ChatGPT Search

Original Article ↗ Hacker News Discussion ↗

OpenAI’s new ChatGPT Search, which layers a fine-tuned GPT‑4o model on top of Bing’s index, is seen as a direct challenge to Google and Perplexity by offering conversational, source‑linked answers instead of traditional link lists. Commenters are split: some find it dramatically better for complex, multi-step queries and up-to-date coding help, while others worry about hallucinations, SEO‑driven garbage being laundered into fluent summaries, privacy concerns, and the long‑term impact on publishers whose content may be consumed without clicks. Many expect advertising and pay‑to‑rank incentives to appear eventually, raising fears that AI search will repeat—and possibly intensify—the ad-driven distortions that have degraded web search over the past decade.

Backend and architecture

  • Several commenters note ChatGPT Search is not a standalone crawler yet, but largely relies on Bing’s index and other third‑party search providers, per OpenAI’s own help docs.
  • Some expect this to help sidestep robots.txt blocking of OpenAI’s own crawler; others stress OpenAI currently claims to respect robots.txt, though details (e.g., crawl‑delay) are unclear.

Comparison with Google, Perplexity, Kagi, etc.

  • Many see this as a direct shot at Google and Perplexity; some think OpenAI is “late”, others cite Chrome vs. late browsers as evidence timing may not matter.
  • Users compare it to Bing Copilot, Perplexity, Kagi Assistant, Phind, Brave Search; some say Perplexity/Kagi still feel better, especially for research and citation quality, others report ChatGPT Search did better on fresh code/library tasks.

Result quality, hallucinations, and reliability

  • Mixed reports: some are “super impressed” (e.g., handling a new library, code for niche FOSS), others show obvious hallucinations (fictional book titles, wrong finance models, wrong language versions, weather off by 20+ degrees, made‑up links).
  • The value is seen mostly in multi‑step or fuzzy queries (“plan a trip”, “integrate docs across libraries”) rather than precise facts where errors are more glaring.
  • People emphasize that without strong source‑level grounding and transparency, LLM answers can be less trustworthy than simply reading the underlying pages.

SEO, spam, and gaming

  • There is broad concern that if the underlying web is SEO‑polluted, LLM summaries may just compress garbage.
  • Some hope LLMs can learn to down‑rank SEO slop (using model‑level filters, user feedback, or even identifying AI‑generated spam from their own logs), but others expect an arms race: “SEO‑LLMs” trying to game “search‑LLMs”.

Ads, business model, and profitability

  • Intense debate over whether OpenAI will eventually add ads:
    • One side: search at massive scale can only be paid for by ads, and investors will demand growth, leading to enshittification similar to Google.
    • Other side: OpenAI already has substantial subscription revenue; some hope they can avoid or at least compartmentalize ads.
  • Several note the huge compute cost of LLM‑based search; question whether ads can cover it if queries are truly chat‑grounded.

Impact on the web and publishers

  • Strong worry that LLM search is parasitic: summarizes answers so well that users don’t click through, undermining ad‑funded publishers and long‑tail blogs.
  • Others argue much high‑quality content has always been hobbyist and will persist; some see this as a chance to kill SEO‑driven “content farms”.
  • People anticipate more paywalls, access deals, and lawsuits; some think search will balkanize around who pays for access.

UX, latency, and access

  • Many like the integrated chat + search UX and citations sidebar; others dislike wordy, slow, streaming answers compared to Google’s near‑instant results and simple blue links.
  • Currently limited to Plus/Team and waitlist users (with slow rollout to free); some see login‑requirement for search as a privacy red flag.
  • There is interest in using it as a browser search engine (custom URL parameters, Chrome extension, Alfred integration), but latency and rate limits are concerns.

Who benefits / use cases

  • Power users with strong traditional search skills are split: some see little value, others use LLMs to discover terminology, narrow research space, or stitch together multi‑source answers, then verify via classic search.
  • Many foresee this as a building block toward “agents” that not only search but execute tasks (reservations, purchases), raising worries about hidden commercial steering.