How we measure AI visibility
Two questions, measured separately. Does AI know you exist, and does AI pick you when someone asks?
In the 30 days ending September 19, 2026, AI knew the business 84.5% of the time (891 of 1,054 judged answers) and recommended it 26.5% of the time (1,945 of 7,352).
That gap is the work. Here is how we count it.
Recognition, the score we used to call GEO, is whether AI knows you exist. We ask engines about your category and count how often they can identify your business at all.
Pick Rate is our score for whether AI picks you. Same questions, different count: how often you are actually recommended, not just recognized.
They move independently. A business can score 100 on Recognition and 1 on Pick Rate, which means AI knows exactly who you are and still sends the customer somewhere else.
AI knows who you are 85% of the time.
It recommends you 26% of the time.
Those are real numbers from AI answers judged in the 30 days ending September 19, 2026, across the nine businesses we measure: 891 of 1,054 for knowing you, 1,945 of 7,352 for recommending you. Being known and being chosen are not the same thing, and the gap between them is the whole problem.
891 correct of 1,054 judged
1,945 of 7,352 judged
Pew Research, 900+ people tracked
How we measure it
Every cycle we put the same questions to ChatGPT, Claude and Perplexity for every business we work with. A second model then reads each answer and decides whether the business was genuinely named, rather than trusting the first model's word for it. Only answers that second reader has judged are counted. Since June 5, 2026 that is 20,765 judged answers.
The market calls this work GEO, AEO, AIO, or AI SEO. We measure it as two questions. Recognition (the score we used to call GEO) is whether AI knows you exist. Pick Rate is whether AI picks you when someone asks who is best.
Almost everyone selling AI visibility measures whether AI knows you and reports it as though it answered whether AI picks you. Those are different numbers and they move for different reasons.
On the question a buyer actually asks
- Perplexity names the business 30.6% of the time (659 of 2,153 judged)
- Claude 25.3% (640 of 2,532 judged)
- ChatGPT 24.2% (646 of 2,667 judged)
Asked about the business by name, the same 30 days: Perplexity knew it 87.6% of the time (310 of 354), Claude 85.4% (299 of 350), ChatGPT 80.6% (282 of 350).
We report these separately and never blend them into one score. A single averaged number hides which engine moved, and that is the easiest place to accidentally mislead someone.
Why being named is the whole game now
Pew Research Center tracked the real browsing of more than 900 people for a month, covering 68,879 searches. When an AI summary appeared, only about 1 in 100 people clicked a link inside it. The click is largely gone. The mention is what is left. (Pew Research Center, 2025)
What actually raises it
A peer-reviewed study from Princeton, IIT Delhi, Georgia Tech and the Allen Institute for AI tested nine different ways of changing a page across 10,000 queries. Five of them worked. The top three lifted citation rate by 30 to 40 percent: cite your sources, quote real people, and put real numbers on the page. (KDD 2024)
Not one of the five is keyword matching. This page is built the way that study says to build one, which is also why it is full of numbers and citations instead of adjectives.
What these numbers are, and are not
- Measured across nine businesses we host or manage, 30 days ending September 19, 2026, on infrastructure we run. The numbers on this page are refreshed as we publish, and each one carries its window and its denominator.
- Aggregated. No individual business is identified.
- AI engines are not deterministic. The same question asked twice can answer differently, so single-day movement is noise, not a result.
- These are our accounts, not a random sample of Kansas City businesses.
- We measure whether AI names a business. We do not promise what happens after that, and anyone promising you phone calls from an AI answer is guessing.
Published research
Our studies are pre-registered: the predictions are written down before the data comes in, and the misses are published next to the wins. Two papers so far, on Zenodo under ORCID 0009-0005-3040-9995:
- Discovery, not adoption: crawler retrieval of an llms.txt resource announced through robots.txt, across seven sites on one server, and what a server log can and cannot show (September 17, 2026). Announced file: 1,549 fetches across six sites in 165 site-days, 9.4 per site-day. Un-announced copy at the root: two fetches.
- Detecting undisclosed model changes from local session logs: a method, and a worked example in which the operator was one day off (August 26, 2026).
Written by Nathan Hall, founder of Lefty Media Co, Kansas City, Missouri.