Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

1. The 2026 AI Search Map — Answer engines / Paid search / AI browsers

Between 2024 and 2026, the word "search" changed meaning. OpenAI announced SearchGPT in July 2024, ChatGPT's Browse mode matured the same year, and "a page of ten links" stopped being the default surface for a query. As of May 2026 the market splits into three currents.

First, **answer engines**: Perplexity, ChatGPT Search, Claude Web's browse, You.com. Instead of a search engine results page (SERP), they return a single paragraph with citation chips. You stop clicking ten links and start verifying one synthesized answer.

Second, **paid search and independent indexes**: Kagi ($10/month), Mojeek, Marginalia, Stract, Brave Search. They reject the ad model — you pay directly, or they fund themselves through API licensing — and they differentiate on result quality and privacy.

Third, **AI browsers**: The Browser Company's Arc Search and its successor Dia (launched May 2025), SigmaOS, Brave + Leo, Edge with Copilot. Search lives inside the browser — in the sidebar, the address bar, the new tab — not at a separate site. Arc Search's old tagline "Browse for me" captures the mindset.

The three currents overlap. Perplexity makes both a mobile app and a desktop browser (Comet). Brave bundles its own search engine and its own AI inside a single browser. But the question "where do you type the first query of your day?" still splits users three ways.

Google and Bing live in the middle of this earthquake. Google bolted AI Overviews on top of the SERP; Bing fused Copilot into Edge. This article focuses on the **AI-first newcomers**, not on retrofits by the existing giants.

2. Perplexity — the answer-engine pioneer

Perplexity AI was founded in August 2022 by Aravind Srinivas (ex-OpenAI) and team. Early demos were a thin webapp that synthesized a paragraph answer with citations. By 2024–2025 the company had raised a Series C and shipped a mobile app, a desktop browser called Comet, a Pro subscription, and a wiki-like "Pages" feature for sharable answers.

Perplexity's signature is **citation chips**: every paragraph ends with small numbered references; clicking a number opens the source URL. That pattern then propagated to ChatGPT Search, Claude's browse, Brave Leo, and Google's AI Overviews. Perplexity normalized the idea that "no citation, no trust."

The product splits into "Pro Search" (a few seconds, one paragraph) and "Deep Research" (5–15 minutes, report-length synthesis across dozens of pages). Pro users can switch the underlying model among GPT-class, Claude-class, and Perplexity's own Sonar.

On the browser side, Perplexity shipped **Comet** — a Chromium-based browser with the Perplexity answer sidebar callable on every tab. "Summarize this page," "compare this price across other sites" are first-class verbs. The mindset overlaps with Arc Search / Dia, but Perplexity uniquely owns a search index.

Revenue is a mix of ads and subscriptions. Pro at $20/month unlocks model selection and unlimited Deep Research; Enterprise adds SSO and data isolation. A 2025 ad rollout sparked some backlash but settled into small "sponsored" cards next to answers.

One-liner: **the first name that comes to mind when you want an answer, not links** — but always click a citation if the stakes matter.

3. Kagi — paid search ($10/month) with strong privacy

Kagi is an **ad-free paid search engine** founded in 2018 by Vladimir Prelovac. It opened in public beta in 2022 and entered mainstream awareness around 2024. In 2026 it sells Starter at $5/month, Professional at $10/month (unlimited searches), Ultimate at $25/month (heavier AI features).

Kagi's identity is **"I pay, so no ads and no tracking."** Without ad inventory it has no incentive to skew results toward advertisers. Users personalize via **lenses** — domain-level rules that pin, boost, downrank, or block specific sites. A "Programming" lens that boosts GitHub and downranks Pinterest is a common starter.

Technically Kagi is a hybrid: its own crawlers (Teclis, Tinygem), partial own index, and licensed APIs from Google/Bing. That makes it different from Mojeek's 100% own-index approach and different from meta-engines that only proxy Bing.

AI features arrive as "Quick Answer" (an AI paragraph above results) and "Kagi Assistant" (multi-model chat: Claude, GPT, Gemini, etc.). Ultimate ships unlimited Assistant.

