Platform
The measurement platform for AI answers about materials.
Reference Grade combines a research-grade measurement of AI answers with the technical content engineering that changes them, in four stages that run continuously.
01 · Measure
Answer Assay
Your panel is the questions your customers' engineers actually ask, not keywords.
- Applications and industries you sell into, and the grades and product lines that serve them.
- The roles who ask: materials, design, process and packaging engineers; formulators; quality and regulatory; sourcing; sustainability.
- The intents behind the questions: selection, comparison, troubleshooting, process selection, compliance, substitution, supplier discovery, cost-down.
- Evidence of what's really asked: technical-service tickets, sales FAQs, RFQs, trade-show conversations, Search Console queries.
Panels are versioned and every change is logged, so trends stay comparable. Each question runs 3–5 times per cycle on ChatGPT, Claude, Gemini and Google AI Mode, with model versions recorded, because engines vary from run to run. Every answer is parsed for the companies, trade names, grades and material classes it recommends, normalized to their current owners, with the rank and qualifiers around each mention (“best for…”, “avoid if…”).
Metrics
What we report
Every metric is a rate across repeated runs, broken out by engine, segment, intent and competitor.
| Metric | What it measures |
|---|---|
| Share of Answer | % of runs that name you: company, brand or grade |
| Recommendation Rank | Median position when named |
| Citation Share | % of runs citing at least one of your pages, and your share of all citations |
| Sourced-Mention Rate | % of your mentions backed by your own pages |
| Source Mix | Citations by source type: yours, competitors', distributors', databases, patents, papers, trade press, market research, forums |
| Spec Accuracy | % of checked claims about your products that match your current data |
| Search Rate | % of runs where the engine searched at all instead of answering from training data |
| Crawl Coverage | % of priority pages fetched by verified AI crawlers in the period |
02 · Diagnose
Source Map, access diagnostics and Spec Watch
- Source Map. For each question and segment: which pages, passages and source types win; which pages were retrieved but not cited; how old they are; HTML or PDF; and what the winning pages have in common: structure, numbers, test conditions, comparisons.
- Access diagnostics. Can engines reach your content? We check robots.txt rules for each AI agent, firewall and bot-challenge behavior, JavaScript rendering, gated or PDF-only datasheets, product finders crawlers can't use, and duplicate regional sites.
- Spec Watch. We check AI answers against your TDS for property values, discontinued or renamed grades, brand ownership after an acquisition, certifications (UL 94, FDA food contact, ISO 10993, REACH/RoHS) and sustainability claims. Every flag shows the answer, the claim, the source it came from and the fix.
03 · Engineer
Content Engineering and Technical Library
A team led by a materials scientist writes and structures the pages engines cite, from your data and with your engineers, and publishes them on your site or in a Technical Library hosted on your domain.
04 · Verify
Crawler Telemetry, re-measurement and experiments
- Crawler Telemetry. Visits from OpenAI, Anthropic, Perplexity and Google crawlers to your pages, verified against published IP ranges, from your server or CDN logs or your Technical Library. We report which pages they fetch, how often, and how long new pages wait for their first fetch.
- Re-measurement. We re-run the panel and report changes with their run-to-run spread, so you can tell signal from noise.
- Controlled experiments. For high-stakes questions, we test content variants in a controlled retrieval setup where we write every search result the model sees. That isolates what about a page changes the answer. The method comes from our published research.
Principles
How we work
- Rates, not screenshots. One answer is an anecdote. We report rates across repeated runs, with their spread.
- Evidence you can audit. Every number links to the underlying answers and sources.
- No manipulation. No fake reviews, paid “top 10” placements, press-release farms, hidden text or instructions aimed at AI. We make true information easy to cite.
- Claims backed by data. We don't publish a property value without your TDS or test report behind it.
- Your data stays yours. NDAs are standard, and client data isn't used to train models.
The engine
Measured from both sides.
Built on Assay
Reference Grade runs on Assay, the AI-answer measurement engine we built and also run for Homiere, our product for real estate. Assay captures AI answers on a recurring schedule and traces their sources. It also verifies crawler activity from both sides: what an engine says it read, and what actually arrived at the server. It is the instrument behind our published research on how answer engines choose their sources.
- Where does a ChatGPT answer come from? Cache, index and live fetch in OpenAI's search stack, measured from both sides.
- When does the model look it up? Parametric knowledge, controlled retrieval and how answer engines arbitrate between them.
Start with evidence
See what AI tells engineers about your materials.
We'll run 25 of your customers' questions across three engines and show you who gets named, which sources get cited, and what's wrong. Free, in five business days.