The End of Google? How AI Search Is Rewriting the Rules for Writers, Readers, and the Open Web

The Quietest Disruption in Tech Is Happening to Search

If you asked a product manager in 2022 which tech category was least likely to be disrupted, "web search" would've been near the top. Google had roughly 90% global market share, an $80B-a-year ad business, and two decades of index depth nobody else could match.

Then, in about eighteen months, the floor shifted.

Google launched AI Overviews at I/O in May 2024, starting in the US, and expanded to 100-plus countries by 2025. A dedicated "AI Mode" tab now behaves less like search and more like a chatbot with web access. OpenAI launched ChatGPT Search on October 31, 2024 (Plus and Team first, free tier in early 2025), after demoing SearchGPT in July 2024. Sam Altman called it an "answer engine." In late 2025 OpenAI released the Atlas browser, wrapping ChatGPT around the browsing experience itself.

Perplexity was handling roughly 500 million queries per month by mid-2025, raised at a $9B valuation in late 2024, and reportedly $18B by mid-2025. Anthropic's Claude added web search in March 2025 for paid users. Microsoft's Copilot kept iterating on Bing. Brave, You.com, Kagi, and Arc Search all shipped variants.

It's the biggest shift in how people find information since PageRank. And yet, the most honest description is still "quiet." There's no iPhone moment. Most users slipped in gradually: a ChatGPT tab replaced a Google tab, an AI Overview made page-one results feel redundant, a Perplexity link showed up in a group chat. One query at a time, default habits moved.

For Glasp and the broader community of curious readers and writers, this is the most important platform shift of the decade. The rules for how knowledge reaches people, and how the people who produce it get paid or credited, are being rewritten in real time.

---

From Ten Blue Links to One Paragraph Answer

The classic Google result page had a distinct contract. You typed a query, Google returned ten-ish blue links plus ads and widgets. You clicked a link, arrived on a site, and the publisher had a chance to show you its writing, its design, its ads, its newsletter signup. The publisher got a visit. You got context. Google got an ad impression.

AI search collapses that. You type a question. You get a paragraph. The paragraph usually has footnote-style citations, but the whole interaction is designed to end at the answer. You don't need to click. You often don't want to.

This is genuinely useful for a lot of queries. "Capital of Paraguay." "Convert HEIC to JPG on a Mac." "Difference between RSA and ECDSA keys." For questions where the right answer is short, stable, and already written down in a hundred places, AI search is a strict UX upgrade: fewer clicks, fewer SEO-farmed listicles, fewer cookie banners.

It's also useful for research-style questions that used to require eight tabs. "Compare the side effects of GLP-1 drugs." "History of Japanese bullet trains in a paragraph." The AI does the tab-opening for you.

The trade-off is that the paragraph is the destination, not a waypoint. The web that produced the underlying material is reduced to cited raw inputs. That's a different contract, and it's one most publishers never signed.

---

What the Data Actually Shows on Traffic

The debate about whether AI search "kills" publishers is noisy, so let's stick to what's measured.

Zero-click was already winning. Pew Research's 2025 analysis showed that about 60% of Google searches ended without a click to any non-Google destination, and that figure rose when an AI Overview was present. SparkToro's Rand Fishkin has been tracking this trend for years and noted that AI Overviews accelerated a curve that started long before them. The internet was already drifting toward answers-on-Google via featured snippets and knowledge panels. AI Overviews were the next step, not the first.

CTR on top organic positions is down meaningfully for informational queries. Similarweb and third-party studies through 2024-2025 reported 30-60% click-through drops on the top organic result when an AI Overview was attached. Raptive and Mediavine, which represent thousands of mid-size publishers, publicly warned that the changes posed an existential threat. Commercial and transactional queries held up better; how-to and definitional content got clobbered.

Citation is not referral. Ahrefs analyses in 2024-2025 showed that even when pages were cited inside an AI Overview or ChatGPT Search answer, click-through didn't meaningfully rebound. A user who's satisfied by the answer rarely clicks the footnote. The industry narrative ("AI search cites sources, so publishers are fine") conflates attribution with traffic.