On privacy, Kagi promises not to use queries for ads and offers anonymous-token sign-in. The hard truth is that no provider with a payment relationship can mathematically prove zero linkage — Kagi sits in the "trust" zone, but a credible one.

One-liner: **the most serious option for "I'm sick of ads and want to tune result quality myself"** — if $10/month is acceptable.

4. ChatGPT Browse + SearchGPT (July 2024) — OpenAI's answer

OpenAI announced a standalone **SearchGPT** prototype in July 2024. In parallel they upgraded ChatGPT's "browse with Bing" path. By 2025 a "Search" toggle below the ChatGPT input fused both into one surface. As of 2026 the SearchGPT brand is essentially the search mode inside ChatGPT.

ChatGPT Search combines **OpenAI's models + OpenAI's own crawler + licensed publisher content**. OpenAI signed deals with Associated Press, Axel Springer, News Corp, Vox Media, Condé Nast and others, so the model can cite and summarize their pages directly. That contrasts with Perplexity's earlier public friction over citation usage.

The UX is simple. Toggle the "Search" icon in ChatGPT, and before answering the model retrieves pages through OpenAI's search infrastructure and uses them as context. Citation cards line up horizontally below the answer, each showing publisher and headline. Follow-ups in the same thread keep context.

"Browse" is the more active mode: while answering, the model can decide "I need to actually fetch this URL" and pull the page body. ChatGPT's expandable tool-use log shows which URLs it opened.

A dedicated SearchGPT entry point exists, but the trend is absorption into the main chatgpt.com surface. Free users get limited access; Plus / Pro / Team / Enterprise unlock the full thing.

One-liner: **the lowest-friction answer search if you already live in ChatGPT** — and publisher licensing gives citations decent stability.

5. Claude Web browse — Anthropic's answer

Anthropic added web search and browse to Claude in phases across 2024–2025. By 2026 every paid claude.ai user (Pro $20/month, Max, Team, Enterprise) can enable Web search as a tool; Claude decides when to call it, fetch pages, and inline citations in the answer.

Claude's browse is the **"tool called only when needed"** model. You don't toggle a mode — when Claude judges its training data is insufficient, it calls web_search, fetches results, and weaves them into the answer with inline source links.

Claude's edges are **long context** and **analytic synthesis**. A 200K–1M token window lets it compare and synthesize across dozens of pages — "how did these five outlets cover the same event differently?" is a Claude-shaped question. The trade-off is that browse can be slower than Perplexity or ChatGPT Search because the model calls tools conservatively.

Claude also ships "Computer Use" — a stronger tool-use mode where the model drives a browser directly. Pro/Max users see it in some workflows. Stability and cost keep it off the default answer-search path for now.

For developers, Anthropic exposes web_search as a first-class tool in the official API. You can build "Claude searches for itself" patterns in your own app via the Anthropic SDK.

One-liner: **strongest when the answer needs to be analytic, long-form, or carefully cited** — pick Perplexity or ChatGPT for fast one-shots.

6. You.com — pivoted to enterprise

You.com was founded in 2020 by Salesforce's Richard Socher. Early on it played in the consumer answer-engine bracket with Perplexity, shipping "Smart" / "Genius" / "Research" modes for general users.

Between 2024 and 2025 You.com **pivoted to enterprise**. A 2025-era Series C anchored the messaging on "an AI answer platform that combines your company data with web search." The consumer you.com site still works, but the company's center of gravity moved to You.com Enterprise, You.com API, and You.com Search API.

On the API side, You.com Search API became a popular pick for developers adding "fresh web search" to LLM apps. JSON responses with clean snippets and URLs slot well into RAG pipelines as a retriever. You.com joined Exa, Tavily, Brave Search API in the "search API for LLMs" category.

The enterprise product indexes internal sources (Confluence, SharePoint, Notion, internal wikis) so employees query "internal + web" in one shot. SSO, permission scoping, and audit logs are standard.

For consumers, You.com is still a workable Perplexity alternative. But marketing and new-feature weight shifted to enterprise, so consumer UX evolves more slowly than Perplexity or ChatGPT.

