Answer Engine
The term names a behavior, not a product category. The same company can operate both a search engine and an answer engine, and increasingly does, with both running on the same index. What changes is the final step: a search engine hands the user a list and lets them do the reading, while an answer engine does the reading and hands the user a conclusion.
The pipeline: retrieval, then synthesis
An answer engine works in two phases that are worth keeping mentally separate, because they fail differently and are optimized differently.
The retrieval phase is classic information retrieval. The engine takes the user's query, typically reformulates it into one or several internal queries, and pulls candidate documents from an index. This phase decides which sources are even available to the answer. A source that is not retrieved cannot be cited, no matter how good it is, which is why indexability remains a hard precondition for everything discussed on this site.
The synthesis phase is generation. A language model receives the retrieved passages along with the question and composes a response in prose, usually attaching citations to some of its sentences. This phase decides which of the available sources actually shape and get credit for the answer. The model can merge three sources into one sentence, paraphrase a source beyond recognition, or cite one page for a fact that three pages stated.
The practical consequence of the split: traditional SEO work targets the retrieval phase, while the newer work of answer engine optimization extends into influencing passage selection and surviving synthesis intact. The two phases together are what make citation a different kind of outcome than a ranking.
Systems that exist today
Four named examples show the range of the category as of late 2026.
Google AI Overviews generates an answer above the classic results for some queries, drawing on Google's own search index. Google's documentation for site owners describes the feature and confirms that inclusion rides on ordinary indexing, with no dedicated markup.
ChatGPT with search answers in a conversational interface and browses the web when the question needs current information, citing the pages it drew on.
Perplexity was built around the sourced-answer interaction from the start: every response is a composed answer with numbered citations, with the link list demoted to a supporting role.
Claude answers conversationally and cites sources when it searches the web to ground a response.
The differences between them matter for measurement, because each has its own index access, its own citation behavior, and its own update cadence. An answer about your brand can differ across all four on the same day.
Linking out versus answering in place
The deep difference between the two engine types is where the user's reading happens. A search engine is a router: its product is a decision about where to send you, and the content consumption happens on the destination site. An answer engine is a destination: its product is the answer itself, and the source sites become footnotes to a text the engine wrote.
This reverses the economics of a click. In the search model, satisfying the user and sending traffic are the same event. In the answer model they decouple: the engine can satisfy the user completely while the cited sources receive nothing but attribution. Qualitatively, when the generated answer resolves the query in place, there is less reason to click any link at all, and the clicks that do happen skew toward users who want depth, verification, or a transaction the answer cannot complete. Publishers should expect the composition of their traffic to shift accordingly, with the reference click thinning out before the transactional click does. Precise figures for this shift vary by study and by query class, and this page deliberately cites none; measuring your own exposure is covered under citation rate.
Boundary cases
Not everything that answers is an answer engine in this sense. A featured snippet quotes one page verbatim rather than synthesizing several, though it foreshadowed the interaction. A voice assistant reading out a single result is a delivery format, not a synthesis pipeline. The term as used in this glossary requires both halves: retrieval across sources and generative composition of the response. When only one half is present, the optimization problem is different enough to deserve its own name.