Conversational Search Engines

By Bill Hartzer · Updated 19 August 2026

What a conversational search engine is

A conversational search engine is a search product where the interaction is a dialogue rather than a series of independent queries. You ask a question in ordinary language, you get a written answer, and your next message is understood in the context of the previous one. "And what about the older model?" is a valid follow-up, even though it contains no searchable terms at all.

That single property — carried context across turns — is what separates a conversational search engine from a search engine that happens to accept a long query. A conventional engine treats every search as a fresh event. It has no idea that the reader's last three searches were about the same camera. A conversational engine keeps the thread and resolves "it", "that one" and "the cheaper version" against it.

Under the interface, three things are happening in sequence: the system works out what to search for, retrieves documents from a web index, and generates an answer over what it retrieved. Only the first and third are new. The middle step is ordinary search, and it is usually somebody else's.

The products commonly described this way as of August 2026 are Perplexity, ChatGPT with web search, Microsoft Copilot, Google's AI Mode, Yahoo Scout and Kagi Assistant. They differ far more in the retrieval layer than in the conversation layer, which is the reverse of how they are usually compared.

The idea is much older than the technology

Natural-language search is not a 2023 invention. It is one of the oldest recurring promises in the industry, and it failed twice before it worked.

Ask Jeeves launched in 1996 on exactly this pitch: type a question the way you would ask a person, and a butler brings you the answer. Behind the mascot was something far more modest: a bank of human-written question templates matched against the query, backed by conventional keyword retrieval. The interface promised conversation and the machinery delivered pattern matching. The gap showed, the butler was dropped in 2006, and Ask.com ceased to operate as a general search engine.

The second attempt was voice. Assistants on phones and speakers from the early 2010s accepted spoken questions and returned a single spoken answer, which is conversational in form. But they were mostly front ends that routed a parsed query to a conventional engine and read back one result, and the failure mode was familiar: ask a follow-up and the assistant started again from nothing.

What changed after 2022 is not that machines began to parse questions — Ask Jeeves parsed questions in 1996. It is that a language model can now compose an answer from retrieved documents rather than pick one, and can hold the thread of a conversation while doing it. The novelty is in generation and in memory, not in accepting a sentence.

How the multi-turn loop actually works

Four stages, each a place where the product can differ from what the reader thinks it is doing.

Rewriting. The question a reader types is rarely the query that runs. Microsoft's documentation states that Copilot "generates a search query that it sends via secure connection to the Bing search service." OpenAI's help documentation likewise says queries are rewritten before being sent to third-party providers. On a follow-up turn this step also has to resolve pronouns and implied subjects against the earlier conversation.
Retrieval. The rewritten query — often several of them, issued in parallel — hits a web index. Google describes its version as a "query fan-out": one question becomes a set of related searches whose results are pooled.
Selection and reading. Retrieval returns far more candidates than the model can use, so some are opened and read and the rest are discarded. Independent measurement of ChatGPT's retrieval stack published on 17 August 2026 found pages the system actually opened were cited about 74% of the time, against about 7% for pages merely retrieved into the candidate set.
Generation. The model writes prose conditioned on the selected passages, attaching citations. It is generating, not quoting, so the sentence can be true, false, or a plausible blend of two sources.

Almost every complaint about conversational search traces to stage one or stage four. If the answer misses the point, the rewrite probably did too. If it is confidently wrong with a citation attached, generation is the culprit.