One-liner: **developers want You.com Search API; enterprises want You.com Enterprise; consumers can pick it, but it's one of many**.

7. Arc Search (The Browser Company) — mobile leader

The Browser Company, founded in 2019 by Josh Miller, shipped its desktop browser **Arc** in 2022 and earned a devoted design-and-UX following with "the browser is the OS"-style positioning. **Arc Search** for iOS launched in January 2024 as their mobile move.

Arc Search's signature is **"Browse for me."** Type a query on the mobile keyboard, tap the button, and Arc opens multiple web pages on your behalf, reads their content, and synthesizes a single tidy answer page with section headers and citation links. No SERP — just a built answer from page one.

Other strengths are mobile-native gestures like **pinch-to-summarize** (two-finger pinch turns a page into a one-screen summary) and aggressive default blocking of ads, cookies, and trackers, which strips a lot of mobile-web friction.

Arc Search is free. Desktop Arc is a separate app, and the two sync tabs and bookmarks via your iCloud account.

In late 2024 The Browser Company announced it would dial back new feature work on desktop Arc and focus on its next-gen AI-first browser **Dia**. That shook the Arc community. Arc isn't shut down, but it's effectively in maintenance mode. Arc Search on mobile continues to receive attention.

One-liner: **the lowest-friction "give me an answer fast on mobile" browser** — though the future of desktop Arc remains a concern.

8. Dia browser (May 2025, Arc's successor) — AI-first

**Dia** is the AI-first desktop browser that The Browser Company unveiled in May 2025. It inherits Arc's design DNA but instead of bolting AI into a sidebar, it puts the LLM into **the browser's address-and-command bar itself**.

Dia's center of gravity is the **command bar**. Like Chrome's omnibox or Arc's command bar, but it parses input as natural-language commands instead of just URLs and queries. "Find that company funding article I read last week," "summarize this page and paste it to Notion," "compare prices across my five open tabs as a table" — these become single-line commands.

Commands compose into **Skills**. Users and teams define Skills that capture frequent workflows. The pattern resembles ChatGPT's GPTs and Claude's Projects / Skills thinking.

Dia is also strong on **tab-context awareness**. It feeds the bodies of your currently open tabs to the model, so "what claim shows up across these five tabs?" works, with the answer pointing to which paragraph of which tab.

Dia opened by invite in 2025 and expanded access gradually through 2025–2026. Pricing is free + paid Pro: free covers basic answers, heavier models and unlimited usage unlock with Pro.

The Browser Company explicitly framed Dia as Arc's successor. That's two signals to Arc users at once — the good signal is "AI-era browser design is finally starting"; the bad signal is "Arc gradually stops receiving new features."

One-liner: **the first serious attempt to put AI into the center of the browser rather than its sidebar — the command-bar paradigm's testing ground**.

9. SigmaOS — Spaces-based

**SigmaOS** is a macOS-only browser launched in 2020. Its core concept is **Spaces** — each workspace owns a permanent group of tabs, and work, side projects, and personal life never mix. Each Space can isolate sessions, cookies, and extensions.

SigmaOS adopted AI early. **Airis**, the sidebar AI, summarizes pages, translates, and answers questions over current content. **Air**, the command bar, supports natural-language tab search, opening, and tidying.

Pricing is freemium with a paid Pro: free caps active tabs; Pro unlocks unlimited tabs and stronger AI models.

The audience fits information workers — consultants, PMs, designers, lawyers — who juggle many parallel projects. The standard loop is: isolate per Space, summarize and query within each Space via Airis, navigate via the command bar.

Compared to Dia, SigmaOS reads as "workspace-first with strong AI assistance," whereas Dia leans more on "AI-first." Dia is Chromium-based and targets both macOS and Windows; SigmaOS is macOS-only with deeper system integration.

One-liner: **the smoothest AI browser for macOS users juggling many projects in parallel**.

10. Brave Search + Leo AI

**Brave** is a privacy-first browser founded in 2016 by Brendan Eich. Ads and trackers are blocked by default, and Brave is a rare integrated stack with its own search engine (Brave Search) and its own AI assistant (Leo).