News discovery is shifting too. The Reuters Institute's Digital News Report 2025 found the share of respondents discovering news via Google was declining, while the share using AI tools was rising year over year. The absolute numbers are still small for AI, but the trend line is what matters.

Here's a simplified view of the CTR picture drawn from published analyses:

Two honest caveats. First, every study uses slightly different methodology, so exact numbers vary. Second, Google itself disputes some of the more dramatic claims, arguing AI Overviews send "quality traffic." Both can be true: fewer clicks in aggregate, better qualified when they happen. Publishers running the actual dashboards are not reassured.

If you're a writer who relies on search as a distribution channel, the takeaway isn't "panic." It's "diversify the channel mix now, not later." We wrote about the underlying economics in The Knowledge Creator Economy.

---

The 2024-2026 AI Search Stack

The ecosystem moved from "one dominant product" to five or six serious ones in eighteen months. The lay of the land as of April 2026:

A few patterns stand out.

Every major product cites sources, but the citation UX varies wildly. Perplexity treats citations as first-class UI. ChatGPT Search shows them but buries them. Google's AI Overviews cite, but the citation is secondary to the answer block. None of them have solved attribution at the level a publisher ad business would want.

All of them are converging on a similar feel: conversational, iterative, follow-up-friendly. The "one query, one result page, retry with different keywords" dance is dead. The replacement is "one question, one answer, refine with a follow-up."

And all of them need the open web. The AI doesn't know anything that wasn't written somewhere by a human first. That's the leverage point publishers still have, and it's one we return to in The Human Curator in the Age of AI.

---

The "Hallucinated Answer" Problem

AI search's biggest weakness is the same weakness every LLM has: it's confidently wrong, sometimes in public.

The most famous early failure was Google AI Overviews suggesting non-toxic glue on pizza to keep cheese from sliding off (The Verge, 404 Media, May 2024). The "recipe" was traced to an 11-year-old Reddit joke. Other Overviews that week told users to eat rocks for minerals and misattributed cancer advice. Google tightened the guardrails, but the damage to trust was real.

Hallucination isn't just embarrassment. It's real-world harm.

• Legal: In Mata v. Avianca (2023), attorneys submitted a brief citing cases ChatGPT had invented. The court sanctioned them. AI-assisted legal research without verification is now a known malpractice risk.
• Medical: Stanford HAI and multiple JAMA studies through 2024-2025 found general-purpose LLMs gave incorrect or outdated medical advice at meaningful rates, especially for rare conditions and drug interactions.
• Finance: AI-generated summaries of 10-K filings and earnings calls have confused numbers, dates, and even which company is being discussed.

AI search products are adding retrieval (RAG), model-level fact-checking, and "I'm not sure" behavior. But the underlying issue is architectural: an LLM generates the most plausible next tokens, not the most true ones. Retrieval narrows the gap; it doesn't close it.

For readers, the lesson is boring but important: don't trust any single AI answer for anything that matters. Click through. Compare.

For writers, the ironic silver lining: accurate, specific, verifiable writing is more valuable now, because it's the scaffolding the AI has to cite.

---

GEO Is the New SEO (Sort Of)

GEO, short for Generative Engine Optimization, went from near-zero search volume in 2023 to mainstream practice by 2025. It's the discipline of making your content more likely to be read, ingested, and cited by LLM-based answer engines.

The early playbook:

1. Write in clean, self-contained chunks. AI systems retrieve passages, not pages. Short paragraphs, clear topic sentences, and explicit definitions help a retriever grab your text without surrounding confusion.
2. Be specific and verifiable. Numbers, dates, named sources, and concrete examples are citation bait. Vague opinions aren't.
3. Answer the question in the first two sentences, then elaborate. AI engines prefer content where the "answer" sits near the top of a section.
4. Keep schema and metadata tight. Structured data (FAQPage, Article, HowTo) still helps retrieval pipelines understand structure.
5. Mention your own name. Attribution works better when your brand is inside the text. "According to Glasp's analysis of reading habits..." survives ingestion better than an anonymous paragraph.

