What is AI visibility (GEO), and how do you measure it?
AI visibility, also called Generative Engine Optimization (GEO), is how often AI engines such as ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews name your brand when someone asks them for a recommendation.
Key facts
- Engines differ, and the same question can get a different answer on a different day, so one answer proves very little.
- Engines with web search answer from pages they retrieve, so the pages they already cite matter.
- Some work is well supported, such as being listed on pages engines already cite. Some is speculative: no major engine has confirmed it reads llms.txt.
- A number is only as good as the count of answers behind it. Ten or fewer answers is an early read.
How do AI engines decide which brands to recommend?
An engine answers from two places: the pages it retrieves when it searches the web, and what it learned during training. Engines that search lean on the first. Engines that do not lean on the second.
Each engine retrieves different pages and words its answer differently. The same engine can also answer the same question differently from one run to the next. This is why a single answer is a weak guide and why a measurement needs many answers.
What can you influence, and what can you not?
You can influence the pages engines retrieve: your own product and comparison pages, your listings on review sites, articles about you on publications, and the accuracy of community pages that mention your brand.
You cannot edit what an engine learned in training, and you cannot make an engine name you. No one can promise that a change will move an answer.
What is well supported, and what is speculative?
Contested labels every ticket strong, moderate or speculative. The label describes how well that kind of work is known to move answers that use live web retrieval. It is not a forecast for your brand.
- Strong: being listed on the pages engines already cite (digital PR). A change typically shows in 2 to 8 weeks.
- Moderate: clear claims on your own comparison and product pages, current profiles on review platforms, and a clear answer to a buyer objection. A change typically shows in 4 weeks to 3 months.
- Speculative: llms.txt and structured data. Structured data helps Google understand a business, which feeds AI Overviews indirectly. No major engine has confirmed it reads llms.txt. Treat both as cheap hygiene, not a growth lever.
How do you measure AI visibility honestly?
- Ask the questions your buyers ask, reviewed by you. Contested uses about 12 buyer questions plus 4 about your brand.
- Ask them on several engines, not one.
- Count how often your brand is named, how often it is in the top three, and how it is described. Contested combines these into a score from 0 to 100.
- Say how many answers a number rests on. Contested calls 20 or more answers high confidence, 11 to 19 fair, and 10 or fewer an early read.
- When the sample is very small, show the counts and say the result is not defensible. A brand named in 3 of 3 answers has not shown that it is always named.
- Compare with named competitors on the same questions, and call a gap too close to call when it is.
What do you do with the result?
A score alone does not change anything. Contested turns each finding into a case-file ticket: what the engine said, the page it probably came from, the fix, and when you are done. After the next audit it checks the tickets you finished and tells you Happened, Not yet or Can't tell. It does not claim a ticket caused a change.
Questions
What is GEO?
GEO means Generative Engine Optimization. It is the work of improving how often and how well AI answer engines name your brand when buyers ask for a recommendation.
Is AI visibility the same as SEO?
They overlap but are not the same. Engines that search the web retrieve pages, so pages that are found and cited matter to both. The result of AI visibility is an answer that names brands, not a list of links, and it can change from one run to the next.
Does llms.txt help my brand get recommended?
It is speculative. No major engine has confirmed it reads llms.txt. Contested labels llms.txt and structured data work as speculative evidence and calls it cheap hygiene, not a growth lever.
How many answers do I need before I trust a number?
Contested treats 20 or more answers as high confidence, 11 to 19 as fair, and 10 or fewer as an early read. An early read is a first look, not a finding.