Search the news
The same recall with the clock turned up. Not a second search: it builds the same query and runs the same ranking, and differs in exactly three ways.
The freshness window defaults to a week rather than unbounded. An unbounded news search returns the best-ranked article on a subject, which is routinely years old.
Publishers can be filtered by list, with sources and exclude_sources.
sort_by=date is available and means what it says.
Sending category here is rejected rather than ignored, because overriding it silently would leave you believing the parameter did something.
Body
512Accepted as an alias for query.
Search depth, which sets how many pages one search reads: instant up to 10, deep up to 50, deep_reasoning up to 100. The budget is absolute rather than a multiple of limit, and a search never reads more pages than it recalled.
instant, fast, auto, deep, deep_reasoning 1 <= x <= 50Fetch and extract page bodies for the top results. Off is roughly ten times faster and returns engine snippets only.
json, markdown, text Restrict to one host.
any, day, week, month, year general, news, science, images, videos, it Ask each enriched page for what it declares about itself: JSON-LD, OpenGraph, Twitter cards, microdata, feeds and contacts. Implies enrich, because there is nothing to read declarations off a page that was never fetched. Free: pages read is what is billed and the page was fetched either way.
Only these publishers. Accepts a bare hostname or a pasted URL, and matches subdomains. Unlike site, which goes upstream as a search operator, a list is applied after recall, because no engine accepts one as a single query.
Never these publishers. An exclude always beats an include.
date orders newest first, applied after ranking and before the list is cut to the limit, so it means the newest of everything that matched rather than the newest of the handful already on top. Results with no date sink below dated ones rather than being dropped.
relevance, date