General intelligence
HELM Capabilities
Stanford CRFM evaluates models on a versioned collection of capability scenarios and publishes prompts, responses, metrics and reproducible configurations.
Evidence policy
How we curate AI benchmarks
Every published number should answer four questions: who measured it, what exactly was tested, which version was used, and when the result was checked.
Curation principles
Accuracy before volume.
Shareof.ai is a curator, not the original evaluator, unless a dataset is explicitly labelled as original Shareof.ai research.
- 01
Use primary sources
Scores come from official research repositories, versioned leaderboards or the organization that conducted the evaluation.
- 02
Preserve model configurations
Dates, reasoning modes, quantization and other configuration details remain attached to the published model label.
- 03
Separate benchmark versions
We do not rank results from incompatible versions, dataset splits, judges or evaluation setups in one table.
- 04
Disclose limitations
Automated judges, contamination risks, specialist scope and other material caveats appear alongside the leaderboard.
- 05
Retain provenance
Each dataset keeps its source URL, version, metric, date checked and original organization.
Current source registry
Where the data comes from
Each source remains linked so readers can audit the original results.
Safety
HELM Safety
Stanford CRFM combines five safety benchmarks across six risk categories and publishes the evaluated prompts, responses and scores.
Long context
HELM Long Context
The benchmark combines five long-context tasks including RULER, InfiniteBench and multi-round co-reference resolution.
Healthcare
MedHELM
MedHELM evaluates model performance across a taxonomy of real-world healthcare tasks using public, gated and private clinical datasets.
Open-ended prompts
Arena-Hard v2.0 Preview
Arena-Hard v2.0 uses 500 hard prompts and 250 creative-writing prompts, with GPT-4.1 and Gemini 2.5 used as automated judges in the published configurations.
Original visibility benchmarks
Shareof.ai industry statistics are calculated from completed, measured scans. We aggregate results by category and never publish an individual prospect, domain, prompt or private report.
Public cohorts require at least 50 businesses and 200 sampled answers. Every page displays the number of companies and answers represented.
Corrections and updates
If an original source changes a result or methodology, we update the relevant page and its checked date. We do not silently combine the replacement result with an older benchmark version.
Corrections can be reported to hello@shareof.ai.
Measured, not guessed
Benchmark your own AI visibility.
See the sampled answers, competitors and cited sources behind your company’s result.
Run a free visibility check →