Whose index is answering

This is the fact that most coverage of conversational search gets wrong, and it is the same distinction that separates a search engine with its own crawler from one that resells another engine's results. Applied to conversational products, as of August 2026:

Perplexity — its own. It runs a documented crawler, PerplexityBot, with a published user-agent and published IP ranges, and its September 2025 engineering write-up describes an "exabyte-scale index and crawling apparatus" tracking over 200 billion unique URLs. This was not always true: for roughly its first two years Perplexity was, in substance, a Bing API reseller with a summarisation layer, and it has never published a dated statement of when that stopped.
ChatGPT — a mixture, and it depends what you pay. OpenAI runs OAI-SearchBot feeding an internal index; its help documentation names Bing and Shopify as partners and says Bing is the sole third-party provider for Enterprise and Edu workspaces. The August 2026 measurement found the free tier served roughly three-quarters from OpenAI's own index, and the paid "thinking" mode served roughly three-quarters from scraped Google results purchased through commercial data providers.
Microsoft Copilot — none of its own. No crawler, no index, no ranking. It grounds on the Bing search service, which is Microsoft's own index rather than a rival's, and which Microsoft's documentation notes "operates separately" under different data-handling terms.
Yahoo Scout — somebody else's index and somebody else's model. Launched 27 January 2026, Scout uses Anthropic's Claude as its foundation model and Microsoft Bing's grounding API for the web. Yahoo's own framing at launch was that the underlying search index is Bing's while the responses, ranking and experience are Yahoo's. Yahoo has not operated a general web crawler since around 2010.
Google AI Mode — its own. Grounded on Google's index via query fan-out, with no separate crawl.
Kagi Assistant — blended. Kagi crawls with Kagibot into its own indexes (Teclis for web, TinyGem for news) but supplements essentially every query with anonymised API calls into third-party commercial indexes. Its Assistant is model-agnostic: the user picks which vendor's model generates.

Two consequences follow that no amount of interface polish changes. First, trying a different conversational engine is frequently not a second opinion — Copilot, Yahoo Scout and Bing itself are three products over one index. Second, when a conversational engine cannot answer, the limit is often the index rather than the model, and no better prompt will fix it.

The forcing event behind much of this was commercial rather than technical. Microsoft retired the Bing Search APIs entirely on 11 August 2025, decommissioning the cheap retrieval substrate a large part of the field had been built on. Companies that had built their own index carried on; the rest paid up, bought scraped data elsewhere, or found a first-party arrangement.

What conversation buys, and what it costs

What it buys. Questions with several clauses, an implied comparison, or an ambiguity a human would resolve by asking back. Refinement without reformulation — "no, the European version" rather than retyping the query. And synthesis across sources, which a list of links leaves entirely to the reader.

What it costs. Five things, and they are not opinions.

The ranked list, as an instrument. A results page lets a reader notice that the top eight links belong to one company, that the only primary source is fourth, or that everything is from 2019. A conversational engine makes those judgements silently.
Determinism. The same question asked twice can produce different answers. A ranked list is stable enough to cite; a generated paragraph is not the same object twice.
Inspectability. The rewritten query is not shown. The candidate set is not shown. The reason one source was preferred over another is not shown.
A consistent view of the web. Free and paying ChatGPT users being served from materially different retrieval backends is the sharpest version of this. That is not a quality tier — it is a different web, undisclosed in the interface.
An exit. Pure conversational products have no ten-blue-links mode. Kagi is the exception that proves the rule: it is a search engine that added an assistant, so the results page never went away.

Conversational search and a chatbot are different things

The dividing line is grounding. A chatbot answers from a model's trained weights: it has a knowledge cutoff, no citations, and no mechanism for knowing whether what it says is currently true. A conversational search engine retrieves live documents first and generates over them.

The distinction is easy to lose because the same interface does both. Microsoft describes web grounding in Copilot as a toggle — "only applicable if web grounding is enabled" — so the same chat window can return either a grounded answer or a model-memory answer, and the two look identical. ChatGPT decides on its own whether a question warrants retrieval, and OpenAI documents no clean switch to disable web search entirely.

The practical test is the citations. Grounded answers carry links to pages that exist and can be opened. An answer that sounds specific but cites nothing is the model's memory, not the web.

