Reference rate

"Reference rate" is a name some AI visibility reports give to the broadest possible count: in how many answers did the engine refer to your brand at all? Named in the text counts. A link to your site counts. Either one is a reference. It is a reasonable first question, and a poor number to manage by, because it adds together two events that have different causes and different fixes.

How is a reference rate calculated?

reference rate = answers that named OR linked the brand
÷ all successful answers in the window

The denominator is the same as every other rate over AI answers: each question, asked of each engine, counted once per run that came back. The numerator is any run that referred to the brand at least once, however many times.

So far this is just a wider share of voice. The trouble is in the "or".

Why does blending a mention and a citation lose information?

A mention means the engine thinks your brand belongs in the answer. A citation means it used one of your pages as evidence. They fail for different reasons:

• Named but not linked: the engine knows you from training or from other people's pages, but did not read yours. The fix is a page of yours that answers the question.
• Linked but not named: the engine read your page for a fact, then recommended someone else. The fix is the passage that says what you do, and the third-party pages that say it too.

A reference rate puts both in one column. A rise could be either, and you cannot tell which lever moved.

What does it look like on real data?

This is Attensira's own workspace: 61 questions about our category, asked of ChatGPT and Perplexity daily. In the 30 days to 4 October 2026, read over the Attensira MCP, the engines gave 4,524 answers. Across both, 4.4% named Attensira and 5.3% linked attensira.com.

Split by engine, the two rates cross. ChatGPT linked us in 6.6% of 2,349 answers and named us in 4.2%: the "read but not recommended" pattern, where it uses our pages as evidence, often on pricing questions, and then names other products. Perplexity ran the other way, naming us in 4.5% of 2,175 answers and linking us in 3.9%. A single reference rate would have shown one number per engine and hidden that the two engines need opposite work.

How does Attensira measure it?

It does not blend them. Every engine gets two rates, each a {value, n} pair:

• Share of voice: answers naming you, over all answers.
• Citation rate: answers linking your own domain, over the same answers.

Both use the same denominator, so they can be compared directly, and the gap between them is the finding. Mention detection is a whole-word match of your brand name or a host-aware match of your domain; it has no entity resolution, so a brand named after a common word reads high. Every change between two windows has to pass a two-proportion z-test at 95% before it is reported as a change.

What do people get wrong about it?

• Calling it a share. Rival brands' reference rates do not sum to 100%. One answer can refer to five brands.
• Comparing it across tools. Which questions, which engines, which countries and how many answers per day are all vendor choices. A rate only means something beside the sample it came from.
• Treating a link as an endorsement. A citation can sit under a sentence recommending a competitor.
• Expecting it to track traffic. It measures presence in answers, not visits. Nothing joins an answer someone read to a later visit.

What should you do with it?

Ask the broad question once, then split it. Find the questions where you are named but not linked, and give the engine a page of yours worth citing. Find the ones where you are linked but not named, and fix the passage and the deciding sources that describe you. Attensira's crew works the two lists separately, and the citation tracking view shows both rates per engine with their sample sizes.