Real-Time Search

Real-time search is the retrieval of current information from the web or other live sources at the moment a question is asked, so that the answer reflects what is true now rather than what a model learned during training. In AI search, it is the step that lets ChatGPT, Perplexity, Google AI Mode or Copilot answer questions about today's prices, news, product releases or opening hours, and it is what produces the cited sources attached to their answers.

The alternative is answering from parametric knowledge, the information stored in a model's weights. That knowledge stops at the model's knowledge cutoff and cannot cite sources. Real-time search closes that gap through retrieval-augmented generation: the system searches, reads the results and grounds its answer in them.

Each major platform does this differently:

• ChatGPT searches automatically when a question would benefit from current information, or when the user selects Search. Its help documentation explains that it rewrites the user's question into one or more targeted queries, may send follow-up queries after reading the first results, and works with third-party search providers, naming Microsoft and Shopify (OpenAI).
• Perplexity searches the web for essentially every question and cites numbered sources inline.
• Google AI Overviews and AI Mode draw on Google's own index, including its freshness systems that favor recent content for queries where recency matters.
• Microsoft retired its public Bing Search APIs on August 11, 2025 and now offers live web data to developers through Grounding with Bing Search in Azure AI Foundry. Copilot uses Bing directly.
• Grok adds public posts from X to its web results, which gives it a different view of breaking events.

Two details explain much of the behavior. First, 'real time' usually means 'from a recently updated index', not a fresh crawl of every page at query time. Pages still have to be discovered and indexed by the platform's crawler, such as OAI-SearchBot, PerplexityBot or Googlebot, before they can be retrieved. Some systems also fetch a specific page live when a user asks for it, through agents such as ChatGPT-User or Perplexity-User. Second, the search queries the AI runs are not the user's prompt. A single question is often rewritten and expanded into several sub-queries (query fan-out), so what gets retrieved depends on those generated queries.

For GEO, real-time search is the reason new or updated content can start appearing in AI answers within days rather than waiting for the next model release. It also means visibility can disappear just as fast when a competitor publishes something more current or a platform changes its retrieval. The practical work is mostly about access and freshness signals. Keep search crawlers allowed in robots.txt and at the CDN, and check AI crawler logs for successful fetches. Keep sitemaps accurate with honest last-modified dates, and use IndexNow to notify Bing and other participating engines of changes. Put dates, version numbers and 'as of' statements on facts that change, update pages instead of leaving stale numbers in place, and serve main content in the initial HTML so crawlers that do not render JavaScript can read it.

How long a page keeps being cited varies by content type. Our analysis of the half-life of content in ChatGPT found that landing pages and documentation stay cited longest, while news pages lose half their citations within 36 days. See also content freshness.