Citation rate denominators explained: why two teams get different numbers

Two teams get different citation rates from the same AI answers because "citation rate" is a fraction, and each team divided by a different thing. One counts answers that cited your domain out of all collected answers. Another counts your citations out of every citation on the screen. A third counts your pages out of all cited pages, or your domain out of all cited domains. All four calculations are arithmetically correct and all four answer different questions. The fix is not to pick a winner but to write the numerator, the denominator and the window beside every number you report.

In short

• A rate is only as meaningful as its denominator. Name the unit being counted (answers, citations, URLs or domains) before you compare two figures.
• In DiscoveredBy, citation rate is collected answers that cited your domain, divided by collected answers. It counts an answer once however many of your pages it links.
• The Citations screen headline is a different fraction: your share of the citations in view, capped at the newest 500. It is not meant to match citation rate.
• URL-level and domain-level fractions are legitimate, but they are variants you compute yourself, so label them as such.
• Window, engine, filters and provisional sample size change the population as much as the formula does. Record them too.

Why does the same data produce different citation rates?

Because a citation is a link, and one answer can hold many links, so "how often are we cited" has several defensible readings. You can count how many answers cited you, how many links point at you, how many distinct pages are yours, or how many distinct domains are yours. Each reading uses a different unit, and a different unit means a different denominator.

Some definitions first, since AI answers tend to lift them.

• A collected answer is a completed prompt execution: the engine ran and returned a result. A run where Google showed no AI answer is not a collected answer.
• A citation is one link an AI engine puts in its own answer, pointing at a specific URL as a source. A cited URL is the page that link points at; the same page cited many times is one URL.
• A domain, for attribution, is the registrable domain of the URL, such as example.com from blog.example.com/post. A subdomain of yours counts as you.
• A denominator is the population you divide by. It is the part of a rate that people most often forget to state.

DiscoveredBy's key terms page puts the general warning plainly: the metrics do not share a denominator, and which one you read changes what the number is telling you. This post applies that warning to citations specifically. If you want the wider picture of mentions, citations and visibility first, read AI visibility, brand mentions, and citations.

Enlarge image

What does DiscoveredBy's citation rate divide by?

Citation rate is collected answers that explicitly cited your domain, divided by collected answers, as the metrics reference defines it. The unit is the answer. If one answer links three of your pages, that answer counts once in the numerator.

Three details in the docs decide what the number means.

The denominator is collected answers, not analysed answers. Citations are recorded during collection, at the same moment the answer text is saved, so citation rate does not wait for the later step that extracts brand mentions. Brand visibility divides by analysed answers instead. That is why visibility and citation rate can read very differently on one screen for one window, and both can be right.

Citing is not retrieving. Citation rate counts a domain the answer named as a source. Where an engine reports the pages it retrieved, a page it read and never linked shows up in domain coverage (called retrieval coverage in the Explorer), which is a separate metric. The Retrieved vs cited page covers that gap.

Some engines save more than the sources an answer credits. For Claude, every page its search returned is stored as a citation, and the docs say the same of Grok and of Perplexity answers collected before Sonar. Citation rate counts saved citations, so on those engines it can include pages the answer text never credited. Check the engine mix before comparing rates.

A run with no answer is in neither side. A run where Google showed no AI answer has no text and no citations, so it is in neither the numerator nor the denominator. Separately, "answers citing sources" leaves Claude, Grok, pre-Sonar Perplexity and some Gemini (app) answers (those whose only sources are Google product or place links) out of its division entirely, for the reason above.

Which other fractions get called "citation rate"?

Several more. Four of them are real screens or metrics in the product, and two are variants you compute yourself. Knowing which is which settles many disagreements in a single conversation.

The Citations screen makes the difference explicit. Its headline is what share of the citations currently in view point at your domain, out of everything shown there, not out of every collected answer. The docs state that this makes it a different question from citation rate, so the two percentages are not meant to match.

