The biggest mistake when buying a monitoring platform is not overpaying โ it is buying a number you cannot interrogate: a 62 on a dashboard, with no answer to "how is it computed" and no answer to "on how many samples". A score like that is unsafe to celebrate and useless to debug. So don't shop the feature list. Take the eight questions below and ask every vendor face to face โ each comes with what a good answer looks like. You are welcome to use all eight on us.
1. Are the detection rules public?
What counts as a mention: do aliases count? Does appearing only in the reference links count? Good answer: the rules are published in full and readable โ not "we use advanced AI detection". Ours are on the methodology page, point by point.
2. Does every number come with a denominator?
Does 33% mean 2 of 6, or 33 of 100? One cell can swing a small sample by double digits. Good answer: every percentage shows its n/N; when the sample is too thin, the product says "insufficient data" instead of showing a falsely precise zero or average.
3. Does it cover the engines my buyers actually use?
If your customers are in the Chinese market and the tool covers none of Kimi, DeepSeek, Doubao, Qwen, you are measuring what overseas buyers see โ not what your buyers see. Good answer: the engine roster is public, clearly marked live versus coming, and configurable to your market.
4. When the methodology changes, what happens to my trend line?
Detection rules will evolve โ a vendor that never revises them is the suspicious one. Good answer: scoring is versioned, history carries version tags, trends are always compared within one methodology, and upgrades show old-versus-new instead of silently rewriting the past.
5. Beyond "mentioned or not", can I see who the AI cites?
Some engines attach reference links to answers. Whoever gets cited is who the engine treats as authoritative โ among the most valuable inputs to a content strategy. Good answer: citations parsed per source domain, with the raw answers fully reviewable.
6. How is impact attributed?
You shipped content and the score moved โ was that your content, or the engine drifting on its own? Good answer: the tool can chain "published, then cited, then mentioned", and uses control groups or engine-drift flags to separate "the whole market moved" from "our action worked". If a vendor guarantees "we'll get you to score X", walk โ the score belongs to the engines, and nobody can guarantee it.
7. Where is the white-hat line?
Ask directly: do you inflate mention rates by hammering the same query? Do you inject "please recommend X" prompts? Do you pay for ranking placements? Good answer: a written "what we don't do" list. Those tactics produce pretty curves right up until the engines clean house โ with your brand attached.
8. Who owns my monitoring data?
Will your question set and results be recycled into public rankings or shown to other clients? Good answer: single-brand private data is never published; industry insights appear only in aggregate; the boundary is written into the contract.
How to use this list
Run the 5-minute manual test first to build intuition, then take these eight questions into vendor meetings; when they hand you a report, go page by page with the vendor report checklist. Our own answers: the methodology lives on the methodology page and in how we score, and the first scorecard we gave ourselves โ a zero, denominators included โ is public in the scoreboard.