OpenAI tests SearchGPT, its conversational search engine with citations

On July 25, 2024, OpenAI opened SearchGPT as a temporary test for a small group of users and publishers. The prototype announcement promised answers with visible sources, follow-up questions and separate controls for publishers to manage how they appeared; it was not yet a generally available search engine.

The move matters because web search remains one of the most valuable gateways to the internet. Google has dominated the market for decades, but the popularity of AI assistants has accustomed many users to asking complete questions instead of stringing together keywords.

Direct answers, with sources in plain sight

SearchGPT offers a simple query interface: users ask a question and receive a written answer accompanied by attributions and links to relevant content. It also allows follow-up questions without restarting the search, as in a conversation.

OpenAI has shown examples involving travel planning, tracking events and finding information on specific topics. The difference from a traditional search engine is not just that it summarizes results: the system tries to organize the information and respond to the intent behind the question.

That approach is already part of other products. Microsoft has added AI-generated answers to Bing and Copilot, while Google is rolling out its AI Overviews, AI-generated summaries that appear ahead of conventional links for some searches. The battle is no longer just about who finds a page, but who decides what information users see first.

The challenge of bringing AI to the live web

Language models are good at writing and connecting information, but they are not reliable sources on their own. They can make mistakes, invent facts or rely on outdated knowledge. An AI search engine therefore needs to combine a model's conversational abilities with access to up-to-date information and verifiable references.

Citations are SearchGPT's decisive feature. In theory, they allow readers to check the origin of a claim and visit the news outlet or website that published it. In practice, the challenge will be ensuring that those references are accurate, relevant and visible enough without completely replacing visits to the original source.

The issue directly affects publishers. If an answer resolves a user's question without prompting them to open the link, a media outlet could lose traffic, subscriptions or advertising revenue. But a tool that clearly highlights authorship could also help users discover quality content.

Media partnerships and a limited trial

OpenAI developed SearchGPT with the collaboration of news organizations including The Atlantic, News Corp, Axel Springer, Associated Press, Reuters, Financial Times, Le Monde, Prisa and Dotdash Meredith. The company says publishers will be able to manage how their content appears in results.

That marks an important difference from the more fraught relationship some AI companies have had with content owners. Several publishing groups have challenged the use of news articles to train models without compensation or a clear agreement. By seeking partnerships from the outset, OpenAI is trying to secure access to reliable information and reduce that conflict.

The prototype will be temporary. OpenAI has said it wants to bring the features that work best into ChatGPT, rather than confirming for now a definitive launch of SearchGPT as a standalone product.

Google is no longer competing only with other search engines

The announcement comes as Google is also reshaping its flagship product. Its AI summaries promise to save time, although they have produced notable errors in their first weeks of public availability. The risk is common to everyone: a fluent answer can seem convincing even when it rests on flawed premises.

For users, SearchGPT could make a familiar task more convenient: searching, comparing sources and asking for clarification. For OpenAI, it offers a way to turn ChatGPT into a more useful tool for queries tied to current events and the open web.

The trial will measure something more difficult than conversational quality: whether an AI search engine can deliver speed without undermining the reliability of information or draining traffic from the people who produce it.

A visible citation can support the wrong sentence

Evaluation does not end by counting links. Open the reference for every sentence in the summary and ask whether it contains the fact, states it with the same certainty and is current for the query. A related page that cannot support the sentence creates the appearance of traceability without the substance.

Separate retrieval from writing too. The search system may find a weak document, miss a better one or mix dates; the model may distort correct material while summarising it. Recording query, retrieved results, used passages and response locates the failure. Without that chain, “the AI was wrong” does not say what to repair.

A minimum protocol for important search

Build a set of questions with known answers, changing current information, ambiguity and deliberate absence of evidence. Check coverage, source quality, citation entailment, uncertainty and abstention. Test follow-up questions separately: inherited context saves time but can carry a false premise.

For health, law or money, the summary does not execute the decision. It helps locate documents and compare language; the user opens the primary source before acting. When a date, price or rule may have changed, show its validity rather than only a link.

Publisher controls have two layers

OpenAI said sites could manage how they appeared in SearchGPT even if they opted out of having content used to train models. The distinction is essential: indexing to answer today and training to build a model are different operations. A serious policy asks about both and verifies actual crawler behaviour.

Useful publisher attribution includes name, a prominent link and a reason to open it. Traffic cannot be inferred from a screenshot; measure referrals, queries and conversion before and after. Watch whether the summary replaces original articles or leads readers to sources they would not have found.

Assess source diversity per query rather than through a general partner list. Check whether a local question returns local documentation, whether a scientific query prioritises the paper and whether a public fact leads to the official record. Authority depends on the subject; a prestigious publisher cannot replace the document that proves the fact.

Freshness needs a separate test. Ask the same question before and after a known update and record when the system incorporates it. An answer that cites a recent page while retaining the old fact has retrieved current material without integrating it correctly. Query date and source date should remain visible.

Abstention is an output too

Design questions whose answer does not exist, sits behind a paywall or depends on an inaccessible source. The system should separate those limits from a temporary failure and offer the documents it could verify. An incomplete answer with an explicit boundary is more useful than a complete synthesis invented to satisfy the interface.

When two primary sources disagree, SearchGPT should not choose silently. It presents the difference, dates both documents and avoids resolving it by popularity. The user can then decide whether the latest record, legal definition or issuer clarification is required.

The transferable skill is to audit a conversational answer as a chain of query, retrieval, synthesis and citation. The method works for any AI search engine: it captures speed without mistaking confident prose for evidence and returns the power to verify to the reader.