Two properties of that screen matter for anyone exporting or screenshotting numbers. It builds both tabs from the newest 500 citations in the window that match the filter bar, and says so on the page when the window holds more. It also narrows by owner and by prompt on top of that. So a figure from that screen describes the visible slice, and you should record the filters that produced it.

Enlarge image

How do you tell which denominator someone used?

Ask what one unit of the denominator is, then reproduce it. If the unit is an answer, the number should never exceed 100% and one answer can contribute at most once. If the unit is a citation, a single answer with many links contributes many times. If the unit is a distinct URL or domain, repeated citations of the same thing collapse to one.

A quick diagnostic sequence works well in a review meeting.

1. Ask for the numerator in words: "answers that cited us" or "links pointing at us".
2. Ask for the denominator in words, and whether it is a count of answers, citations, pages or domains.
3. Ask for the window and filters (engine, country, persona, language, tags, prompt).
4. Ask whether any answers were excluded, and why.
5. Recompute the figure from the counts. If the counts reproduce the percentage, the two numbers were never in conflict.

Many "your number is wrong" disputes are settled at step 1 or 2.

The calculation worksheet

Copy this into a spreadsheet or a report appendix. Fill one row per figure you publish. The last column is the point of the whole exercise: no number leaves your team without its definition next to it.

CITATION RATE CALCULATION WORKSHEET

Project / client:            ______________________
Reported by / date:          ______________________
Window (start to end):       ______________________
Engines included:            ______________________
Filters applied (country, persona, language, tag, prompt, owner):
                             ______________________
Collection channel(s):       app / API / both (state which)

FIGURE 1
Name used in the report:     ______________________
Unit counted:                answers / citations / URLs / domains
Numerator (words):           ______________________
Numerator (count):           ______
Denominator (words):         ______________________
Denominator (count):         ______
Result (count / count):      ______ %
Exclusions and why:          ______________________
Fewer than 30 observations?  yes / no  (if yes, label provisional)
Source screen or export:     ______________________
Definition line to print beside the number:
  "[Name]: [numerator] divided by [denominator], [window], [engines]."

FIGURE 2 ... repeat

And this is the one-line definition template to print under a chart or in a table footer:

Citation rate = collected answers citing example.com / collected answers,
28 days ending 2026-09-28, all engines, no filters.

Worked example: one window, four answers to "what is our rate?"

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to legal and compliance teams. It tracks its domain against three fictional competitors, Brieflane, Clausewise and Docket North. In a 28-day window its monitoring collected 40 answers. Suppose these are the counts behind the window (assume none of the answers came from an engine that "answers citing sources" leaves out):

Enlarge image

Four colleagues each report a figure.

Nobody made an arithmetic error. Priya and Marcus disagree by a factor of two because they count different things: Priya counts the 12 answers that cited Quillstone, while Marcus counts the 18 links those answers held, out of 120 links across all 40 answers. Elena's figure is lower still because Quillstone's 18 citations land on only 6 pages, so counting each page once shrinks its side of the fraction more than it shrinks the total. Tom's figure says almost nothing about performance: one domain out of 25 will always look small, and it moves only when a new domain enters or leaves the cited set.

Now add three more that a colleague might quote.

• Answers citing sources: 34 / 40 = 85%. This says how often the engines cited anything at all. It describes the engines, not you.
• A "citation rate among answers that cited anything": 12 / 34 = 35.3%. This is not a product metric. It is a defensible variant that removes answers with no sources from the denominator, so label it as your own.
• Domain-citation share of voice: the citation counts of the tracked domains are 18 (Quillstone), 30, 24 and 12, which sum to 84. So Quillstone's share is 18 / 84 = 21.4%. The other 36 of the 120 citations sit on domains nobody added as a competitor, so they leave this denominator. Adding a fourth competitor would lower this figure without any answer changing.

Three of these numbers are honest descriptions of the same 40 answers, and the report should print all the ones it uses, each with its definition. What it should not do is put "30%" in one quarter and "15%" in the next because someone switched screens.

