SIGNALS
Measurement

What is a good AI citation rate?

The short version

There is no portable benchmark for a good AI citation rate, and the published figures prove it rather than contradict it. MaxAEO's 2026 AI visibility benchmarks reported a median brand mentioned in roughly 31% of relevant non-branded answers across eight engines, with the top quartile clearing 58% and the top decile at 74% or higher, while AirOps' 2026 retrieval study found only 15% of the pages ChatGPT retrieved were cited at all. Those numbers are not measuring the same thing: the denominator is whatever prompt set the publisher chose, and changing the prompts changes the rate more than any site change will. Before comparing anything, check which event you are counting: BrightEdge measured ChatGPT naming brands 3.2 times more often than it cited them in August 2025, so a citation rate understates how often a buyer actually sees you. Two comparisons do hold. Your own rate against your own earlier runs on an unchanged question set, and your rate against the named competitors inside that same set. Everything else is somebody else's prompt list.

What is a good AI citation rate?

A good AI citation rate is one that is higher than your own last measurement on the same question set, and competitive with the companies named alongside you in that set. No absolute figure survives contact with a different prompt list, which is why the published benchmarks are not comparable with one another, let alone with your own set.

The reason is structural rather than a flaw in anyone's research. Citation rate is the share of prompts in a set where you are named or cited, so the denominator is a choice. A set built from narrow questions your product genuinely answers will produce a high rate. A set built from broad category questions dominated by review sites will produce a low one. Both are honest measurements of different things.

That makes the number useful in exactly two directions. Compared against itself over time, with the questions held fixed, it reads as a trend. Compared against the other companies appearing in the same answers, it reads as share of voice. Compared against a figure from an industry report, it reads as nothing at all.

What people are usually asking when they ask this question is whether they should be worried. A more answerable version is: of the questions a buyer in our category would ask, how many name us, and who is named instead? That has a specific answer for any company and it takes an afternoon to produce.

Our methodology page keeps citation outcomes in a separate series from the page-level readiness score for the same reason. Folding an outcome rate into a page score makes a week when the engines rotated their sources look identical to a week when a page got worse.

Why do published AI citation benchmarks disagree so much?

Published AI citation benchmarks disagree because they measure different units over different windows on different engines, and almost none of them share a prompt set. Reading four of them side by side makes the incompatibility obvious, which is more useful than picking one.

Published figure What it measured Source
Median brand mentioned in roughly 31% of answers, top quartile 58%, top decile 74% Non-branded prompts across eight engines, rolling 90 days to mid-2026 MaxAEO, 2026
Median half-life of 11 days from peak Citation share per page per engine, 883,000 pages Profound, 2026
Median brand at half its peak in 31 days Brand-level citation counts, 10,991 brands over 10 months Trakkr, 2026
15% of retrieved pages cited 548,534 pages retrieved across 15,000 ChatGPT prompts AirOps, 2026

The two decay figures in that table are the clearest illustration. Profound's 2026 citation decay study measured citation share per page per engine across 883,000 pages and seven engines over twelve months, producing 1.19 million page-and-engine lifecycles, and found the median page lost half its citation share within 11 days of peaking. Trakkr's "The Half-Life of AI Citations", which tracked 108,650 citation URLs at the URL level and 10,991 brands across 857,000 daily reports over ten months, found the median brand fell to half its peak citation count in 31 days.

Neither is wrong and the gap is not a contradiction. One measures a page losing its share of a category's citations; the other measures a brand losing its total count across all its pages. A brand can hold its count while every individual URL behind it churns, which is exactly what the Trakkr study found: 73.5% of citation URLs appeared once and never returned.

So when a benchmark is quoted at you, the question to ask is what the denominator was. If the answer is a prompt set you have not seen, the figure cannot tell you whether your own rate is good, only that somebody measured something.

What does a citation rate actually measure?

A citation rate measures the share of prompts in a defined set where your domain is cited, or where your brand is named, and those two are different metrics that often get the same label. Being named in the text of an answer and having a URL of yours cited as a source are separate events, and a brand can be mentioned in answers that cite nobody's pages at all.

The denominator deserves most of the attention because it is the part under your control. A prompt set is a commercial document: it encodes which buyers you care about, which use cases, which stage of the purchase, and whether competitors are named in the question. Two suppliers measuring the same company with different sets will report different rates and both will be defensible.

The numerator has a subtlety too. AirOps' 2026 retrieval study examined 548,534 pages retrieved across 15,000 ChatGPT prompts and found only 15% of retrieved pages were cited, which means a page can be fetched and read on every run and still never appear. Retrieval and citation are separate stages, and a rate that counts only citations will show a zero for a page that is one sentence away from winning.

There is also the question of what counts as an appearance. Some tools count a mention anywhere in the answer, some only count a linked citation, some count position within the answer, and some weight the first-named company more heavily. None of these is wrong; they are simply not comparable, which is why switching tools mid-programme usually produces an apparent jump or drop that reflects nothing.

The practical upshot is that a citation rate is only meaningful alongside its definition. A report that gives a percentage without stating the prompt set, the engines, the window and whether mentions count is not reporting a measurement.

Why is your citation rate lower than the number of times you are mentioned?

Your citation rate is lower than your mention rate because engines name companies far more often than they link to anyone, so a rate built on citations undercounts appearance by construction. It is the usual reason a citation rate looks alarming while the underlying position is fine.

The gap has been measured. BrightEdge's analysis of thousands of ChatGPT prompts, published on 14 August 2025 using its AI Catalyst product, recorded an average of 2.4 brand mentions and 0.74 citations per prompt, and summarised it as ChatGPT mentioning brands 3.2 times more often than it cites them. The same analysis found 44% of prompts contained no brand mention at all. BrightEdge sells software in this category, it does not publish the exact number of prompts behind the figures, and the data is from 2025 on one engine, so treat the ratio as an order of magnitude rather than a constant.

