Mentions, citations and share of voice

What do mentions and citations look like on a real brand?

Here are both, for our own brand, side by side. Attensira tracks attensira.com on ChatGPT and Perplexity; these are the 30 days to 4 October 2026, read over the Attensira MCP that morning.

• ChatGPT

  Answers read

  2,274

  Named Attensira

  90

  Linked to attensira.com

  142

• Perplexity

  Answers read

  2,100

  Named Attensira

  90

  Linked to attensira.com

  77

Two things in that table run against intuition.

On ChatGPT we were linked more often than we were named. Most guides assume the citation count is always the smaller one. Ours was not.

The two engines named us equally often and linked us very differently. The mention counts are indistinguishable statistically. The citation counts are not: the gap between 142 of 2,274 and 77 of 2,100 clears a two-proportion test comfortably.

Both make sense once you separate three different objects that most dashboards blur together.

What is a mention?

A mention is your brand named in the text the engine wrote. No link is needed. An engine can recommend you at length from what it learned in training and attach no URL at all, and that still counts.

Attensira counts mentions per run, not per occurrence. An answer that names you nine times is one mention, exactly like an answer that names you once. Enthusiasm is not frequency.

Detection is a whole-word match of the brand name, plus a host-aware match of the domain. The first chapter covers its one systematic error: brands named after ordinary words read high.

What is a citation?

A citation is one URL the engine attached to its answer, pointing at a domain you own. Engines show sources differently: numbered footnotes, a sources panel, inline links, a list at the end. Attensira normalises all of them to the same thing, a URL attributed to a domain.

The citation rate counts runs again: the share of answers that attached at least one link to your domain. An answer that links four of your pages counts once toward the rate; the four URLs still count toward the source data below.

How can an engine link you without naming you?

Here is the receipt. One of our tracked questions is "AI search visibility starting price". In the 30 days, ChatGPT answered it 57 times. It attached a link to attensira.com in 42 of those answers. It named Attensira in none of them.

We opened the answers. Each one is a short list of vendors and prices, roughly "tracking tools start around this much, done-for-you services around that much", with Semrush, Otterly and others named. Our pricing page sits in the list of sources the answer drew on. It supplied context for a price range. The recommendation went to other brands.

So a citation is not a vote. It means the engine read your page while writing the answer. Whether it then put your name in front of the reader is a separate decision, and the mention rate is the only number that tells you whether it did. Attensira's documentation puts it in one line: mentioned is not cited, and the reverse happens too.

The opposite case is common as well. An engine names a brand from memory and links a review site, a marketplace listing or a comparison article instead. That is a mention with a citation to someone else's domain, and it means the engine is describing you through other people's writing.

What is share of voice?

This is where tools disagree most, so it is worth being exact.

In Attensira, share of voice is a rate: answers that named your brand divided by all answers read in the window. Competitors get the same calculation with the same denominator. Because one answer can name four brands or none, the rates do not add up to 100%. A window where you sit at 0.40 and three competitors sit at 0.55, 0.30 and 0.20 is ordinary.

Many tools use the name for a pie: your mentions divided by all mentions of the brands on your competitor list. That version always sums to 100%, and it has a flaw built in. Its denominator is the list you typed. Add a competitor and your share falls; remove one and it rises. Nothing in the answers changed. A number that moves when you edit a settings page is not a measurement of the engines, which is why Attensira does not compute it.

The practical consequence: a share of voice figure from one tool is not comparable with one from another until you know which version each uses. Ours will usually look larger, because it is answering a different question: in what fraction of answers were you named at all?

What is source share?

Source share is the one metric here that really is a share of a whole. It is a domain's citations divided by all citations recorded in the window. Every citation belongs to exactly one domain, so the shares add up.

It is the map of who supplies the answers in your category. Ours, for the same 30 days:

• Third-party sites: publishers, blogs, review sites, forums

  Citations

  72,350

  Share of all citations

  about three in four

• Tracked competitors' own sites

  Citations

  14,004

  Share of all citations

  about one in seven

• AI platforms' own sites, such as help.openai.com

  Citations

  8,241

  Share of all citations

  about one in twelve

• Our own domains

  Citations

  1,442

  Share of all citations

  about one in sixty-seven

• Reference sites

  Citations

  1,112

  Share of all citations

  about one in eighty-seven

The two most cited domains across our 61 questions were help.openai.com, with 4,978 citations, and developers.google.com, with 4,154. Our questions are about AI search, and the engines lean hard on the platforms' own documentation to answer them.

The lesson for us is blunt. Our own pages supply a sliver of the material these answers are built from, and that sliver comes from answers to just 9 of our 61 questions, 3 of them with our name in the question. The pages that decide our category belong to other people. The third-party presence chapter covers what to do about that.

What does position mean in an AI answer?

Two different numbers share the word, so keep them apart.

Your position is where your brand is first named in the answer, counting only the names you track: you and your tracked competitors. It is always shown with its base, as in "second of five tracked names". It is useful and it is relative to your list: track one more competitor that appears before you and your position drops with the answer unchanged.

Average citation position is the average place of a domain's URLs in the engine's list of attached sources. It answers "when this domain is cited, how far down the list does it sit". It is not a rank of your brand, not a search ranking and not a leaderboard. Treat it as a weak ordering signal and only read it alongside source share or citation rate.

How should each metric move after you ship a change?

In a predictable order, if the change works, and knowing the order stops you calling a result too early.

Citations move first. A new or rewritten page has to be fetched, indexed and retrieved before it can be cited. When that happens, the page starts appearing in the sources of answers to the questions it addresses. Source share and citation rate are the first numbers to show it, and a per-page view of citations shows which URL did it, as the measuring chapter of the visibility guide explains.

Mentions move second, if at all. Being read is not being recommended. Our pricing page shows that an engine can lean on a page for months without putting the brand's name in the answer. A mention needs the page to carry a passage that names you as the answer to the question, and often needs other sites saying the same thing.

Position moves last. It only exists in answers that already name you, so it needs mentions first and enough of them to average.

So when a report says "citations up, mentions flat" a month after a change, that is not failure. It is the first half of the sequence. When it says "mentions up, citations flat", look for a third-party page that started naming you, because your own page is not what changed.

Which metric answers which question?

• Do engines name us when buyers ask?

  Metric

  Share of voice on unbranded questions

  Watch out for

  Branded questions inflate it; report them apart

• Are our own pages the material answers draw on?

  Metric

  Citation rate

  Watch out for

  A citation is not a recommendation

• Who supplies the answers in our category?

  Metric

  Source share by domain

  Watch out for

  Depends on which questions you track

• Who beats us on a question?

  Metric

  Competitors' share of voice, same denominator

  Watch out for

  Rates do not sum to 100%

• Where do we come when named?

  Metric

  Position among tracked names

  Watch out for

  Moves when you edit the competitor list

How should you read the gaps between them?

The gaps are the useful part.

• Named often, linked rarely: engines know you but describe you through other people's pages. Check which domains they cite when they name you, and work on being accurately present there.
• Linked often, named rarely: your pages are being used as evidence while someone else gets the recommendation. That was us on pricing. Make the page state plainly what you are and who you are for, so the passage the engine lifts carries your name with it.
• Neither: you are not part of the answer at all. Fix share of voice first; no amount of linkable content helps in answers where you are never considered.

Before reading any gap as meaningful, check the count on both rates and whether the difference clears a significance test. Two rates built on a handful of answers can differ by a wide margin through chance alone, which the sampling chapter shows with our own numbers.

The next chapter, building a prompt set, covers choosing the questions all of these rates are averaged over.