AI Hot Concepts Handbook

Appearance

Perplexity and AI Search

At a glance How Perplexity turned Retrieval-Augmented Generation into a mass-market product: the mechanics of its "search + citations" conversational answer engine, the paradigm gap with traditional search, and AI search's impact on the content ecosystem and traffic distribution.

This page contains time-sensitive material, accurate as of 2025-06; job listings, leaderboards, and product features may have changed since. Verify against the original source before citing.

Perplexity and AI Search ​

Background: An "Answer Engine" That Made RAG a Mass-Market Product ​

In December 2022, Perplexity AI, founded by Aravind Srinivas and colleagues, went live. Its product shape was unlike anything else at the time: instead of a screen full of blue links, you typed a question into a box and got a direct answer — with a clickable source citation hanging off every key claim, so you could jump to the original text and verify it for yourself. This conversational answer engine built on "search + citations" was the first time the Retrieval-Augmented Generation (RAG) technique long known in academia became a product ordinary people could pick up and use. For the technical details of RAG, see Retrieval-Augmented Generation (RAG) — the principle is to retrieve first, then generate, so the large language model (LLM) "speaks with the material in hand."

In the two or three years that followed, AI search went from novelty to table stakes. In July 2024, OpenAI unveiled the SearchGPT prototype and folded it into ChatGPT Search in November 2024; in May 2024, Google launched AI Overviews at Google I/O and pushed it across the US market; in China, Metaso AI Search went viral overnight in March 2024, and products like Quark and Doubao baked AI search into their own offerings. By 2025, "AI search" had become one of the most crowded races after ChatGPT and Conversational AI. At its core, this wave moves the entry point for information retrieval from "keywords + a list of links" to "natural language + generated answers" — a paradigm-level shift.

Anatomy of the Product: An Answer Pipeline with Citations Attached ​

Perplexity's experience looks deceptively simple; underneath runs a multi-stage pipeline:

User asks a question (natural language)
      ↓
① Query rewriting: rephrase the colloquial question into several searchable sub-queries
      ↓
② Multi-source retrieval: query the web index, knowledge sources, even real-time information in parallel
      ↓
③ Re-ranking: order results by relevance and credibility, extract the best passages
      ↓
④ LLM generation: use the retrieved passages as context to produce a cited answer
      ↓
⑤ Citation tracing: every answer segment is bound to a source link, clickable for verification

Stage by stage:

  • Query rewriting: User questions are often colloquial and vague ("what's the latest big-model news lately"), and the system must rewrite them into several precise sub-queries suited for search engines — the engineering embodiment of the prompt engineering idea of "turning a vague request into an executable instruction."
  • Multi-source retrieval: Under the hood this relies on vector databases and semantic retrieval for semantic matching — web passages are turned into embedding vectors, then similarity-searched against the question vector, solving the "no keyword overlap but semantically related" problem.
  • Re-ranking: The initial recall should be broad even if precision is low; the re-ranking stage then applies a finer-grained model (a cross-encoder) to push the genuinely relevant passages to the top. This step directly determines "the quality of the knowledge fed to the LLM."
  • Generation and citations: The LLM's job is to organize the retrieved passages into a coherent, direct answer. Citation tracing is Perplexity's killer feature — every claim can be traced back to its source, which both mitigates hallucination and gives users a path to verify things themselves. That is what most separates it from a bare chatbot.

The one-sentence verdict

The formula for AI search success is "good retrieval + good generation + traceable sources." Citations are the piece most easily overlooked, yet they are exactly what builds user trust — an AI search without citations is, in essence, just a more expensive chatbot.

Answer Engine vs. Search Engine ​

Put Google and Perplexity side by side in one table and the paradigm difference is obvious:

DimensionTraditional search engines (Google, etc.)AI answer engines (Perplexity, etc.)
InteractionKeywords + a list of linksNatural-language question + a direct answer
Information formatTen blue linksOne synthesized answer with citations
User mindset"I'll find the answer myself""The AI finds it and tells me"
Multi-step tasksRepeated searches, manual assemblyOne question, synthesized across sources
Time-sensitive infoIndex updates, direct linksReal-time retrieval + generation
Trust mechanismDomain, rankingCitations, verifiability
Main riskAds and SEO pollutionHallucination and generation bias

The answer engine's challenge to the search engine is paradigm-level: traditional search hands ranking power to the algorithm and judgment power to the user; AI search takes over the judgment as well. This has two direct consequences. First, search behavior shifts from "multiple rounds of clicking links" to "one round, one answer." Second, the logic of traffic distribution is rewritten — users no longer click through ten results one by one, they read a single summary, which shakes the foundations of a web ecosystem that bills on clicks. To place this shift in a larger frame, see What Is AI: Hot Concepts and Concept Boundaries.

The Technical Crux: Retrieval Quality Determines Answer Quality ​

Technically, AI search rests on the combination of "real-time retrieval + LLM summarization": retrieval is responsible for "knowing the latest facts," and the LLM for "shaping them into a credible statement." But the pipeline's Achilles' heel is this —

Retrieval quality determines answer quality

However strong the LLM's generation is, it cannot conjure away errors or noise introduced at the retrieval stage. "Garbage in, garbage out" is amplified in AI search: if the retrieved passages contradict each other, are outdated, or are polluted by SEO, the final answer will be wrong. The core engineering competition among AI search products is, at bottom, a competition over retrieval quality and data freshness. For how to build an evaluation system, see LLM Evaluation and Benchmarks.

