Conversational search: when the answer replaces the list of results

How does it differ from semantic search?

They are different layers. Semantic search is a technique: it compares the meaning of the query with the meaning of the content to find results. Conversational search is the experience built on top: it uses search to gather sources, a model writes an answer, links the sources and keeps context for the next question. These systems resemble the RAG pattern: sources are retrieved and then a model uses them to answer, although each product documents its internals differently.

What do the main products do today?

This comes from official documentation consulted on November 4, 2026. It changes fast, so check it before relying on a detail.

For ChatGPT search and Perplexity, the table covers only what their crawler pages state; interface or plan details are not claimed here.

What changes for the user?

In classic search, a person types keywords, opens several results and builds their own answer. It is slow, but they decide which sources to read. In conversational search, the system does that work. Silvia Podesta, in a March 2024 UXmatters column, sums it up this way: the person no longer has to evaluate each individual result. Citations give a sense of control, but it is hard to judge the reliability of those sources, and the tool does not communicate well why it chose some and not others. Her central warning is trust: people could accept a ready-made, authoritative-sounding answer without questioning it.

There is data on the effect on clicks. In March 2025, Pew Research Center measured the browsing of 900 U.S. adults: with an AI summary on Google, 8% of visits included a click on a traditional result, versus 15% without a summary, and only 1% included a click on a link inside the summary. It is a Google measurement in the U.S.; it says nothing about other products.

Follow-up questions are the other half of the change. They let people refine without retyping everything, but they only help if the interface keeps context, as Google does when moving from an AI Overview to AI Mode.

How do you design it?

Five decisions worth making explicitly:

1. Sources next to the key claims. The documentation for OpenAI’s API web search tool asks that, when web results are shown, inline citations appear as prominent, clickable links. A list at the bottom is not enough. More in citations and sources in AI answers.
2. A way out when there is no good answer. Google does this in AI Mode by showing web links when it is not confident in the quality of its response. More in showing uncertainty.
3. Tappable follow-ups with context. Suggesting related questions helps people who do not know what to ask, as long as they depend on the answer and are not generic. This is practitioner advice, not measured data.
4. Easy recovery. Google admits its responses can be wrong and asks users to check and send feedback. Your interface should offer something concrete: rephrase, ask for another source, correct. Invented answers are explained in hallucinations.
5. Choosing the format. A chat creates a different mental model than a list of results. A short answer next to a list of sources preserves the habit and makes it easy to visit the original sites. Compare with chat vs. structured UI before deciding.

When should you NOT use it?

This is a design judgment, not a data point:

• When the person needs to explore. Comparing products or looking for images usually works better with a list or a gallery.
• When the exact figure matters. Legal deadlines, doses or prices: lead straight to the official source, as Google does when it asks people to check important information.
• When sources are few or low quality. A long list of citations does not create trust by itself.
• When search is a shortcut. If someone is looking for a specific site or page, a search box with direct results is faster.

Example

Illustrative example: a travel guide site with conversational search. Someone asks “which area of Lisbon is best to visit without a car with kids?”. The answer summarizes three neighbourhoods with one reason each, and each neighbourhood links to the source of the recommendation. Below are three follow-ups: “Which one has the most parks?”, “Show me hotels near the metro” and “And in winter?”. When the visitor picks the last one, the answer remembers they travel with children and covers each neighbourhood’s climate without asking them to repeat the context. If there is no good answer, it offers the list of sources instead of making one up.

Common mistakes

• Presenting the answer as absolute truth, with no signal that it can fail.
• Putting sources at the end as decoration instead of next to the claims.
• Losing context on every follow-up question.
• Suggesting generic follow-ups that have nothing to do with the answer.
• Always replacing the list of results without checking whether people need it.
• Stating how a third-party product works internally without an official source.

Sources

• Google Search Help: AI Mode
• Google Search Help: AI Overviews
• Google Blog: AI Mode in Search (May 20, 2025)
• OpenAI: Overview of OpenAI crawlers
• OpenAI: Web search (API)
• Perplexity Docs: Perplexity crawlers
• Anthropic: Claude web search
• Pew Research Center: Google users are less likely to click on links when an AI summary appears (July 22, 2025)
• UXmatters: Conversational AI Search Engines: Implications for Usability and the User Experience (March 4, 2024)
• UX Design Institute: Designing for AI Search in 2026: A UX Designer’s Guide to LLM Interfaces and Machine Experience (practitioner guide; no studies or data, its recommendations are opinion)