How to measure GEO: mentions, citations and inquiries

To measure GEO, fix the test conditions first, then look at whether the brand appears, whether it’s described accurately and whether there’s a cited basis. Record site visits and qualified inquiries separately — a rise in AI brand mention rate can’t be read straight as sales growth.

Build metrics you can recalculate

Metric	Definition used here	Evidence to keep
AI brand mention rate	Answers correctly mentioning the target brand ÷ valid answers	Raw answers and a check of the brand entity
First-recommendation rate	Answers recommending the target brand first under the agreed rule ÷ valid answers	Whether there was an explicit recommendation, ordering or tie rule
Citation rate	Answers citing an agreed target domain ÷ valid web-search answers	Cited URLs, page content and how they map to the answer
Description accuracy	Check business, product, price and other facts; mark each error	Current fact source and the wrong statement
Qualified inquiries	Deduplicated real inquiries meeting the agreed requirement criteria	Known source, requirement details and follow-up status

When the denominator is zero, mark “not calculable”. Failed requests, CAPTCHAs or unfinished tests aren’t the same as the brand not appearing; list failures separately. Citing your site doesn’t necessarily mean the answer recommended your brand, and a correct mention doesn’t necessarily come with a citation.

Keep branded and non-branded questions apart

“What does Brand X do?” checks information accuracy; “How should a company like ours choose a supplier?” shows whether you appear in buying scenarios. Mix the two and adding more branded questions can make mention rate look like it went up.

When the same question is tested repeatedly, each answer is one record and the number of rounds is fixed; if you summarize at the question level, state the aggregation rule. Report each platform in its own table; don’t merge them into an “all-platforms” score without saying how.

A log you can use as-is

Field	What to enter
Sample	Question ID, full wording, branded or buying question
Environment	Platform entry point, date, language, region, web-search state
Result	Completed or not, correct mention, first recommendation, cited URLs
Review	Archived answer, factual errors, human reviewer
Version	Question-set version, page edit dates, retest batch

This is a blank template, not a measurement. Keeping a group of questions you deliberately didn’t optimize for helps you check whether content was only tuned to the known test questions — though that control alone still can’t prove causation.

How to read historical project numbers

Our Kechuang Yihao case records a 15-question snapshot on Tencent Yuanbao on June 15, 2026: mentioned 15/15, recommended first 12/15. The Shanghai Industrial Transformation institute case records a 10-question snapshot on June 14: mentioned 9/10, recommended first 7/10. These are dated project records; they don’t represent current results and aren’t an expected value for any other brand.

A proper GEO evaluation keeps the full time series, including declines, no-change rounds and failures; when the sample, platform or entry point changes, mark a new baseline. Sampling changes, competitors’ content and platform updates can all affect results.

Connect retest results to the next step

If the description is wrong, check the fact source first; if a citation is outdated, check the document version; if you appear but get no clicks, ask whether users need to visit your site at all; if there are visits but no inquiries, check whether the page answers buying questions and whether the hand-off to sales is smooth. Don’t treat every problem as a reason to publish more articles.

To check your site first, see the GEO audit scope; if you’re buying a service, use the GEO pricing checklist to confirm the retest workload, then contact Moments.