How we measure.
Every figure CiteGraph publishes comes from the same procedure. It is written down here so you can argue with it.
The questions
For each product we write ten questions the way its buyers type them into an assistant, not as keywords: a straight pick (“best X for a two-person team”), a comparison, a cheapest-option, an alternative-to-a-named-rival, and specific use cases. Questions are generated from the site’s own content and category, then shown to you before anything runs — you can delete any that miss and add your own. The set is fixed between scans so movement means something.
The engines
Each question goes to three production answer engines through their official grounded-search APIs — the same retrieval path a real buyer triggers. We never scrape a chat interface, never use a cached index, and never simulate an answer. Failed calls are excluded from denominators rather than counted as a miss.
The runs
Answers vary between identical prompts, so each question runs three times per engine. A scan is therefore 10 × 3 × 4 = 120 sampled answers. Higher plans raise runs to five for tighter intervals.
Named versus cited
Two different things get counted. Named means the product appears in the answer’s prose — URLs are stripped before this test so a link never inflates it. Cited means a page was used as a source. We record both, because a product can be recommended constantly while its own words never reach the buyer.
The margins
Every rate is a sample, so every rate is published with its margin. Leaderboard positions carry Wilson 95% intervals, and a change between scans is only reported as movement when it exceeds the two margins combined. Anything smaller is noise, and we say so rather than dress it up as a trend.
Source grading
Cited pages are fetched and graded by what they actually say about the product — SOLID when the page describes it substantively, THIN for a passing mention, NOISE when the page is cited but not really about it. Pages a competitor controls are flagged, so you never waste a pitch on a rival’s blog.
What we don't do
We do not generate llms.txt files: we tested them and found no measurable effect on citation rates. We do not post on your behalf, fabricate reviews, or write claims your site cannot support — unknown facts stay marked [FILL IN] in every draft. Leaderboard positions cannot be bought, and there are no sponsored rows.
Limits worth knowing
Engines change without notice, and so can these results. Ninety answers is enough to separate a real move from noise but not enough to resolve two products a point apart. Questions are English-only for now. Region-specific answers are not yet modelled. Where a number is uncertain we print the uncertainty rather than round it away.
See it run on your own domain.
Check your site →