Answer Engine

An Answer Engine is a system that synthesizes a direct, cited answer to a user's query instead of returning a ranked list of links for the user to click through and read themselves. Where a traditional search engine indexes the web and lets the reader do the synthesis, an answer engine delegates that synthesis to a language model: it retrieves relevant passages, generates a prose response, and attaches inline citations to the sources it drew from. The category was popularized by Perplexity, which launched its answer product in 2022, and now includes ChatGPT Search (rolled out to all free users in early 2025), Google AI Overviews (general availability from May 2024), and Microsoft Copilot in Bing. The economic consequence for publishers is direct: when the answer is rendered on the results page, the click that would have gone to the source domain often never happens, compressing referral traffic for informational queries.

How it works

A query enters a retrieval pipeline — typically a vector search over a web index or a re-ranked BM25+embedding hybrid — that surfaces the top passages. Those passages are packed into a model's context window alongside a system prompt instructing it to answer using only the retrieved evidence and to cite each claim. The model generates a grounded response with inline reference markers, and a post-processing step maps those markers back to source URLs. Some systems add a follow-up retrieval pass when the first answer is low-confidence or the query is multi-hop.

Why it matters for AI engineers

Building an answer engine forces hard trade-offs between latency, cost, and grounding quality: every extra retrieval pass or reranking step improves citation accuracy but adds round-trip latency and token spend. Reliability hinges on the retrieval layer — a wrong or stale passage produces a confidently wrong answer, so eval harnesses must score citation faithfulness, not just fluency. Security exposure is real: retrieved web content can carry prompt-injection payloads that hijack the synthesis step, requiring input sanitization or sandboxed retrieval. Shipping decisions center on whether to own the index (expensive, defensible) or rent one, and how aggressively to cache semantically equivalent queries to cut inference cost.

Answer Engine vs. alternatives

Related terms