majorReviewed Mon Sep 14

Perplexity introduces Q2D-Web for large-scale retrieval evaluation in agentic RAG

Perplexity released Q2D-Web, a production-shaped retrieval benchmark built around 190 million web documents and 69,721 agent-reformulated queries across 10 languages.

Context

Perplexity released Q2D-Web, a production-shaped retrieval benchmark built around 190 million web documents and 69,721 agent-reformulated queries across 10 languages.

What changed

Q2D-Web evaluates machine-written search reformulations produced inside agentic RAG workflows rather than only human-written queries. It also provides separate relevance signals from agent citations, production rankings, and expanded LLM judgements.

Why it matters

The visible user prompt is not necessarily the query a source competes for. AI visibility therefore depends on several distinct stages: query reformulation, first-stage retrieval, ranking, evidence selection, and eventual citation.

Visibility OS interpretation

Visibility OS treats Q2D-Web as strong evidence for a multi-stage visibility model. Retrievable does not mean selected; selected does not mean cited; cited does not necessarily mean prominently represented in the final answer.

What remains uncertain

How stable are agent reformulations across repeated runs? Which retrieval signals best predict later citation? How transferable are the findings to Google AI Mode, ChatGPT Search, Gemini, Claude, and Copilot?

Evidence claims

EstablishedConfidence 99%introduces

Q2D-Web contains 190 million web documents and 69,721 agent-reformulated queries in 10 languages, sampled from nine months of PII-free production search traffic.

Related historical records

Google expands AI Mode and Deep Search using query fan-out

This record related to the linked event.

Perplexity launches its answer engine

This record builds on the linked event.