What an answer engine is
An answer engine is a product that responds to a question with a written answer rather than with a list of pages that might contain one. Links are still present, usually as numbered markers attached to individual sentences, but they function as supporting evidence for prose rather than as the product itself.
The clearest example is Perplexity, which launched its search product on 7 December 2022 and has described itself as an answer engine from the start. Ask it when a company was founded and it returns a paragraph with small numbered citations, each linking to a page it drew from. There is no ranked list to scan and no mode that turns one on.
The term is older than the current generation of products. Wolfram Alpha, launched in 2009, called itself a computational knowledge engine. It does not search the web at all — it computes answers from curated structured data. Ask a conventional search engine for one country's population divided by another's land area and it will look for a page containing that sentence; Wolfram Alpha does the arithmetic. That is an answer engine in the strict sense, and it is a useful reminder that "answer engine" describes an output format, not a technology.
What changed after 2022 is that large language models made that output format possible over the open web instead of over a curated database. The answer is now generated rather than looked up — which is also why it can be wrong in ways a database lookup cannot.
Answer engine, search engine, chatbot
These three words are used interchangeably in ordinary coverage and they describe genuinely different things.
- A search engine crawls the web with a program of its own, stores what it finds in an index (the engine's own stored copy of the web), and ranks that index against a query. The output is a list. Google, Microsoft Bing, Brave Search, Mojeek, Yandex and Baidu all do this.
- A chatbot generates text from a model's trained weights. With no retrieval step it has a knowledge cutoff, no citations, and no way of knowing whether what it says is currently true. ChatGPT was exactly this, and only this, from its November 2022 launch until web retrieval was added.
- An answer engine is the combination: retrieval from some index, then generation over what was retrieved. Retrieval is what makes the citations possible and what makes the answer current.
The technique of retrieving documents and then generating over them is usually called retrieval-augmented generation, or simply grounding. Everything in this category does it. The interesting question about any of these products is therefore never whether it is "AI" — they all are — but which index it is grounded on.
Whose index is underneath
Almost every answer engine on the market is generating prose over an index somebody else built. This is the least-understood fact about the category, and it is the same distinction that separates a search engine with its own crawler from one that resells another engine's results.
- Perplexity built its own. It runs a documented crawler, PerplexityBot, with a published user-agent string and published IP ranges, and in September 2025 it published engineering figures describing an index tracking "over 200 billion unique URLs" (Perplexity Research, 25 September 2025). It did not start that way. In its first years it was widely and credibly reported to be serving results through the Bing Search API, with the summarisation layer as the differentiator. Perplexity has never published a dated statement of when that changed.
- ChatGPT Search is a mixture. OpenAI runs its own search crawler, OAI-SearchBot, feeding an internal index; it also names Bing and Shopify as search partners in its own help documentation, and says Bing is the sole third-party provider for Enterprise and Edu workspaces. Independent measurement published in Search Engine Land on 17 August 2026 found the free tier served mostly from OpenAI's own index and the paid "thinking" tier served mostly from scraped Google results purchased through third-party data providers.
- Microsoft Copilot has no crawler and no index at all. Microsoft's own documentation describes the mechanism plainly: Copilot parses the prompt, generates a search query from it, sends that query to the Bing search service, and writes prose over what comes back. The qualifier that matters is that Bing is Microsoft's own index, so Copilot is a front end onto its parent's crawl rather than onto a rival's.
- Yahoo Scout is the clearest case. Launched 27 January 2026, it is an AI answer engine with Anthropic's Claude as the model and Microsoft Bing's grounding API supplying the web. Yahoo's own summary at launch, as reported by Search Engine Land, was that "the underlining search index is from Bing, but the responses, ranking, and experience is all Yahoo." Yahoo has not run a general web crawler since around 2010.
The practical consequence is worth stating bluntly: a reader who consults Copilot, then Yahoo Scout, then Yahoo's classic results page has used three interfaces and one index. Switching answer engines is frequently not a second opinion.
What answer engines are genuinely good at
Messy questions. A ranked list is an excellent instrument for "find me the page about X" and a poor one for "which of these three things has the property I care about, and why." Questions with several clauses, an implied comparison, or an ambiguity that a human would resolve by asking back are where the format earns its keep. On a conventional engine these take three or four reformulations; on an answer engine they take one question and a follow-up.
Synthesis across sources. Pulling a specification out of a manufacturer page, a review and a forum thread and putting them side by side is genuine work, and it is work a list of ten links leaves entirely to the reader.
Follow-ups that keep context. "And what about the previous version?" is a query no keyword engine can serve, because it carries no keywords. This is the single clearest advance over the search box.
Freshness, in some products. Perplexity's indexing pipeline is tuned for recency and it shows on news-adjacent questions. That is a property of the index, not of the interface, which is precisely the point of asking whose index it is.
Where they fail
Citations are a claim about provenance, not a guarantee of accuracy. A citation means a document was retrieved. It does not mean the sentence it is attached to reflects that document. In June 2024, Forbes accused Perplexity of plagiarising its reporting and Wired published parallel findings on scraping and inaccurate summaries; the News Corp suit filed against Perplexity in 2024 included allegations that it produced quotes that did not exist and attributed them to Dow Jones and New York Post reporting. Citations create an appearance of verifiability that the generation step does not always earn.
