Read these before reading a single night's number as signal.
Assistants are non-deterministic. The same question gets different
answers. Each is asked more than once (see “repeats” per story) and the spread is shown per
assistant. A single night's number is a sample, not a fact — the 7- and 30-day views are the
ones to read.
Citing is not linking. “According to Reuters” sends nobody anywhere.
Both are tracked and reported separately. Linked share — the number with a URL attached — is
the one closer to referral traffic, and it is consistently much lower than cited share.
Citations can be hallucinated. Every cited domain is checked in DNS
and flagged if it does not resolve. A domain that resolves can still host a made-up URL; this
probe does not fetch pages to verify.
Assistants without web access cite from memory. A model with no
search draws on what it believes it would cite — a real signal about model priors, and a
different quantity from what a search-grounded assistant does. Each assistant is labeled
accordingly.
This measures assistant answers, not user traffic. It says nothing
about how many people ask assistants about the news, or click through. It is an upstream
indicator, not an outcome.
Story selection has a bias. Top stories favor wire-heavy national
and world coverage — the stories carried by the most outlets — over local or lifestyle stories
that one outlet covers strongly but few others pick up.