What to watch

Advertising. The premise that conversational search escapes the incentives of ad-funded search did not survive. OpenAI announced advertising in ChatGPT on 17 January 2026 for free-tier adult users in the United States, with ads in production by March 2026; an OAI-AdsBot agent in OpenAI's crawler documentation corroborates it. Copilot has always sat inside Microsoft's advertising division. Perplexity ran the experiment in reverse, launching ad formats in late 2024 and discontinuing them in February 2026. An ad in a ranked list is a labelled row that can be ignored; its equivalent inside a paragraph of prose is not observable from outside.

Agents. The conversational surface is being extended from answering to acting. That has already produced litigation: Amazon sued Perplexity in November 2025 over its Comet browser's shopping agent accessing customer accounts, won a preliminary injunction on 10 March 2026, and had it overturned by the Ninth Circuit on 4 August 2026. The case continues, and is the leading United States authority on whether an AI agent may act on a website on a user's behalf.

Consolidation of the index layer. The number of organisations that crawl and rank the open web at scale is small — Google, Microsoft, Yandex, Baidu, Brave, Mojeek and a few others, joined recently by Perplexity and OpenAI. Every conversational product is downstream of that list, and which of them are building indexes rather than renting them is a better predictor of where the category goes than which model they use this quarter.

Frequently asked questions

What is a conversational search engine?

A conversational search engine is a search product where questions are asked in ordinary language and each turn is understood in the context of the previous one, so a follow-up like "what about the cheaper model?" works without repeating the subject. It retrieves documents from a web index and generates a written answer with citations, rather than returning a ranked list.

What is the difference between conversational search and a chatbot?

Grounding. A chatbot answers from a model's trained weights, with a knowledge cutoff and no sources. A conversational search engine retrieves live web documents first and generates over them, which is what makes citations possible. The same interface often does both — Microsoft documents web grounding in Copilot as a toggle — and the two are hard to tell apart without looking for links.

Which conversational search engines have their own index?

Few. Perplexity runs its own crawler and index and publishes figures on its scale, though it began as a Bing API reseller. Google's AI Mode runs on Google's index. OpenAI runs its own crawler and index for ChatGPT but also buys web data elsewhere. Microsoft Copilot has no crawler and grounds on Bing. Yahoo Scout grounds on Bing's API with Anthropic's Claude as its model.

Was Ask Jeeves a conversational search engine?

It was marketed as one from 1996, but the machinery was human-written question templates matched against the query, backed by conventional keyword retrieval. It parsed questions; it did not compose answers and did not carry context across turns. The idea was right and the technology was not there, which is why the promise had to be made twice more before it worked.

Is Google a conversational search engine?

Partly, through AI Mode — a chat surface with follow-up turns that launched as a Search Labs experiment in March 2025 and reached nearly 120 countries in English by December 2025. AI Overviews are not conversational: they are a generated block on an ordinary results page with no memory of a previous query. As of August 2026 AI Mode sits alongside classic results rather than replacing them.

Do conversational search engines give the same answer twice?

Not reliably. The answer is generated rather than retrieved, so wording and sometimes substance vary between runs. Retrieval can also vary: independent measurement in August 2026 found free and paying ChatGPT users being served from materially different backends. This is the main reason a generated answer is a poor thing to cite and the linked source is a better one.

Does a conversational search engine search the query I typed?

Usually not exactly. Microsoft's documentation states that Copilot generates its own search query from the prompt before sending it to Bing, and OpenAI says queries are rewritten before going to search providers. On a follow-up turn the system also has to resolve pronouns against the earlier conversation. The rewritten query is not shown and cannot be edited.

Are there conversational search engines without advertising?

Kagi takes no advertising anywhere in its product and funds itself by subscription, and its AI answers are invoked rather than automatic. Perplexity introduced ad formats in late 2024 and discontinued them in February 2026 in favour of subscriptions. OpenAI announced advertising in ChatGPT in January 2026, in production by March, and Microsoft Copilot sits inside Microsoft's advertising business.

Conversational search engines