**Brave Search** launched in beta in 2021 and went GA in 2022. Its own index has grown to dominate result generation; Brave announced around 2023 that it had cut Bing API dependency and now serves an overwhelming share from its own crawl. **Goggles** are user-defined ranking rules — you decide which domains to boost or downrank, the same spirit as Kagi's lenses.

**Leo** is Brave's built-in AI sidebar inside the browser. It summarizes pages, translates, answers questions, and explains code. Free tier with usage caps, plus Premium at $14.99/month. Model choices include Claude, Mixtral, and Llama families. Brave emphasizes that Leo chats aren't used for training and are forwarded anonymously to model providers.

Brave Search also offers an API. The Brave Search API is now a reliable pick for LLM app developers building RAG pipelines, frequently compared with Exa, Tavily, and You.com Search API.

Big picture: **a single company that owns search, AI, and the browser — and is structurally ad-hostile**. Without Google-level ad dependency, advertiser bias has little path into answer synthesis.

One-liner: **the most consistent pick when you dislike ads and tracking and want search, AI, and browser all from the same vendor**.

11. DuckDuckGo + DDG AI Assist / Mojeek — independent indexes

**DuckDuckGo (DDG)** is a privacy-first search engine that started in 2008. It blends its own index, Bing API, and many "instant answer" sources. In 2024 it added **DDG AI Assist** — a short AI-generated snippet above the SERP — and the standalone "Chat" page, where you can call GPT, Claude, Mixtral, and others anonymously. DDG proxies you to the model, so your IP and account never reach OpenAI or Anthropic directly.

DDG's message is **"search + AI + tracker blocking + email, all privacy-first."** Browser (DuckDuckGo Browser on Mac/Windows/iOS/Android), search, and AI Chat live under one umbrella. Everything is free.

**Mojeek** is a UK company with a **100% own index** — no dependency on Google, Bing, or Yandex. Its own crawler powers a multi-billion-page index. That makes it one of the most serious independent-index candidates. Index size and freshness lag Google, and long-tail Korean/Japanese queries can show weakness, but the differentiation is real.

Mojeek also offers an API; some meta engines (including Kagi) use it as one source. No ads — revenue comes from paid users and API licensing.

Adjacent: **Marginalia Search** (a Swedish solo project, non-profit, biased toward text-heavy older blogs and documents) and **Stract** (a Sweden-origin open-source search engine with user-weighted ranking experiments). Both reward "I want results that are different from Google."

One-liner: **for diversity, pair a main engine with one of these as the secondary** — no ads, or an independent index, or a strong privacy story.

12. Phind — developer-focused

**Phind** is an answer engine specialized for code and engineering questions, launched in 2022. It markets itself as "programmer-focused search," preferring Stack Overflow, official documentation, and GitHub as citation sources, and returning answers as code-plus-explanation.

Phind's strength is **executable code answers**. Answers come as step-by-step code blocks with concise prose; citations usually link to official docs or GitHub issues/repos. ChatGPT and Claude can produce similar output, but Phind's answer format is shaped for a developer workflow from the start.

UX-wise Phind ships a **VS Code extension**, a web chat, and at one point experimented with a dedicated browser. In 2024–2025 it blended its own model line (Phind-CodeLlama family) with OpenAI/Anthropic models, on a free + paid Pro structure.

Phind targets "I need to write and debug code right now," not just "I need to learn about a topic." For questions with heavy environment context — OS dependencies, broken builds after a library bump — Phind frequently beats Google + Stack Overflow on round-trip time.

Competitors include IDE-embedded assistants (GitHub Copilot Chat, Cursor, Sourcegraph Cody, JetBrains Junie) and the general ChatGPT/Claude products. Phind fits best in the slot "I want to open a separate tab and get a quick code answer."

One-liner: **the search you open instead of a Stack Overflow tab when chasing a bug**.

13. Bing Copilot — Microsoft

**Bing** is Microsoft's search engine, and "Bing Chat" with GPT-4 in early 2023 was the first big shock to AI search. Through 2023–2024 Bing Chat was reorganized under the **Microsoft Copilot** brand. As of 2026 the lineup is **Copilot in Bing** (search-side AI), **Microsoft 365 Copilot** (Office-side), and **Copilot for Windows** (OS-side).

