Verified Research

Verified research with free-search

Use search results to find pages, and take facts only from the pages. A snippet is an engine's summary of a page as it looked when it was crawled: it drops qualifiers, merges editions, and is often years old. Use this workflow whenever someone will rely on the answer.

Tool names below are the bare names on the search MCP server; your host may show them with a prefix.

0. Fix today's date

Know today's date before judging anything as current. Take it from your system context. Failing that, the retrieved … stamp on a fresh (uncached) search result is the server's clock. "Latest", "this year" and "upcoming" are relative to it. Do not infer the current year from search results or from memory.

1. Locate: search is for URLs only

• search(query) finds candidate pages. Add freshness="month" (or "day"/"week"/"year") when recency matters, include_domains=[...] once you know the official domain, category= for papers, filings, news, datasets.
• Some facts have a registry, and the search asks it on its own when the question names the thing plainly: a package ("latest fastapi version"), a CVE id, two currencies ("100 usd to cny"), a place and a weather word, a country and an indicator ("china gdp"), a city and a time word, a domain, a coin, an iOS app. The record comes back as the first result with the publisher's date, marked "Record sources", and can be cited as it is. category= ("software", "security", "weather", "finance.fx", "stats", "calendar", "reference") asks the same sources explicitly.
• Pick the primary source: the organizer's, vendor's, regulator's or author's own page, the filing, the paper, the repository. Aggregators, listicles, reposts, forums and AI-written summaries are useful only for finding it.
• Never copy a date, amount, name, rule or number from a snippet or the Lead: line into your answer.
• Search in the language of the source. A Chinese query is routed to a Chinese index automatically; to force more of them use engines=["so360", "baidu"].

2. Read the source itself

Divisions, regions, fee tables, prize breakdowns and deadlines usually sit on a sub-page or in a linked PDF (rules, FAQ, 章程, 通知). If the landing page lacks them, follow its links with fetch or read_doc before you go back to search. If fetch returns a near-empty body, retry with render="browser".

3. Check that it is current

For every source you rely on, establish three things.

1. When it was published or last updated. Look for published <date> in a fetch header, datePublished/dateModified from extract_structured, or a date in the body. no publication date found means the page is undated. Say so, and do not assume it is recent.
2. When this copy was retrieved. Pages and searches are cached for up to 7 days, and cached N days ago in the output means you are reading a stored copy. For anything that changes (deadlines, prices, availability, standings, versions, breaking news) pull it again with fetch(url, force_refresh=True), or max_age_hours=0 on search and research.
3. Which edition it describes. The year, season, version or round on the page must match the one asked about. Last year's page for an annual event is the most common wrong answer, because it is official and detailed and so looks right. If only an older edition is published, report it as that edition and say the current one is not announced.

freshness= is best-effort and keeps undated results, so a result that survives freshness="month" may still be old. The search says how many results it could date (dated: 3/10) and marks each one.