Mention rate vs citation rate: named in the answer, or linked under it

How is each one measured?

Differently, and the difference matters for how far to trust each.

Citations are read from structured data. OpenAI's web search returns url_citation annotations carrying the URL, title and location of each cited source. Perplexity returns a search_results array with the URL, title and date of each source. Nothing has to be inferred: the platform says which URLs it used.

Mentions are recovered from text. In Attensira, a mention is a case-insensitive whole-word match of your brand name, or a host-aware match of your domain, against the answer. There is no entity resolution. Word boundaries stop "Arc" matching "architecture", but they cannot stop a brand called Linear matching "linear time". The error runs one way: a brand whose name is an ordinary word reads high, never low. That is why every answer is stored verbatim, so a surprising rate can be checked by reading the answers behind it.

Why do the two diverge?

Because several ordinary mechanics separate them. An engine can name you from what it already knows, with no fresh page to link. It can recommend you and cite a review site or a comparison article that talks about you, which is a mention with a citation to someone else. And it can quote your explainer as the source for a definition, or your pricing page as the source for a price, while recommending your competitors in the same answer, which is a citation with no mention.

Our documentation says citation rate is almost always lower than mention rate, because a model has to consider you before it links you. For most brands on most prompt sets that holds. "Almost always" is not "always", and our own data is a clear exception.

Which prompts move which rate?

In our data the split is clean. Prompts that ask for a list of tools, such as "best GEO tools", are where mention rate lives: the answer is a shortlist of names, and the engine rarely needs anyone's explainer. Prompts that ask how something works or what it costs, such as "what is GPTBot" or "AI search visibility starting price", are where citation rate lives: the answer is an explanation with sources, and on those two prompts ChatGPT linked our pages in 12 of 33 and 42 of 57 answers while naming us in none.

So the two rates are, in practice, also two halves of a prompt set. A company that only tracks shortlist prompts will see mentions and conclude its content does not matter. A company that only tracks explainer prompts will see citations and conclude it is recommended. Both conclusions come from choosing half the questions, which is why a prompt set worth tracking holds both kinds.

What should you do when they diverge?

Read the direction first.

If mention rate is high and citation rate low, you are known but not read: engines describe you through other people's pages. The work is publishing pages that answer the prompts directly, getting them fetched by the engines' crawlers, and checking which of your pages, if any, are picked up.

If citation rate is high and mention rate low, as on ours, you are read but not recommended: engines use your pages for facts and name others as the answer. The work is making the recommendation case on the pages engines already trust: comparison and alternatives pages that state plainly what you are for, and accurate placements on the third-party pages that answers lean on for shortlists.

If both are low, fix mention rate first. No amount of linkable content helps in answers where you are never considered.

What does share of voice mean here?

In Attensira, share of voice is the mention rate: the fraction of successful runs in which a model named your brand. It is a rate, not a slice of a pie, and competitor rates on the same denominator do not add up to a whole. Many tools in this category use the name for a pie version instead, dividing a fixed total of mentions between brands. Neither is wrong, but they are not comparable, and a report that puts one tool's share of voice next to another's is comparing two different quantities.

The difference shows up quickly in practice. On a pie version, your share can rise because a competitor dropped out of the answers, even if you were named no more often than before. On the rate version, your number moves only when the answers that name you change. If you are reporting progress, the rate is the one that cannot flatter you.

Ask any tool which one it means before you compare its number with anything.

How many answers do you need before either rate means anything?

More than most dashboards show. Each prompt, engine and country is read at most once per workspace-local day. One sample per daily reading (n=1), which is what Starter takes, gives a rate that can only be zero or one for that prompt on that day, so it takes weeks before a per-prompt rate settles. Three samples per daily reading (n=3), on Growth and Business, settle faster because each day contributes three draws. Across a whole prompt set the numbers stabilise sooner still, which is why our 4,524-answer figures are steadier than any single prompt's.

Read weeks, not days, and compare prompt sets, not individual prompts, unless a prompt has accumulated a few hundred answers. When a tool shows a rate without its n, ask for the n before you act on it.