Bing itself still ships a classic SERP with a Copilot-synthesized answer at the top. The Edge browser ships Copilot in the sidebar — summary, translation, writing assistance, image generation (DALL-E based) all from one panel.

Microsoft's edge is **deep integration with Microsoft 365**. Outlook mail, Word/Excel documents, Teams chat and meetings flow into Copilot as context. Combining external web search with internal SharePoint/OneDrive via "Microsoft Search" is a different game from consumer answer engines.

Pure consumer search share for Bing still trails Google. But Edge + Windows 11 + Office bundle Windows users into Bing's results via Copilot whether they explicitly chose Bing or not.

Pricing: consumer Copilot is free with caps; Copilot Pro at $20/month adds GPT-4 priority and Office integration; M365 Copilot is around $30/month per seat on an annual contract.

One-liner: **the most natural AI search if you already live in the Microsoft ecosystem**.

14. ExaSearch / Komo / Andi / Yep / Marginalia / Stract — the rest

**ExaSearch (Exa)** is a **semantic search API** for LLM app developers. Instead of keyword matching, it ranks "documents semantically similar to this seed" via embeddings. It's frequently used in RAG pipelines, research automation, and competitor monitoring. Distinct category from the keyword search APIs of Brave / You.com / Tavily.

**Komo** positions itself as a collaborative answer engine — Perplexity-like UX plus "team workspaces" for jointly curating answers and sources.

**Andi** is a small answer engine whose mission is "search that answers like a friend." Simple UI, free, no ads, with a reputation for natural-sounding answers.

**Yep** is built by Ahrefs. It uses Ahrefs's massive own-crawler index for search and experiments with a "creator-first" model where revenue flows back to content authors. Closer to a classic SERP than to AI synthesis.

**Marginalia** — already mentioned. Swedish solo non-profit search, biased toward text-heavy older blogs and documents and against SEO spam. The best tool when you want to rediscover the "old web."

**Stract** — also mentioned. Open-source search where you can tune your own ranking weights.

What they share: **second and third options for users who refuse to outsource all search to a single Google**. When one engine returns nothing useful, another returns something different — that's the 2026 reality of search.

One-liner: **"one main engine + one or two secondaries" is the standard 2026 setup for information workers**.

15. Korea / Japan — Naver Cue:, Kakao i, Yahoo! Japan, LINE AI

In Korea, **Naver Cue:** opened in beta in 2023–2024. It's Naver Search's answer-style AI: type a natural-language query above the search box, and it returns a paragraph plus Naver-native sources (blog, café, news, encyclopedia). It's powered by Naver's own LLM (HyperCLOVA X family).

Naver's edge is the **depth of its own Korean-language index**. Korean blogs, cafés, and Knowledge-iN content are indexed elsewhere too, but Naver indexes them deeper and fresher inside its own walls. Cue: repackages that as an answer surface.

**Kakao** has wrapped Kakao's own LLMs (KoGPT, KoLLM family) into "Kakao i" branding, with answer/summarization features rolling step by step into KakaoTalk and Kakao Search. The center of gravity is "answer or summarize directly inside chat."

In Japan, **Yahoo! 検索 Japan** is the LY Corporation (LINEヤフー) search engine, and along with Google forms a two-engine market in consumer search. Through 2024–2025 LINEヤフー has been mixing its own and external LLMs (including OpenAI) into search, the LINE messenger, and LINE WORKS for stepwise AI answer/summarization features.

The **LINE AI** zone splits into in-messenger features (bots, summary, translation inside LINE) and the LINE WORKS business AI. Mobile-messenger share in Japan tilts overwhelmingly toward LINE, so a major entry point for AI answer search is "inside the messenger," differing from Korea or the West.

Japan also has Japanese-language LLMs like **NTT's tsuzumi**, adopted for government and enterprise use, advancing the "own-language answer search" agenda.

One-liner: **East Asia runs strongly on "own-language index + own messenger + own LLM" stacks, parallel to global Big Tech**.

