Cloudflare's Radar Researcher lets users query public Internet data in plain language

Why it matters

Radar Researcher gives journalists, researchers and operators a faster route into Cloudflare's public Internet measurements while preserving access to the charts, datasets and query trace. Its conclusions remain bounded by what Cloudflare's network and related data sources can observe.

Cloudflare's Radar Researcher is a beta natural-language interface for investigating Internet traffic, outages and network quality.

Cloudflare's product lead Lai Yi Ohlsen said the tool is designed to make Radar's public Internet measurements easier to query without requiring users to navigate its dashboards or API documentation. The beta accepts typed or spoken questions and returns explanations with interactive charts. Users can also open the assistant from an existing Radar visualization, preserving the chart and its current context.

Ohlsen joined Cloudflare in 2025 after spending six years leading research and data work at Measurement Lab, an open Internet measurement project. Her public resume also lists teaching roles at Parsons Design & Technology. That background maps closely to Radar Researcher's intended audience: journalists, researchers, network operators and human-rights advocates who need to interpret Internet measurements under time pressure.

The product addresses a familiar problem in public data. Publishing an API still requires users to understand its structure, parameters and vocabulary. Cloudflare Radar's documentation covers DNS activity from the company's 1.1.1.1 resolver, HTTP traffic observed across its network, Internet quality, outages, routing and security. Radar Researcher translates a user's question into queries against that material.

Keeping the chart tied to its source data

Cloudflare says all information provided to Radar Researcher comes through the Radar API. The assistant returns the same interactive chart formats used elsewhere on Cloudflare Radar, accompanied by a plain-language explanation.

Its existing-chart workflow gives the model three inputs: a screenshot of the visualization, the exact data behind it from the API, and the active view parameters, including location, date range and filters. According to the launch announcement, the screenshot supplies visual context while the API data preserves the numerical values instead of forcing the model to estimate them from pixels.

That approach protects the values shown in a chart from transcription and rounding errors. It cannot guarantee that the assistant selected the right dataset, geography or comparison period for an ambiguous request. Radar Researcher remains in beta, and Cloudflare's announcement does not provide accuracy, latency, usage or failure-rate measurements. It also does not identify the model provider, model name, system architecture or inference cost.

Users can open an audit trace showing how the assistant interpreted a request, which datasets it queried and how it worked through the results. The beta also includes concise and extended response modes, suggested follow-up questions, voice input, searchable conversation history and shareable links that expire after 30 days.

Cloudflare's examples include comparing household Internet quality in Portugal and Spain and investigating an Internet disruption. In each case, the assistant queries fresh Radar data and presents the result with charts. For a journalist reconstructing an outage timeline, the trace provides a way to inspect the datasets behind the answer before citing it.

Cloudflare's vantage point sets the boundary

Radar's evidence reflects the Internet activity visible through Cloudflare's infrastructure. Its public documentation says the service draws from Cloudflare network traffic and aggregated, anonymized DNS data from the 1.1.1.1 resolver, alongside other sources used in specific Radar products. Those measurements offer broad coverage, while remaining bounded by the traffic and tests available to Cloudflare.

That limitation matters when an answer uses global language or compares countries. A chart may faithfully reproduce Radar's API response while still describing Cloudflare's view of an event. Researchers and operators must assess whether that dataset and vantage point support the conclusion they want to draw.

Radar Researcher's central test is whether a conversational interface can reduce the expertise required to find relevant measurements without obscuring where the numbers came from. Its interactive charts and audit trace give users a path back to the source data. The beta's usefulness will depend on whether users can consistently verify that the assistant queried the right underlying evidence.