Three sites made 215,128 “best software” pages for AI. Perplexity cites them

Original Article ↗ Hacker News Discussion ↗

AI-driven “best software” review sites are being mass-produced to game generative search engines, with Perplexity cited as relying on thousands of such low-quality pages as sources. Commenters see this as part of a broader shift toward “AI SEO” or Generative Engine Optimization, where content is written in LLM-like prose to win favor with AI answer engines, further polluting the web and risking models training on their own output. Many also report declining usefulness and trust in Perplexity and other AI search tools, worrying that incentives for speed, growth, and ad-like monetization are eroding information quality across the internet.

AI-Generated “Best Software” Pages & GEO/AEO

  • Many see these mass-produced “best X software” sites as a natural evolution of SEO: now optimized for LLMs instead of humans, sometimes called GEO/AEO (“Generative/Answer Engine Optimization”).
  • Commenters expect a growing industry around manipulating AI training and retrieval so models recommend specific products.
  • Some note VC-backed companies already doing this at scale, including spammy posting on forums and iterative rewriting to match LLM embeddings.

Perplexity-Specific Reactions

  • Several former paying or trial users report quality decline: faster but worse answers, misaligned citations, and unreliable “Computer”/deep research features.
  • Complaints about billing practices (auto-charges after trials, poor support) and mid-subscription downgrades of paid features.
  • Some say Perplexity was once better than Google for search-like tasks but has lost its edge; others now prefer Claude, ChatGPT, Gemini, or even Google’s AI mode despite higher hallucination rates.
  • A minority still use Perplexity to avoid lock-in to a single underlying model vendor.

LLMs Training on LLM Output & Source Bias

  • Strong concern that AI-generated content is flooding the web, creating a feedback loop where models train on their own “slop,” amplifying errors.
  • Discussion of “self-preference bias”: models often rate their own style, code, or passage structure as superior to human edits or alternative solutions.
  • Experiments and papers are cited (in-thread) suggesting LLM judges prefer LLM-generated text, raising issues for evaluation and SEO-like gaming.

Search Degradation & Coping Strategies

  • Many say product search and general web search are “ruined” by spam, ads, and AI slop.
  • Some revert to pre-2020s books, niche forums, or curated domain whitelists/“lenses” (e.g., with alternative search engines) to avoid AI content.
  • Others use filters and blocklists in browsers to hide “AI slop” and SEO-heavy sites.

Skepticism About the Article & Meta-Astroturfing

  • Multiple commenters find the report itself AI-written, with tedious prose and low-credibility “research firm” branding.
  • Suspicion that the anti-Perplexity pieces from similar-looking “independent research” sites are themselves GEO experiments or astroturf meant to influence LLM outputs.
  • Some call for flagging or removing such AI-heavy submissions from discussion sites.

Broader Concerns: Incentives, Propaganda, and Small Players

  • Ads and monetization are seen as core drivers of spammy behavior; some argue propaganda actors are also seeding AI-friendly misinformation sites.
  • Smaller SaaS vendors report being excluded from LLM recommendations unless they pay “best software” sites for placement, sometimes facing retaliatory negative content if they refuse.
  • A few speculate about “human-only” networks or darknets, but acknowledge it’s unclear how to enforce human-only participation or remove economic incentives.