16. Who should pick what — general / dev / privacy / mobile

Too many options. Persona-by-persona:

**General user, wants quick answers**: Perplexity or ChatGPT Search. Perplexity has the smoothest "answer + citation" flow; ChatGPT Search has the least friction if you already live in ChatGPT. Both free tiers are enough to try.

**General user, tired of ads**: Kagi at $10/month is the most serious option. If $10 is too much, Brave Search (free) + Brave Leo is a strong alternative.

**Privacy first**: DuckDuckGo (browser + search + AI Chat) or the Brave stack. If you also care about an independent index, Mojeek. Don't put everything in one basket — run two in parallel.

**Developer, code search and debugging**: Phind as the main answer search. Claude/ChatGPT inside IDE/Cursor for coding. ExaSearch in your own app's RAG pipeline. Stack Overflow remains valuable as a citation source.

**Developer, search APIs for LLM apps**: Brave Search API, Exa, Tavily, You.com Search API, SerpAPI. Pick two or three to compare against your workload before standardizing.

**Mobile, fast answer**: Arc Search is still the lowest friction. The Perplexity mobile app is excellent, and ChatGPT mobile's Search mode is fast too.

**Want to shift your workflow to an AI-first browser**: Dia if you can get access, SigmaOS on macOS, Brave for the conservative-and-stable choice. Desktop Arc, with maintenance-mode worries, requires more care.

**Need analytic, long-form answers**: Claude Pro/Max with web search. Comparing five outlets or summarizing a 50-page PDF while pulling in more search is Claude's zone.

**Enterprise deployment**: Microsoft 365 Copilot (if already on M365), You.com Enterprise (internal + web), Glean (enterprise search, out of scope here), Perplexity Enterprise. Or build your own RAG on top of an external search API.

**Korean**: Naver Cue: for everyday Korean queries; reinforce with Perplexity / ChatGPT / Claude for global research.

**Japanese**: Yahoo!検索 + Google + ChatGPT/Claude combined. AI answers inside LINE are growing fast, making mobile-messenger workflows another search entry point.

Boiled down: **one main answer engine + one or two secondaries + one AI browser** is the 2026 standard kit.

17. References

- Perplexity — https://www.perplexity.ai/

- Perplexity Comet browser — https://comet.perplexity.ai/

- Kagi Search — https://kagi.com/

- Kagi pricing — https://kagi.com/pricing

- OpenAI — SearchGPT prototype (July 2024) — https://openai.com/index/searchgpt-prototype/

- ChatGPT — https://chatgpt.com/

- Anthropic — Claude — https://www.anthropic.com/claude

- claude.ai — https://claude.ai/

- You.com — https://you.com/

- You.com API — https://you.com/api

- The Browser Company — Arc — https://arc.net/

- Arc Search (iOS) — https://arc.net/search

- The Browser Company — Dia — https://www.diabrowser.com/

- SigmaOS — https://sigmaos.com/

- Brave — https://brave.com/

- Brave Search — https://search.brave.com/

- Brave Leo — https://brave.com/leo/

- Brave Search API — https://brave.com/search/api/

- DuckDuckGo — https://duckduckgo.com/

- DuckDuckGo AI Chat — https://duckduckgo.com/aichat

- Mojeek — https://www.mojeek.com/

- Phind — https://www.phind.com/

- Microsoft Copilot — https://copilot.microsoft.com/

- Bing — https://www.bing.com/

- Exa (search API) — https://exa.ai/

- Komo — https://komo.ai/

- Andi — https://andisearch.com/

- Yep (by Ahrefs) — https://yep.com/

- Marginalia Search — https://search.marginalia.nu/

- Stract — https://stract.com/

- Tavily (search API) — https://tavily.com/

- Naver Cue: — https://cue.search.naver.com/

- Kakao i — https://www.kakaocorp.com/page/service/service/Kakaoi

- Yahoo! JAPAN Search — https://search.yahoo.co.jp/

- LY Corporation (LINEヤフー) — https://www.lycorp.co.jp/

- NTT tsuzumi — https://www.rd.ntt/research/JN202310_tsuzumi.html