he asked a question, and says it "doesn't really know or care about the truth" [4]. He thinks chatbots should report how much of an answer is unreliable [4]. He is firm that proper attribution is simply good intellectual practice and is encouraged by work on grounding answers in quality sources [4]. A demo by Wikimedia's head of machine learning, which retrieves relevant articles and answers only from them, suggests to him a better on-site search experience [4].

His forecast is modest. In ten years he expects Wikipedia to look similar, with AI-assisted search and large growth in developing-world languages helped by better machine translation [4]. He sees AI as a tool for editors: flagging biased wording, checking whether an article reflects its sources, helping non-native speakers [4].

He also brings a rule about money that bears on the whole debate: "business model drives outcomes" [4]. Avoiding advertising was at first almost an aesthetic choice for Wikipedia, but he now sees it as decisive, because organizations follow the money in the long run [4]. Without ads there is no reason for an algorithm to steer a reader somewhere more profitable and no suspicion that a company paid for its entry [4].

Where They Agree, and Where They Don't

Agreement: the list of links is losing its place as the default. Andreessen predicts it, Srinivas builds on it, and Altman reports younger users already starting elsewhere [1][2][3]. Even Wales expects AI-assisted search on his own site [4].

Agreement: Google is not about to vanish. Srinivas says so outright and points to its other businesses [2]. Altman lists how hard it is to dislodge [3]. Andreessen notes Google has been moving toward answers itself [1].

Disagreement: is this a replacement or a different product? Andreessen speaks of search giving way [1]. Srinivas keeps a search engine under the hood and says the new format is better for some questions and worse for navigational and real-time ones [2]. Altman rejects the comparison, saying the aim is a broader way to find, act on and synthesize information [3]. Their positions track what each is building: an investor describing a shift of medium, a founder who needs a wedge against an incumbent, and a lab head who sees search as one use of a general system [1][2][3].

Disagreement: can ads coexist with trusted answers? Altman thinks ad support gets worse with AI and prefers that users pay [3]. Wales holds that refusing ads is what keeps a knowledge source honest [4]. Srinivas is more open. Perplexity has subscription revenue and does not need ads today, he says, but he allows for future ads provided they do not cost user trust or compromise truthful answers, and he likes how some targeted ads barely feel like ads [2]. He also praises the ad auction as a business [2]. Altman's own hedge, that some ad unit probably makes sense, shows the question is not closed even for the skeptic [3].

Disagreement: is making things up a defect? Andreessen calls it creativity under another name and describes a customer who wants it [1]. Srinivas designs against it, restricting the model to retrieved text, and lists four ways it still happens: the model misreads the query or passages, the snippets are stale or thin, irrelevant detail confuses it, or the wrong documents are retrieved [2]. Wales treats it as indifference to truth [4]. Altman expects improvement, not a cure [3]. They are answering for different uses: brainstorming a legal theory tolerates invention, and an encyclopedia does not [1][4].

Disagreement: are citations the answer? Srinivas makes them the product [2]. Andreessen treats them as something the AI can produce on request [1]. Wales supports attribution but supplies the warning: a magazine once printed a false detail about him that came from vandalism on his own Wikipedia page, an example of a citation loop in which an error becomes its own source [4]. A footnote shows where a claim came from, not that it is true.

Why It Works

The incumbent's margin is the opening. Srinivas's argument is economic. Google earns its highest margins when people click links, and an answer that satisfies the user without a click reduces that revenue, so the company best placed to build answer engines has the least reason to [2].

Retrieval plus a language model. The answer engine works by retrieving documents and having a model write from them, and Srinivas notes that language models help because they tolerate imperfect ranking, finding the right link even if it is ninth [2]. He adds that real retrieval still needs old-fashioned term matching alongside newer methods, plus signals like domain authority and recency [2]. Wales's on-site demo uses the same pattern [4].

New media absorb old ones. Andreessen's McLuhan argument explains why links need not die: they become material the new medium draws on and can reproduce [1].

Money shapes the answer. Wales's rule that the business model drives outcomes, Srinivas's insurance story and Altman's fear of product nudges inside answers are three versions of one mechanism: whoever pays influences what is shown [2][3][4].

People want less work. Srinivas's claim that a better product lets the user be lazier explains the pull toward answers, and his view that curiosity is unbounded while skill at asking is limited explains suggested follow-ups [2].

Where It Applies, and Where It Doesn't

Answers fit questions that need synthesis across sources; Srinivas himself grants that links and custom displays win for going to a known site and for live information [2].

The supply side is unresolved. Andreessen's point that fewer web pages mean less training data, and Wales's point that unseen sources lose donors, describe the same risk from two ends, and none of the four offers a tested fix [1][4]. Altman says people who create valuable data deserve some way to be compensated and that the answer to that question is yes, but admits he does not yet know the model [3]. Wales adds that Wikipedia's own reliance on journalism is already strained by the decline of local media and the rise of clickbait [4].

Answer engines can be gamed. Srinivas calls it answer engine optimization, says simple hidden-text tricks are easy to fix, and describes defense as a partly reactive cat-and-mouse game, as it was for Google [2].

Consensus is not truth. Andreessen's Galileo question applies to any system that summarizes prevailing sources [1].

The sources are also uneven. Srinivas and Altman run companies with a direct stake in the outcome and Andreessen co-founded a venture firm, and only Altman's conversation is represented here by verbatim transcript passages; the others come from summaries [1][2][3][4]. Statements about who has or has not adopted ads, which models are used and what the law says reflect the dates of the conversations, so check current terms before relying on them [1][2][3].

What To Do With It

1. Match the tool to the question: use an answer engine for questions that need synthesis, and ordinary search or a dedicated app for reaching a known site or getting live scores and similar real-time facts [2].
2. Open the citations. Treat anything in an answer without a source as opinion, the standard Srinivas took from academic writing, and remember Wales's example of an error that became its own source [2][4].
3. Keep going after the first answer. Use the suggested follow-up questions or write your own, since Srinivas designs the answer as the start of the inquiry [2].
4. Find out how the tool is paid for before trusting it on a purchase or a contested topic. Altman and Wales both argue that paying users or donors are better protected than an audience sold to advertisers [3][4].
5. For anything that matters, check facts separately, as the law firm in Andreessen's account does with citations [1].
6. On disputed questions, ask the model to argue each side hard. Andreessen says he stages deliberately heated debates this way because the model otherwise defaults to making both sides agree [1].
7. If you publish online, expect answer engines to read your pages on behalf of users, press for clear attribution, and watch the unsettled question of compensation that Altman acknowledges [2][3][4].
8. If you rely on Wikipedia through an assistant, remember where the knowledge came from and consider chipping in, since Wales says donations depend on that awareness [4].
9. If you are building a product, do not copy the incumbent's format. Srinivas's method is to find the improvement the incumbent's own margins discourage it from making [2].

Sources

1. Marc Andreessen on the Future of the Internet, Why AI Will Save the World, and the Birth of the Browser · Lex Fridman Podcast
2. Aravind Srinivas: Perplexity CEO on Future of AI, Search & the Internet · Lex Fridman Podcast
3. Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI · Lex Fridman Podcast
4. Jimmy Wales on Building Wikipedia, Neutrality, Community Health and the Business of Trust · Lex Fridman Podcast