You cannot scan. A results page lets a reader notice that the top eight links all belong to one company, that the only primary source is fourth, or that everything is from 2019. An answer engine makes that judgement on the reader's behalf and does not show its working.
The query you asked is not always the query that ran. Microsoft's documentation states that Copilot generates its own search query from the prompt and sends that to Bing. The reformulation is invisible and not editable, so the answer depends on a query the user never typed.
The same question can be answered from different indexes. The August 2026 measurement of ChatGPT's retrieval stack found free and paying users being served from materially different backends. That is not a quality tier; it is a different view of the web, and it is not disclosed in the interface.
There is usually no off switch. A pure answer engine has no ten-blue-links mode. Kagi is the notable exception in the other direction — its AI answers are invoked deliberately rather than fired on every query — but Kagi is a search engine that added AI, not an answer engine.
Who pays for the answer
The early framing of answer engines as an escape from advertising has not survived contact with the cost of running them.
Advertising in ChatGPT was announced on 17 January 2026 and was visible in production by March 2026, aimed at free-tier adult users in the United States; OpenAI's own crawler documentation independently corroborates it by listing an OAI-AdsBot agent that vets pages submitted as ads. Microsoft Copilot has always sat inside the division that also owns Microsoft Advertising and Bing. Perplexity went the other way: it introduced advertising formats in late 2024 and then discontinued them in February 2026 in favour of a subscription-first model — a reversal that received far less coverage than the launch did.
The structural point is not that advertising is present but that it is harder to audit. On a ranked results page, a paid placement is a labelled row a reader can ignore. In a paragraph of generated prose there is no row, no label position and no ranking to inspect, and the influence of a commercial relationship on a sentence is not observable from outside.
The other unresolved bill is the publishers'. An answer that satisfies the question removes the reason to click through to the page it was built from. The two large responses so far are opposite: OpenAI signed content licensing agreements with a long list of publishers — AP, Axel Springer, Condé Nast, the Financial Times, Le Monde, News Corp, Reuters and The Atlantic among those named in its launch post — while Perplexity has been sued by News Corp, Reddit and three Japanese newspaper publishers, and has run a revenue-sharing programme for cited publishers as a partial answer to the complaint. Neither approach has settled the question.
Frequently asked questions
What is an answer engine?
An answer engine is a product that replies to a question with a written answer and citations rather than with a ranked list of links. It retrieves documents from a web index, then uses a language model to generate prose over them. Perplexity, ChatGPT Search, Microsoft Copilot and Yahoo Scout are answer engines; Wolfram Alpha is an older, non-web variety that computes answers from curated data.
What is the difference between an answer engine and a search engine?
A search engine crawls the web, indexes it and returns a ranked list for the reader to evaluate. An answer engine adds a generation step that reads some of those results and writes the answer instead. The retrieval machinery underneath is the same problem; the difference is the output and who does the judging. Most answer engines do not run the retrieval machinery themselves.
Is Google an answer engine?
Partly. Google remains a search engine — it crawls the web with Googlebot and ranks its own index — but AI Overviews put a generated answer above the links on a large share of queries, and AI Mode is a full conversational surface. As of August 2026 AI Mode sits alongside classic results rather than replacing them, so Google is both things at once.
Which answer engines run their own web index?
Very few. Perplexity operates its own crawler and index and publishes figures on its scale. OpenAI runs a crawler and internal index for ChatGPT but also buys web data from other sources. Microsoft Copilot has no crawler at all and grounds on Bing. Yahoo Scout grounds on Bing's API with Anthropic's Claude as the model. Google's AI features run on Google's own index.
Are the citations in AI answers reliable?
They are reliable as a record that a page was retrieved, and unreliable as proof that the sentence reflects it. Reporting by Forbes and Wired in June 2024, and allegations in the News Corp litigation, described summaries and quotations that did not match the cited sources. Independent measurement in 2026 also found pages the system actually opened were cited far more often than pages merely retrieved.
Do answer engines show advertising?
Increasingly yes. OpenAI announced advertising in ChatGPT on 17 January 2026 and it was in production by March 2026 for free-tier United States adult users. Microsoft Copilot sits inside Microsoft's advertising business. Perplexity introduced ad formats in late 2024 and discontinued them in February 2026. Kagi takes no advertising at all and funds itself by subscription.
Is ChatGPT an answer engine?
It became one. From its November 2022 launch until web retrieval arrived, ChatGPT was a chatbot answering from model weights with a knowledge cutoff and no sources. ChatGPT search launched on 31 October 2024 and opened to users without an account on 5 February 2025. Retrieval now fires when the model judges it necessary, and there is no documented way to disable it.
Sources
- research.perplexity.ai/articles/architecting-and-evaluating-an-ai-first-sear…
- openai.com/index/introducing-chatgpt-search/
- searchengineland.com/chatgpt-retrieval-stack-index-cache-pages-485036
- learn.microsoft.com/en-us/copilot/privacy-and-protections
- yahooinc.com/press/introducing-yahoo-scout-a-new-ai-answer-engine
- searchengineland.com/yahoo-scout-yahoos-return-to-search-and-web-discovery-4…
- developers.openai.com/api/docs/bots
- en.wikipedia.org/wiki/Perplexity_AI