Cloudflare launched Radar Researcher on August 7, 2026, with an unusually useful feature for an AI assistant: receipts. It turns plain-language questions into interactive Internet-data charts, then lets you inspect the datasets and tool calls behind the answer. If you want to know how to use Cloudflare Radar Researcher, the trick is not learning a secret incantation. It is asking a bounded question and checking that the machine answered that question—not a more convenient cousin.
That distinction matters. A fluent paragraph can hide a wrong date range with great confidence; a visible API path cannot hide nearly as well. Radar Researcher is most valuable as an inspectable data workflow, not ChatGPT wearing a network-operations lanyard.
What Radar Researcher actually gives you
According to Cloudflare’s August 7 launch post, Radar Researcher is available in beta from the header on every Radar page. You can type or dictate a question, choose a concise answer or a fuller report, revisit searchable conversations, and follow suggested next questions. Shared conversation links expire after 30 days—a sensible shelf life for analysis built on changing data.
The assistant pulls from the same API that powers Cloudflare Radar’s public visualizations, then renders the result with Radar’s existing chart components. That includes maps, time series, stacked areas, bars, donuts, and histograms—not flattened pictures of them. You can inspect the data interactively instead of reverse-engineering a screenshot into a spreadsheet like an archaeologist with a deadline.
- Blank-panel workflow: Best for open-ended investigations across several datasets.
- Explain with AI: Best when a Radar chart already shows the subject you care about.
- Search-bar workflow: Best for launching a quick question without leaving the page.
How to use Cloudflare Radar Researcher step by step
1. Start from a chart when possible
Open Radar Researcher, or browse Radar until you find a relevant chart and select Explain with AI. The second route is usually stronger. Cloudflare sends the model three pieces of context: a screenshot of the chart, the exact underlying API data, and the active location, date, and filter settings.
The screenshot tells the vision-capable model what you see. The raw values stop it from estimating numbers from pixels. The active filters stop a question about Portugal in July from becoming an answer about Europe in whatever month the model finds easiest.
2. Set the answer depth before asking
Choose a concise answer when you need one measure, one comparison, or a chart explanation. Use the fuller report for an outage timeline, a multi-country comparison, or a question that needs several Radar datasets. Starting with a report for a one-number question mostly buys you more prose to verify.
3. Ask one bounded first question
Name the geography, dates, metric, comparison, and output you want. Then run the question and open the expandable trace. Check that the assistant selected the right dataset before reading its conclusion. This is the same discipline that matters when building an MCP server on Cloudflare Workers: a clean interface is useful only if the tool call underneath it is correct.
For higher-stakes work, add a preflight: ask Researcher to restate the planned geography, date range, metric, baseline, and output before it queries anything. If that summary is wrong, correct it there. Five seconds of scope checking beats discovering after publication that “global traffic” quietly became “HTTP requests observed by one dataset.”
4. Change one variable per follow-up
After the first chart, ask for one additional country, a longer time window, or a second metric—not all three at once. Single-variable follow-ups make mistakes easier to spot. When the analysis is sound, pin the conversation or share it, remembering that the public link is temporary rather than a permanent citation.
A prompt formula that produces useful charts
Good prompts for Radar Researcher have six parts: subject + geography + time window + metric + comparison + output. That structure sounds fussy until you compare it with debugging an answer whose definition of “traffic” changed halfway through. Natural language removes API vocabulary from the entrance exam; it does not remove analytical precision from the job.
Analyze [subject] in [geography] from [start date] to [end date].
Use [metric], compare it with [baseline], and return [chart/table].
State the unit, denominator, filters, and Radar API endpoint.
Flag missing data and do not infer causation from correlation.| Goal | Copy-ready prompt |
|---|---|
| Investigate an outage | Find HTTP traffic anomalies in Kenya from August 1–7, 2026. Compare each day with the prior four-week weekday baseline, annotate recorded outages, and do not assign a cause without separate evidence. |
| Compare network quality | Compare median download speed and latency in Portugal and Spain for July 2026. Show side-by-side charts, state units and endpoints, and flag missing observations. |
| Measure bot traffic | Compare verified bot and human HTML requests in the United States and Germany for the last 30 complete days. Return percentages, denominators, and the source endpoint. |
If you do not know which metric Radar supports, ask the assistant to list relevant datasets before requesting analysis, or inspect the Radar API reference. That resembles querying structured documentation through MCP: discover the available source first, then call it. Guessing an endpoint in prose is still guessing.
Audit the trace before you cite the answer
Radar Researcher exposes how it interpreted the prompt, which datasets it searched, and the tool calls it made. Treat that trace as part of the deliverable. Before repeating a number, verify the endpoint, geography, time window, filters, unit, denominator, and comparison baseline.
- Endpoint: Does it measure traffic, DNS queries, outages, bots, or network quality?
- Scope: Is the result worldwide, regional, national, or tied to an autonomous system?
- Window: Are the dates complete, and is the comparison period equivalent?
- Measure: Is the chart showing a count, percentage, index, median, or share of bytes?
- Claim: Does the data show correlation while the prose quietly claims causation?
The machinery underneath helps explain why this is possible. Cloudflare’s unified MCP server says Code Mode represents roughly 2,500 API endpoints with three tools—docs, search, and execute—using about 1,100 tokens of context. Its comparison puts minimal native MCP schemas at roughly 244,000 tokens. Researcher can search the server-side specification instead of stapling an encyclopedia of tool definitions to every prompt.
That architecture makes discovery cheaper; it does not make every conclusion true. Keep the same posture you would use when treating an AI system as auditable: verify inputs, constrain outputs, and preserve evidence. “The chart looked official” is not a control.
What the beta cannot decide for you
Cloudflare calls Radar Researcher a beta and says it will expand the datasets the assistant understands. Missing output may mean the requested dataset is unsupported, the prompt is ambiguous, or the data is unavailable; it does not prove the underlying phenomenon never happened. The service also uses per-IP rate limiting, so it is not a bulk-export replacement for the API.
The assistant can assemble an outage timeline, but traffic data alone cannot establish motive. A country comparison can also look clean while hiding different denominators or incomplete periods, which is why the trace check matters. Because conversations preserve context, a bad premise can travel through several polished follow-ups; when the question changes materially, start a fresh conversation and rebuild the scope.
The real test is inspectability
Will inspectable, site-native agents reduce analytical errors—or simply wrap mistakes in better charts? Cloudflare is also adding controls based on the emerging WebMCP specification, which could let compatible browser agents operate Radar tools directly instead of scraping the page. That raises the stakes: the interface may become machine-facing infrastructure, not just a convenience for humans.
The useful unit of AI data analysis is not the answer. It is the answer plus the path that produced it. Cloudflare’s eventual full launch—and the dataset expansion it promised but had not dated as of August 7, 2026—will show whether Radar Researcher makes that path reliable enough for work beyond an impressive beta demo.
Get the Daily Pulse
Sharp analysis on what's actually moving in AI. No hype, no filler, no weekly digest.
