One measurement: ChatGPT, Claude, Gemini and Perplexity, each with web search,
answered the same English-language questions about metastatic breast cancer,
written from a healthcare-professional, oncology perspective, once each, on
20 October 2026. UK is this publication's market designation. No country
persona was applied when the answers were generated, and the questions were
not localised to UK clinical practice, so the answers are not specific to the
UK.
AI Visibility — the share of answers in scope that name a molecule at
least once, by its international non-proprietary name or a brand name. The
figures are QA-adjusted: mentions the automated scoring found but the answer
does not support were removed, and verified mentions it missed were added.
Named first — within a competitive class, the share of the answers naming
any member of the class in which a therapy is the first member named. The same
count as a share of all answers is shown beside it. AI Visibility and named
first come from two separate measurements of the same answers; no figure
combines them. One answer in the endocrine class could not be scored for
order and is left out of that class's base, not estimated.
Subtypes and biomarker cohorts — read from each question's wording.
HER2-low and HER2-ultralow are combined. The ESR1-mutated and PIK3CA/AKT1/PTEN
cohorts are the HR+/HER2− questions that name the biomarker; one question is
in both. The bases are small, which is why no figure for a single AI system is
shown for a subtype or cohort.
Investigational — the one status this page states, for the emerging
Phase III assets. No regulatory status in any country is stated for any other
therapy.
What this does not show
Order of mention is not a preference, a recommendation or a clinical priority,
and being named says nothing about efficacy, safety or how a treatment is
described. Therapies are compared within a clinical context, not across
contexts. Agents that were not scored, such as margetuximab, cyclophosphamide
and cisplatin, are not ranked and are not zero. One collection date, and
web-search answers change over time. Four lighter-tier AI systems with web
search, one answer per question per system; differences between systems are
mostly how many treatments an answer names, not accuracy.