Running a benchmark
Enter a query in Enter a keyword to benchmark across LLMs… and submit. Results appear as one row per platform. Past benchmarks are kept for reference.The results table
Expand a row for the full Response from that platform and its cited sources, each with a Visit link.
Reading the comparison
The pattern across rows is the finding, not any single row: Cited everywhere. You’re an established authority for this query. Nothing to do. Cited nowhere. A content problem, not a platform quirk. The sources cited instead show what the platforms consider a good answer — compare them against your page. Cited on some platforms only. The most informative outcome. Platforms weight sources differently: some lean on documentation and community discussion, others on conventional web authority. Look at which sources the platforms citing you have in common versus those that don’t, and you have a specific hypothesis about what earns the citation. Sources count varies widely by platform. A platform citing two sources is close to winner-take-all for that query; one citing ten leaves room. Prioritize the narrow ones — being one of two is worth much more than being one of ten.The Model column
Names the specific model behind each result, where the platform exposes it. Worth recording when you’re tracking visibility over time: a change in citation behaviour often follows a model change rather than anything you did.N/A means the platform didn’t report a model.
Variance
AI responses aren’t deterministic. The same query can return different sources on different runs, so a single benchmark is a sample. For queries that matter commercially, run the benchmark a few times over a week before drawing conclusions. Consistent absence is a finding; a single absence is noise.How this differs from the single-platform pages
Benchmark first to find where you’re missing, then use the single-platform page to investigate.