The limits are real. You can't buy rank; there's no AdWords equivalent yet. Citations may not drive clicks. Each engine weights differently, so what Perplexity retrieves isn't what ChatGPT Search or Google's AI Mode retrieves. And prompt injection and page poisoning are real, growing threats.

The honest frame is that GEO is maybe 40% "new thing" and 60% "classic content quality and structure." The biggest GEO advantage in 2026 is still the oldest advantage in publishing: write something true, specific, and actually useful. For more on that mindset, see Learning in Public.

---

What Publishers Are Actually Doing

Publishers aren't sitting still. Roughly four strategies are visible in 2026:

1. Licensing deals. OpenAI has signed content agreements with the Financial Times, News Corp, Vox Media, The Atlantic, Axel Springer, and Condé Nast, among others (deals announced 2023-2025). Google has its own set of deals. Perplexity launched a publisher revenue-share program in 2024. Good news: real money is flowing. Bad news: the deals heavily favor large incumbents. Small and mid-size publishers are mostly not on the list.

2. Walled gardens and paywalls. More content is moving behind logins and paywalls partly to make it inaccessible to crawlers. Reddit started charging for API access in 2023. Publishers are adding selective access for AI crawlers via robots.txt, llms.txt, and server-side filtering.

3. Direct audience channels. Newsletters (Substack, Beehiiv, Ghost), podcasts, YouTube, and private communities are the obvious hedges. If search can't be trusted to deliver your readers, you build channels where you own the list. The Verge's 2024-2025 pivot to newsletter and Discord was explicit about this.

4. First-party products. Tools, data, courses, memberships. The NYT Games division (Wordle, Connections, Spelling Bee) is the poster child, now material to NYT subscriber retention.

Most serious publishers are doing some combination of all four. The common thread is "reduce dependence on any single platform." That's a lesson writers at every scale can internalize. A personal newsletter, a community where your readers actually talk to you, a small paid tier of deeper work, these add up.

---

What Readers Gain and What They Lose

It's easy to frame AI search as purely bad. It isn't.

What readers gain:

• Speed. A good AI answer saves 3-5 minutes on a research question that used to mean opening tabs, skimming, synthesizing.
• Less SEO sludge. The worst of affiliate-farm content ("10 Best Kitchen Knives of 2024, #3 Will Shock You") is being bypassed. For many commercial queries, AI summaries are simply more useful than the SEO pages they replaced.
• Lower friction for non-native English speakers. AI answers adapt to language and reading level. That's a real accessibility win.
• Better synthesis across sources. For "compare X across five perspectives" questions, AI is strictly faster than DIY.

What readers lose:

• Serendipity. You don't stumble into a great blog you'd never heard of. The long tail of the web gets harder to reach.
• Source diversity. AI answers blend sources into a single voice. You lose the texture of writers disagreeing with each other.
• Confidence calibration. The AI's tone is uniformly confident. Real experts hedge. Flattening that into a paragraph distorts the epistemic picture.
• The skill of searching. Being good at queries and source evaluation is a real skill. Outsource it fully and it atrophies. More on that in The Information Diet.

A healthy reading practice in 2026 uses AI search for quick lookups and low-stakes research, but reserves deeper reading, highlighting, and note-taking for what actually matters.

---

What the Open Web Needs Next

The open web isn't going away. It's being repriced. The most interesting design question of the next few years: who owns the reader's relationship to the content?

Three things need to be true for the open web to stay healthy:

1. Credit has to turn into compensation, or at least referral. Right now, citations are cheap for AI engines and expensive for publishers. The incentive is wrong. Better attribution UX, revenue sharing, and browser-level referral tracking would all help. Some of this is regulatory (EU AI Act, copyright law updates), some voluntary (Perplexity's publisher program), some technical (llms.txt, provenance standards like C2PA).

