DEV Community

Cover image for Mention rate or citation rate? How to compute both AI search metrics from captured answers
Ricardo Batista for cloro

Posted on

Mention rate or citation rate? How to compute both AI search metrics from captured answers

Mention rate is the share of AI answers whose text names a brand, whereas citation rate is the share of AI answers that link to a page on the brand's domain. The two numbers do not move together. In 23,625 answers from ChatGPT, Gemini, Google AI Overview, Google AI Mode and Bing AI Mode (20 October to 18 November 2026), 71.3% of the answers that named one of 12 API brands did not link to that brand's domain.

  • Mention rate was 2.0 to 4.3 times citation rate on each of the five engines.
  • The gap can also go the other way: one brand was cited in 11.9% of the answers and named in 6.1%.
  • "Citation rate" has three definitions in use. Say which one a report uses.

What is the difference between mention rate and citation rate in AI search?

Mention rate reads the answer text, whereas citation rate reads the cited sources. Both are rates over the same answers, for one engine and one period.

  • Mention rate = answers whose text names the brand / all answers. Count an answer one time, also when it names the brand five times.
  • Citation rate = answers that cite one or more URLs on the brand's domain / all answers.

Mention rate measures what the reader sees in the answer text. Citation rate measures whether the engine shows your page as a source, and it is the only one of the two that can send a visit.

Two other definitions of citation rate are in use, and they give different numbers for the same data:

  • Per retrieval: citations of a page / times the engine retrieved that page. It is a page diagnostic. It can go above 1.0 when one answer cites the same page two times, so it is not a percentage.
  • Per mention: answers that link to you / answers that name you.

Retrieval is not citation. On ChatGPT, one domain was in the source list of 27.7% of answers (1,475 of 5,317) but in the inline citations of only 9.1% (485).

How far apart are the two numbers?

For 12 search and scraping API brands, as a share of brand-answer pairs, 20 October to 18 November 2026:

Engine Brand-answer pairs Named Cited
ChatGPT 63,804 13.9% 5.0%
Gemini 65,460 15.0% 3.5%
Google AI Mode 65,508 10.3% 4.0%
Google AI Overview 63,612 10.4% 5.1%
Bing AI Mode (from 7 Nov) 25,116 14.6% 3.7%

Of the 35,774 brand mentions, 25,506 (71.3%) came in answers that did not cite the brand's domain. Between brands the spread was larger: the most-named brand was named in 30.1% of answers and cited in 17.5%, and another was named in 12.1% and cited in 0.7%.

cloro, the publisher of this measurement, shows the opposite gap. Over the same 23,625 answers, the engines cited cloro.dev in 11.9% and named cloro in 6.1%. 68.5% of the answers that cited the domain (1,917 of 2,800) did not name the brand. On the prompt "What is the difference between mention rate and citation rate in AI search?" itself, ChatGPT cited a cloro.dev page in 8 of 15 answers from 5 to 19 November, and the monitor counted 0 mentions of the brand in 72 answers from the five engines.

Which metric should you report?

Report both, per engine, on the same answers, and use the gap to choose the work:

  • Mention rate answers "does the engine recommend us?". It feeds share of voice.
  • Citation rate answers "does the engine use our pages?". It is the metric for the people who own the site.
  • Named, not cited: the name comes from training data, from other sites, or from a page the engine read and did not link. Inferred: a page on your own domain that answers the prompt can add the citation.
  • Cited, not named: the engine uses your page as evidence and credits nobody. Inferred, not yet measured: put the brand name in the same sentence as the fact the engine quotes.

Keep each engine as its own row. A pooled rate describes no engine.

How do you compute both from captured answers?

This works on any capture that keeps the answer text and the cited URLs per answer:

from collections import defaultdict
from urllib.parse import urlparse
import re

BRANDS = {  # name -> (aliases, domains)
    "acme": (["acme", "acme api"], ["acme.com"]),
}

def host(url):
    h = urlparse(url).netloc.lower()
    return h[4:] if h.startswith("www.") else h

def rates(answers):
    """answers: dicts with engine, text, cited_urls (the inline citations)."""
    n = defaultdict(int)
    hits = defaultdict(lambda: defaultdict(int))
    for a in answers:
        n[a["engine"]] += 1
        # Drop citation labels such as "acme+2" before matching names.
        text = re.sub(r"\b\S+\+\d+\b", " ", a["text"]).lower()
        hosts = {host(u) for u in a["cited_urls"]}
        for brand, (aliases, domains) in BRANDS.items():
            named = any(re.search(rf"\b{re.escape(x)}\b", text) for x in aliases)
            cited = any(h == d or h.endswith("." + d) for h in hosts for d in domains)
            k = hits[(a["engine"], brand)]
            k["named"] += named
            k["cited"] += cited
            k["named_not_cited"] += named and not cited
            k["cited_not_named"] += cited and not named
    return {key: {m: v / n[key[0]] for m, v in c.items()} for key, c in hits.items()}
Enter fullscreen mode Exit fullscreen mode

Three traps from real captures:

  1. Citation labels look like mentions. ChatGPT puts labels such as "cloro+2" in the answer text. A plain string match counts them as brand mentions.
  2. The source list is not the citation list. On ChatGPT and Bing AI Mode the source list is longer than the inline citations. Use the field the engine shows to the reader.
  3. AI Overview does not always show. Count only the searches that showed an AI Overview. In this window, 98 of 5,399 Google results had none.

For volume: for a rate near 25%, about 300 answers per engine per period give a 95% confidence interval of about plus or minus 5 points. The same model on the same prompt, minutes apart, repeated only about 40% of its cited domains in paired runs, so one answer proves nothing.

Method

This measurement is published by cloro, which sells an API that captures AI answers, and the data comes from cloro's own monitor. The cloro rows above are self-reported and should be read as such.

  • Prompt set: 202 prompts about SERP APIs, LLM scraping, AI tracking APIs and web search APIs, one answer per prompt, engine and day.
  • Window: 20 October to 18 November 2026, Bing AI Mode from 7 November. 23,625 answers: ChatGPT 5,317, Gemini 5,455, AI Mode 5,459, AI Overview 5,301, Bing AI Mode 2,093.
  • Named: the monitor's brand-mention match on the answer text. Cited: the brand domain in the inline citations, or in the AI Overview sources.
  • Panel: 12 API brands with a known domain. Brave is left out, because its name is also a common word.
  • Limits: the prompts cover one market (API buyers), so the rates do not describe all AI search.

The definitions, with a worked example on 14 answers, are in the GEO metrics guide.

Top comments (0)