What changes the denominator even when the formula stays the same?

The population you divide by moves when the window, the engine mix, the filters or the sample size move, even if the formula is untouched. Record these alongside the formula.

• Window. The 7, 28 and 90-day windows (each ending yesterday) contain different answers. A 28-day figure and a 90-day figure are different populations.
• Engine mix. Adding an engine adds answers to both sides. If engines cite at different rates, your rate moves without any change in your content. See keep app and API results separate for the channel version of this problem.
• Filters. Country, persona, language, tag and prompt narrow the population. A rate for one persona is not comparable with an unfiltered rate.
• Collection gaps. Days that did not run shrink the collected population unevenly, so check that two windows you compare were collected comparably.
• Sample size. In the docs, a metric that reports an observation count is still shown below 30 observations, but the sample is too small to read as settled. For a rate over answers, treat fewer than 30 answers in the denominator the same way and label the figure provisional.
• Empty populations. A rate with a zero denominator reads as no data, never as 0%. Do not write a zero into a report for a period that had nothing to measure. The distinction between a true zero, no answer and a failed run is covered in zero mentions, no answer, or failed collection.

Common mistakes and what this cannot tell you

• Comparing across units. Putting an answer-level rate next to a citation-level share in one chart, as if they were the same measure.
• Reporting a share of voice as a citation rate. The Competitors screen's share counts citations across tracked domains. It moves when you add or pause a competitor. See share of voice versus brand mentions for the mention-based version.
• Mixing the two "observations" counts. The docs warn that counts called observations on different metrics are not the same denominator.
• Treating a Citations screen figure as a project-wide truth. It describes the newest 500 citations that match the filters.
• Reading more citations as more influence. A citation is an observation that the answer linked a page. It is not proof that the page shaped the recommendation, and a higher rate is not proof that a content change worked.
• Reading a lower rate as decline. If the denominator grew because you added prompts or an engine, a falling rate may be a change in the population, not in performance.

No denominator choice can tell you why an engine cited what it cited. A citation rate describes what happened in the answers you collected, and nothing here promises that any change to a page will move it.

Frequently asked questions

Which citation rate should we put in the headline of a report?

Use the product's own citation rate (collected answers citing your domain over collected answers) as the headline, because its unit is easy to state and it does not double-count an answer with several links. Put any other fraction beside it under its own name. Whatever you choose, keep it fixed across periods so trends compare like with like.

Why is my Citations screen percentage different from citation rate?

They divide by different things. The Citations screen shows your share of the citations currently in view, capped at the newest 500 matching the filters. Citation rate divides by all collected answers. The docs state the two are not meant to match.

Is citation share the same as citation rate?

Not in general. The glossary defines citation share as the percentage of cited sources that come from your domain, so its unit is the source, not the answer, although its own example counts prompts. Check what the person quoting it counted before comparing it with a rate.

Should I exclude answers with no sources from the denominator?

You can, but that makes it a variant, not the product's citation rate, so name it. On ChatGPT (app) and Gemini (app), an answer with no sources usually means the app did not search, which is information about the engine, not about you. Publishing both versions, clearly labelled, lets a reader see the effect.

Can a URL-level or domain-level rate be more useful?

Sometimes. A URL-level view helps when you care about how many distinct pages earn citations, and a domain-level view helps when you compare which sites dominate a category. Compute them from your exports or a spreadsheet, and label them as your own calculations.

Does a rate under 30 observations mean anything?

It means something, but it is fragile. The docs treat a metric built from fewer than 30 observations as too small to read as settled, so show the count next to the percentage and avoid drawing a trend from it.

Next step

Open the Explorer and query citation rate and answers citing sources for the same window and filters, so you can see the two denominators differ (domain citation rate needs a domain breakdown or filter, so run it as its own query). Then paste the definition line from the worksheet under each figure in your next report. To start monitoring your own citation rate, sign in or create a project. For the related screens and definitions, see Citations, Metrics defined and the glossary entry for citation share.