AI citation rate: 20 definitions read, 6 different denominators

Two tools, the same brand, the same week: one reports a 12% citation rate, the other 47%. Neither got it wrong. They simply do not divide by the same thing.

To find out what the word actually covers, we read what vendors publish about their own calculation. Ten public pages from six vendors, all HTTP 200 on 28 August 2026, twenty metric definitions collected, every quoted sentence then found again in the raw downloaded HTML.

The result overturns the starting question. We went looking for the market's few competing formulas. We found seven definitions out of twenty carrying no formula at all, six different denominators behind the single word "citation", and the most precise vendor of the set publishing two contradictory denominators for the same metric, on two pages read the same day.

The method, and what it does not cover

A curl with a browser User-Agent on each of the ten pages, then a search for the sentence inside the raw HTML. All ten returned 200: zero unreachable pages. That detail matters, because a page we had failed to open would have been filed as "no formula" by mistake, inflating the result in the convenient direction.

The classification rule is mechanical. An entry counts as a published formula when the page writes both a numerator and a denominator. Otherwise it counts as prose with no ratio, however clear the sentence and however obvious the intent.

One limit, straight away: we read what vendors publish about their formula, not what their code runs. A published formula is not a verified formula. Nothing here guarantees the number shown in an interface matches the sentence shown in the documentation.

Seven definitions out of twenty carry no formula

Across the twenty definitions collected, thirteen carry a numerator and a denominator, seven are prose with no ratio. That is the first finding, and it moves the question: before comparing two formulas, both have to exist.

Two vendors out of six publish no citation rate at all, only a plain count. Semrush writes "Citations: The number of AI responses that cite your domain as a source." Otterly.AI writes "The number of times your website is referenced as a source in AI-generated answers." Those are integers. An integer cannot be compared across tools until you know how many prompts were run: 40 citations over 100 prompts and 40 citations over 4,000 prompts display identically.

Profound defines the object without publishing its rate: "A citation is a reference or link to a specific webpage, article, or resource within an answer engine response." The page does publish the Visibility Score and Share of Voice formulas. A citation rate, none.

As for Semrush's composite score, a 0 to 100 mark comparing the brand to its competitors, its formula is not published at all. The page says what the score means, never how it is obtained.

The table below carries all twenty entries. Filter by vendor or by status, click a row to read the exact sentence, its unit and its bound.

Six denominators for a single word

Among the thirteen published formulas sit six different denominators behind what the market calls a citation:

• All responses, at Scrunch AI, metric Citations
• Responses citing at least one source, at Scrunch AI, metric Citation Consistency
• All citations, counted one by one, at Scrunch AI, metric Citations by Owner
• Responses where the source was retrieved, at Peec AI, product page
• retrieval_count, one row per retrieved URL, at Peec AI, API reference
• No denominator at all, at Semrush and Otterly.AI, who publish a count

The first three come from the same vendor, on the same page. That is not an inconsistency on their side: Scrunch names three distinct metrics and defines each one properly. It is the reader who melts them into one when speaking of "the citation rate" in the singular.

Moving from the first denominator to the second is enough to shift a number without anything changing in the answers. Dividing by the responses that cite something, rather than by all responses, removes every non-citing response from the denominator. The same numerator rises, mechanically.

The Peec AI case: two pages, two denominators

This is the vendor with the most precise documentation of the set. That is exactly why it can be caught out: where a vendor publishes prose, there is nothing to contradict.

Its product page writes "Citation Rate (Domain) = Total citations of that Domain / Total responses where Domain was used as a source", and specifies that the denominator is the number of distinct responses that retrieved the domain. It even adds that dividing by retrieval_count would give "the deprecated citation_avg", not the citation rate.

Its API reference, read the same day, writes "citation_rate: sum(citation_count) / sum(retrieval_count)".

Two pages from the same vendor, two denominators for the same metric, and the first one takes care to state that the second describes something else. The difference is not cosmetic: one response can retrieve several URLs from the same domain, so retrieval_count is greater than or equal to the number of responses. The same numerator, divided by the larger of the two, yields the smaller of the two rates.

The same vendor publishes a second point worth reading twice: "Citation Rate (URLs) = Total citations of that URL / Total responses where URL was used as a source", noting that several citations inside one response push the ratio above 1. A rate that leaves the 0 to 1 range is not a bug, it is a deliberate choice: this is not a frequency, it is an intensity, a number of citations per response. The word "rate" still invites you to read it as a percentage, and that is where a dashboard goes wrong.

Four more ways of not talking about the same thing

Same name, two denominators

Peec AI's Visibility divides by "Total responses". Profound's Visibility Score divides by "the total number of responses that include at least one brand". Same metric name, two populations underneath.

On identical observations, the second is always greater than or equal to the first, since its denominator is a subset of the other. Putting both numbers side by side in a table is comparing two thermometers with different scales.

The sentence and the example describe different calculations

Profound's Share of Voice is defined as "dividing the number of responses that mention your brand by the total number of all brand mentions across all responses". Responses on top, mentions underneath: two units in one fraction, which leaves it unbounded by construction.

The example on the same page counts mentions on both sides: "if your brand is mentioned 20 times out of 100 total brand mentions". The sentence and the example do not describe the same calculation. Which one runs in the product is not knowable from outside, and it is precisely the kind of ambiguity a hurried reader never spots.

Closed denominator, open denominator

Peec AI divides brand mentions by those of the brands the project tracks: a closed denominator, whose list you choose. Similarweb takes the opposite stance and says so plainly: "Similarweb counts every brand that appears across tracked AI responses, whether you listed them or not."

