Web research

Your agents are not stuck with what they knew at training time. Any of the 10 can search the live web, open a specific URL and read it, and look up what's trending on a platform in a given country — and then use what they found in the same breath: a document with real citations, a post draft, a summary, a spreadsheet.

There is nothing to connect. No OAuth screen, no API key, no toggle in settings. Web research is built in and on from your first message.

The three things it does

Results are summaries, not link walls

Every search result carries a formatting rule with it, before the agent ever sees it: present it as 3–5 bullets, and every URL must be a real markdown hyperlink — never a raw URL pasted into the chat.

So you don't get this:

Here are some results: https://example.com/very/long/tracking?utm_source=… https://example.org/…

You get a short read with the sources woven in as clickable links, and the agent's own take on which ones matter for what you asked.

Search the web

1. Open a conversation with any agent — the one whose job the research feeds into.
2. Ask in plain language, in any language. You don't need to say "search the web"; if the answer needs live information, the agent goes and gets it.
3. Watch the step appear in the run timeline as it works — you can expand it to see exactly what was searched and what came back.
4. Ask a follow-up. The agent keeps the results in the conversation, so "and the second one — what does it actually say?" works.

Read a specific page

Paste a URL and say what you want from it. The agent fetches the page, strips it down to the actual text, and works from that.

This is the move that makes research-then-write work: read the source, then write from it. Lucy (and eight other agents) can turn what came back into a real document with inline [n] citations and a references appendix that survives editing and reload — see Documents.

What's trending on a platform

Ask what's trending on Instagram, X, LinkedIn, TikTok or Facebook, for whichever country you care about. You get themes grouped together with a suggestion for how you'd actually post about each one — in your brand voice, if your Brand Kit is filled in.

Where research goes next

Research on its own is a tab you close. In Unyo it's an input to the thing you actually wanted:

• Into a document — an article, a press release, a competitor teardown, with real [n] citations. Documents
• Into a post — Ashley turns a trend into an editable draft card with caption, hashtags and an image. Social drafts
• Into a spreadsheet — any agent can turn what it found into a multi-tab .xlsx you can download. Spreadsheets
• Into your memory — say "remember that" and it's saved as a real entry in your Neural Core, which every agent reads on its next run. Neural Core

The boundary

Enforced limits are the point, so here they are plainly.

• It reads, it never acts. There is no tool that logs into a site, fills a form, clicks a button, or buys anything. The capability is search + fetch + extract. It doesn't exist to be misused.
• It reads what you or the search can reach. Pages behind a login, a paywall, or a bot wall won't extract. You'll be told the content couldn't be extracted — not handed an invented summary of a page nobody read.
• It's a read, not an archive. A search snippet is about 300 characters; reading a page pulls roughly 5 000 characters by default. Enough to summarise, compare and cite. Not a full crawl of a 400-page site.
• The web and the code sandbox never meet. When Blake runs Python on your data, that sandbox has no internet at all — it can't fetch a URL, can't pip install, can't phone home. Research happens outside it, and only what you asked for goes in. Data analysis
• Search costs no credits. Neither do reads, drafts or plans — 73 of our 105 priced tools are free, and the web tools carry no tool charge at all. You're billed on the model's real usage for the run — never per message, never per search. Credits and billing

Tips

• Say what you want done with it, not "search for X". "Find three benchmarks and put them in a table" beats "search for conversion benchmarks" — you skip a round trip and get the artefact.
• Give it the URL when you have it. Reading a page you chose is more accurate than hoping a search surfaces it.
• Name the country for trending. The default assumes one, and trends are extremely local. "In Spain", "in Germany", "in the UK" changes the answer completely.
• Ask in your language. Queries work in any language, and the summary comes back in whichever of the 11 agent languages you've set. See Languages.
• Push back on it. If a result looks thin, say so — "that's a vendor blog, find me something independent" works, and the sources are right there for you to check.

Troubleshooting

"Web search is unavailable." Research is temporarily out of service — the agent tells you explicitly rather than answering from memory and pretending it looked. Retry in a moment; if it persists, contact support.

"Could not extract content from this URL." The page has no extractable text for an anonymous reader — a login wall, a paywall, bot protection, or an app shell that renders everything in JavaScript. Try the AMP or print version, a mirror of the article, or paste the relevant text into the chat directly. If it's a PDF or a document you have, attach it instead — every agent can read attached PDFs, DOCX, XLSX and CSV. File attachments

The agent answered without searching. It judged the answer didn't need live information. If you want it grounded, say so: "look this up and give me the sources."

Trending results feel generic. Add the country and narrow the ask. "What's trending on TikTok" is a big question; "what's trending on TikTok in Spain that a bakery could post about" is one the agent can actually answer usefully.