The most famous failure of this weak point came in May 2024: after Google I/O, Google opened AI Overviews up fully, and users flooded in with reports of absurd answers — the AI "suggesting glue on pizza" or "eating a rock a day for minerals" — because AI Overviews had scraped joke posts on Reddit and low-quality web pages and treated them as fact. Google was forced to sharply roll the feature back within weeks, keeping it for only a small share of queries. The incident became the classic textbook case of AI search's twin risks of "retrieval quality + hallucination"; for the related safety and governance discussion, see AI Safety and Governance.

Compared with a bare chatbot, AI search has two structural advantages: fresh knowledge (it fetches in real time on every query, unbounded by the model's training cutoff) and traceability (citations let users verify). That also explains why "AI search + RAG" is the first path enterprises reach for when putting LLMs into production — to build a working RAG application from scratch, see Build a RAG App from Scratch. That said, RAG is not a silver bullet; compare the common engineering traps in Common Pitfalls and Antipatterns.

The Product Landscape (dataAsOf: 2025-06) ​

As of mid-2025, the AI search race has settled into a complete field of "overseas + domestic + big tech + startups":

ProductMakerNotes
PerplexityPerplexity AI (US startup)Conversational answers + citations; pioneered mass-market AI search
ChatGPT SearchOpenAIBuilt into ChatGPT, wired into its chat ecosystem
Gemini / Deep ResearchGoogleDeep Research mode: multi-step retrieval generates a report
AI OverviewsGoogleAI summaries at the top of traditional search results
Metaso AI SearchMetaso Technology (China)Went viral in March 2024; strong in Chinese-language scenarios
Quark AI SearchAlibaba (China)Mobile-first AI search, deeply integrated with the Quark browser
Doubao / ERNIE Bot SearchByteDance / BaiduSearch enhancement built into their LLM products

How to read this map

Big tech (Google, OpenAI, Baidu, ByteDance) takes the "existing entry point + AI enhancement" route, embedding AI search inside products it already owns; startups (Perplexity, Metaso) take the "standalone answer engine" route, betting on a new category of entry point. The former's moat is traffic; the latter's moat is the "answer experience" and citation quality.

From a wider vantage point, AI search and agents are converging: Perplexity has shipped the Comet browser, and OpenAI has released Operator and Deep Research — search is evolving from "finding answers" to "getting things done for you." That is the same paradigm shift described in Manus and Agent Applications, seen from the other side.

Impact on the Content Ecosystem: Traffic, Revenue Shares, and "Zero-Click Search" ​

The more popular AI search becomes, the harder it hits the existing content ecosystem, and three issues are unavoidable:

  1. Traffic distribution: Users stop clicking links, and website traffic broadly comes under pressure. Early data showed AI summaries measurably cutting click-through rates on some queries, and "zero-click search" became the phrase content publishers dread most.
  2. Publisher revenue-share disputes: Starting in 2024, Perplexity, OpenAI, and others signed content licensing and revenue-sharing agreements with publishers such as Time, Axel Springer, and the Associated Press, while The New York Times and others took the legal route into court. The core dispute: where is the line between "citing" and "summarizing," and who pays for the content being cited.
  3. Content production reshaped in reverse: Once "being cited by AI" becomes the new traffic outlet, content production starts bending toward writing that is "easy to retrieve and easy to cite" — the same game as the traditional SEO era, except the play is now "competing for citations" instead of "competing for rankings."

A loop worth worrying about

AI search depends on high-quality content as its corpus, yet intercepts publishers' traffic at the distribution stage. If publishers cut back on content as their revenue dries up, AI search's retrieval quality will deteriorate in step — the sustainability of the citation ecosystem is AI search's real long-term exam. Together with data compliance and content copyright, it forms the core variable of the industry's second half.

Where Answer Engines Go Next ​

The evolution of AI search will not stop at "text answers." Four directions are worth watching:

  • Multimodal search: Launch a search with a screenshot or a voice clip; "searching text with an image" and "identifying things by sound" are entering product form (see Multimodal Models).
  • Agentified search: From "give an answer" to "execute for you" — after gathering the research, it compares prices, books tickets, and drafts the summary. This is exactly the paradigm Manus and Agent Applications represents: search and execution are converging, and the endpoint of an answer engine is "turning answers into actions."
  • Personal knowledge search: Bringing local documents, chat logs, and email into the retrieval scope, expanding the answer engine from "searching the internet" to "searching my world" — still built on the same vector retrieval and RAG underneath.
  • The ultimate business-model question: Advertising clashes head-on with "zero-click"; can subscriptions carry the growth, and can revenue-sharing agreements produce genuine wins with publishers? These questions still have no standard answers.

Expect the competition in AI search to shift from "who can give the smarter answer" to "who can hold the line on accuracy (retrieval quality), truth (verifiable citations), and fairness (a win for publishers) at the same time." The technical foundation is Retrieval-Augmented Generation (RAG); the answers on business and governance will be left to the test of time.

Further Reading ​

References ​