The consequence is arithmetic. The same numerator over a closed denominator gives a higher number, and it rises further if you drop a competitor from your tracking list. A closed share of voice is steered from a form, not from what engines answer. The point is picked up in our definition of AI share of voice.

Otterly counts prompts, everyone else counts responses

Otterly.AI is the only tool here whose unit of observation is the prompt: "The percentage of prompts where your brand is present compared to all monitored prompts." The other five count responses.

One prompt replayed across five engines counts 1 at Otterly and 5 at Peec. Adding an engine to your tracking therefore moves your score at one vendor and not at the other, without any model having said anything new about your brand. That is the first reflex when two numbers refuse to reconcile: ask what is being counted, before asking how it is divided.

Six formulas, twelve rows, six results

To make the spread tangible, we apply six of the collected formulas to a single set of inputs.

That set of twelve observations is built by hand. No prompt was run, no engine answer was observed, this workstation having access to no model API whatsoever. These twelve rows describe no real engine, not vydera.com, not a client. They exist only to run six calculations over identical inputs, which no real dataset would allow cleanly, since every tool observes its own sample.

Across those twelve rows, the six formulas return 22.86%, 41.67%, 55.56%, 58.33%, 75% and 1.14. Between the lowest and the highest of the five percentages, a factor of 3.3. And the sixth value is not a percentage at all.

None of the six is wrong. Each answers one precise question correctly: what share of my responses cites me, what share of citing responses cites me, what share of all citations is mine, how often am I cited per response where I am retrieved, what share of responses retrieves my page, what share of my prompts is covered at least once. Six questions, six numbers, one name on the report.

Why Search Console will not build you a citation rate

The natural reflex without a tracking tool is to look for the signal in Search Console. We did it on our own property, sc-domain:vydera.com, 27 August 2025 to 25 August 2026. Three things came out, and none of them is a citation rate.

One. Google does not label AI preview impressions. The searchAppearance dimension, called live, returns HTTP 200 and zero rows on this site. That is not a failed call, it is a result: the dimension answers and returns nothing here. The control run on a much larger property of the same account does return five values, but none of them separates an AI preview impression from a blue link impression.

Two. The query dimension covers only a fraction of the site. With no dimension, the property shows 1,371 clicks and 109,907 impressions. With the query dimension, 411 clicks and 43,416 impressions spread over 1,579 queries, meaning 29.98% of clicks and 39.50% of impressions. Whatever is missing there is equally missing from any measurement built on Search Console.

Three. The signal closest to a citation is not a citation. 475 queries in the corpus name a third-party AI visibility tool. They carry 32,591 impressions, that is 75.07% of the impressions readable at query level, for 0 clicks, at an impression-weighted average position of 3.70. The site is served near the top, on queries about a competitor, and nobody clicks.

An impression is not a citation. It says a page was served, never that a model reused it in an answer. And the awkward part has to be published too: 421 of those 475 queries name one single competing tool, carrying 30,406 of the 32,591 impressions. This block describes the fan-out of one comparison article on our site, not a market law.

The second awkward part contradicts what we expected: those impressions do not come from long natural-language questions. 22,219 of the 32,591 are carried by queries shorter than six words, of the "meteoria ai visibility tool" kind. The longest query in the corpus runs 77 words, an entire persona prompt, and it returned 1 impression at position 1 for 0 clicks. Fan-out shaped queries generate volume, not traffic: 406 of the site's 411 clicks come from queries under six words with no natural-language marker.

A final warning, methodological this time: Search Console measures the shape of a query, never a surface. Nothing in these numbers proves an impression came from AI Mode rather than a classic results page. And the sample is small: 411 clicks at query level over twelve months generalise to nothing beyond vydera.com.

Five questions to ask before comparing two numbers

They take a minute per tool, and they are enough to tell whether two numbers are comparable.

1. What is the numerator? Responses, mentions, or citations counted one by one. Three objects, three numbers.
2. What is the denominator? All responses, only the citing ones, or the number of retrieved URLs. It is the strongest lever, and the least visible.
3. What is the unit of observation? The prompt or the response. One prompt run on five engines is worth 1 or 5 depending on the tool.
4. Is the denominator closed or open? Limited to the brands you track, or extended to every brand that appears. A closed denominator is steered from a form.
5. Is the ratio capped at 1? If the tool allows several citations per response in the numerator, no. It is then an intensity, not a percentage.

If a tool does not answer those five questions in its public documentation, its number remains usable to track a trend against itself. It is not comparable to another tool's, and it does not belong in the same table. We go through the vendors in our comparison of AI visibility tools, and the prompt sample feeding all these ratios in the prompt tracking definition.

What this measurement does not say

• No real citation was observed. This workstation has access to no model API: zero prompts run, zero answers recorded. The comparator's six results come from a hand-built set, flagged as such inside the tool itself
• The citation rate of vydera.com inside a market tool is not measured, since no project exists on that domain. Zero projects is not zero citations
• We read published formulas, not code. Nothing guarantees the product computes what the documentation describes
• The origin of Search Console impressions remains unknown. No dimension separates an AI preview impression from a blue link impression
• The list of third-party tools searched inside queries is closed and written in advance, and it undercounts: the 32,591 impressions are a floor, not a ceiling
• No client property was read. The scope is vydera.com alone

One method note that applies to everyone. The Search Console export used at first was capped at 1,000 rows out of 1,579 actually available. It cut 579 queries and 8,941 impressions, and it cut in the middle of the alphabetical sort, losing most accented queries. Every total published here was recomputed on the complete file. A truncated export reads exactly like a fresh measurement.

For the short definition, it lives here: AI citation rate. And if the next question is how to get cited rather than how to count it, we covered that in our article on optimising content for AI citations.