2. Readers need a portable, reader-owned layer. Today your highlights, notes, and saved articles live scattered across Kindle, Notion, Readwise, ChatGPT history, and a dozen other silos. If the AI owns your reading history, it owns the leverage. If you own it, you can feed it to any AI you want. This is the thesis behind Glasp's web highlighter: highlights you make while reading become your own knowledge asset, not an AI company's training data.

3. Writers and readers need direct channels. Communities, newsletters, and small paid memberships are the layer AI search doesn't eat. Not because AI can't summarize them, but because the relationship is the product.

Glasp's own angle is specific. The most durable version of the knowledge web is one where readers highlight, curate, and re-share, and where every reader is a little bit of a publisher. Glasp's AI chat runs against your own highlights, not someone else's opaque model. That's consistent with the broader idea of an AI reading assistant: the useful future isn't "AI replaces the web." It's "AI helps you do more with the web you already care about."

---

Frequently Asked Questions

Is Google really dying?

No, but the shape is changing fast. As of early 2026 Google still has roughly 85-90% global search share by query volume. What's declining is the share of those queries that end in a click to an external website. Google isn't going away; Google-as-referral-to-the-open-web is shrinking.

Which AI search engine is the most accurate right now?

There's no single winner. In benchmarks through 2025, Perplexity and Claude web search ranked high for citation quality, ChatGPT Search for conversational depth, and Google's AI Mode for breadth and freshness. All of them hallucinate sometimes. Treat any single answer as a draft.

Does getting cited by an AI engine send real traffic?

Usually not much. Ahrefs studies in 2024-2025 showed citation didn't correlate strongly with click-through. Users who get a satisfying answer rarely click the footnote. Citation is nice for branding, but publishers who budget around it get disappointed.

Should I still do SEO in 2026?

Yes, alongside GEO, direct audience work, and first-party products. Classic SEO still drives real traffic on commercial, branded, and local queries. Informational content has been hit hardest. Shift your mix rather than abandoning any channel.

How do I optimize for AI search without being crawl-bait?

Write clear, specific, self-contained content. Keep paragraphs under roughly 80 words. Answer the question in the first two sentences. Use real numbers and named sources. Keep schema markup clean. Include your brand name inside the text so attribution survives ingestion.

Is it ethical for AI engines to summarize my writing without paying me?

Active legal and ethical debate. Lawsuits from The New York Times (vs. OpenAI, 2023), Getty Images (vs. Stability AI), and authors' groups are still working through the courts in 2026, with jurisdictions landing differently. If you're a publisher, check licensing options, set robots.txt and llms.txt deliberately, and lobby for better attribution standards.

Will small publishers survive?

Some will, some won't. Survivors will almost certainly own a direct channel, have a distinctive voice or specialty, and treat search as one channel rather than the channel. The era of building a business purely on Google organic traffic is probably over. The era when a small, specific publication with 5,000 true readers can sustain a writer is very much alive.

---

Conclusion: The Search Is Dead. Long Live the Answer.

Search didn't die, exactly. It got abstracted one level up.

For twenty-five years, the job of a search engine was to point you at pages that might answer your question. The job of the pages was to actually answer it. The whole web we know was built on that split. AI search merged the two jobs into one box. It's a better product for a lot of questions. It's a worse deal for the people who used to be on the other side of the click.

That's the tension to sit with. Readers got something legitimately good. Writers lost something legitimately important. Neither side is wrong about their experience.

If you're a reader, use AI search where it shines, but keep the habit of reading full sources for what you actually care about. Keep a place where your highlights, notes, and favorite writers accumulate over time. Don't let the answer layer be the only layer.

If you're a writer, don't chase the AI engines the way you used to chase Google. Write for humans first, structure for retrieval second, and build at least one channel where you own the relationship. The durable asset isn't your rank. It's your readers' attention, on purpose.

The open web was built on a handshake. The handshake changed. The work continues.