When you ask an LLM who leads AI drug discovery,
who does it name?
100 ranked questions across 10 categories, put to frontier models with no retrieval and no hints. Every answer is a top-5. What comes back is not ground truth — it is a readout of the media and literature space the model absorbed. That makes it measurable.
Reach versus quality of placement
Horizontal axis: how many of the 100 questions an entity appears in. Vertical axis: its mean rank when it does. Bubble size scales with total visibility score. Top-right is broad but shallow; top-left is narrow but dominant.
Category footprint
Share of each category's total visibility score held by Insilico-affiliated entities — the company, its platforms, its molecules and its people. Ten categories, ten separate reputations.
The finding that reframes this project
insilico.com/robots.txt disallows ClaudeBot, GPTBot,
Google-Extended, CCBot, Bytespider and
meta-externalagent, and sets ai-train=no.
So essentially none of the ranking below comes from the company's own website. It comes from press coverage, peer-reviewed papers, conference reporting and third-party databases. Media strategy is the model-visibility strategy here — the site itself is opted out.
Read this as one sample, not a measurement
These are single-sample answers from one model. LLM rankings move with sampling temperature and prompt phrasing, so an entity moving one or two places is noise, not signal.
What is robust at n=1: presence versus absence, first-place counts, and the shape of the long tail. Treat the rest as directional until repeats land. Method and caveats →