What that does to your own reporting is specific. If an engine names you in six answers out of twenty and links you in two, a citation rate says 10% and a mention rate says 30%, and both are correct measurements of different events. A programme that reports only the first will look like it is failing, while the thing a buyer actually sees is your name in the answer, which is three times more common. Our page on mention rate against citation rate works through which of the two to lead with.

Three things a tool may be counting when it reports a rate
What is counted The event Why the number differs
Mention rate Your brand name appears in the answer text The most common of the three, and the one a buyer reads
Citation rate A URL of yours is listed as a source Far rarer, and it depends on the engine choosing to attribute at all
Retrieval rate Your page was fetched while the answer was built Higher than both, and visible only in your own server logs

Report whichever one you choose consistently and define it on the slide. The failure mode is not picking the wrong metric, it is switching between them without saying so, which produces a jump or a collapse that reflects a change in counting rather than a change in standing. That is also why a tool migration mid-programme needs a parallel month on both definitions before anybody reads a trend off it.

Which AI citation metrics are there, and how is each one calculated?

Five metrics are in common use and they answer different questions, so the first job is to say which one you are reporting. Mention rate and citation rate count appearances against your own prompt set. Citation share and share of voice divide your appearances by everyone's. Citation velocity measures movement rather than level. None of them is interchangeable with another, and most of the disagreement between two tools reporting on the same company comes from quietly reporting two of these under one label.

The metric family, with the arithmetic behind each one
Metric How it is calculated The question it answers
Mention rate Prompts where your brand is named, divided by prompts in the set How often a reader sees your name
Citation rate Prompts where a URL of yours is cited, divided by prompts in the set How often your own pages are doing the work
Citation share Citations attributed to you, divided by all citations in the defined set, times 100 How much of the category's attention you hold
Share of voice Your brand mentions, divided by all brand mentions across the same answers Whether you are gaining on named competitors
Citation velocity No settled formula; see the section below Whether the direction has changed

The citation share formula is published rather than inferred. AirOps gives it as "Citation Share = (Citations attributed to an entity / Total citations in the defined set) x 100", and separates mentions from citations in the same piece: "Mentions without a link still contribute to visibility and trust but are not the same as true citations" (Oshen Davidson, What is Citation Share?, AirOps, 7 August 2026, read at source 13 November 2026). The same page carries two benchmark claims with no method attached, that brands with both a mention and a citation are up to 40% more likely to keep visibility and that only about 30% of pages stay visible from one run to the next, which is why neither is quoted here as a number to aim at.

Sample size decides whether any of them means anything. Kevin Indig's definition puts citation rate at citations divided by relevant queries, with a worked example of twelve citations across a hundred queries reading as 12%, and recommends calculating it "across a meaningful sample size" of at least 50 to 100 queries (Kevin Indig, Citation Rate, Growth Memo, 10 April 2026, read at source 13 November 2026). His version counts citations rather than prompts containing one, which parts company with the table above as soon as a single answer cites you twice, and it is worth deciding which you mean before the first run. He is a practitioner rather than a researcher and publishes no study behind the threshold, but the direction of the sample-size advice is right and matches what the variance figures below imply: a twenty-prompt set moves five points when one answer changes its mind.

Decay belongs in this family too and has a page of its own, because it is measured between runs rather than inside one. Half-life, the one-and-done rate and the reappearance rate come off the same log as the metrics above, and how to measure AI citation decay sets out the arithmetic for all three.

What is AI citation velocity, and is it a real metric?

Citation velocity is the rate at which your citations are changing rather than the level they sit at, and it is real enough to be worth watching and too unsettled to be worth comparing with anybody else's. The published definitions do not agree on the unit, so two vendors can hand you a velocity figure for the same month and both be correct.

One definition makes it a proportion. AnswerManiac gives "Citation Velocity (CV) = New Citation Appearances in Period / Total Tracked Queries", and publishes a benchmark table running from 0.02 to 0.30 by company stage, credited to "aggregated data from AEO campaigns" with no dataset, sample size or method named (AnswerManiac, Citation Velocity, 21 February 2026, read at source 13 November 2026).

The other makes it a growth rate. Writing for Texta, Steve Burk defines it as "the change in citation frequency over time" with the formula "((Current - Baseline) / Baseline) x 100", so fifty answers last month and seventy-five this month reads as a velocity of +50% (Steve Burk, AI Search Citation Velocity, DEV Community, dated 25 March with no year shown on the post, read at source 13 November 2026).

Those are different quantities. The first is bounded by the size of your prompt set and tells you how much new ground you took. The second is unbounded, moves fastest when the baseline is small, and will print a triumphant number for a brand that went from two citations to three. A report that says "velocity" without the formula underneath it is not reportable.

One claim attached to the metric travels further than its evidence. Several pages state that citation velocity spikes precede branded search increases by four to six weeks and credit the finding to Authoritas. We went looking for that research and could not find it published, so the lag is unverified here and we are not repeating it as a number. If you want to test the relationship, you already have the two series: your own velocity log and branded search volume in Search Console, read with enough lag to see a shift.

If you do report velocity, fix the definition before the first run, keep the prompt set unchanged, and report it next to the level. Direction without level tells leadership that a brand cited in two answers out of forty is outperforming one cited in eighteen, which is arithmetic rather than progress.

How much movement is normal before it means anything?

Movement of tens of percent between runs is normal on a small prompt set, which is why a single month-to-month change is rarely evidence of anything. Trakkr's "The Half-Life of AI Citations" reported that a typical brand's week-over-week citation count swings by 51.8%, across 857,000 daily reports covering 10,991 brands.

Volume is what separates signal from sampling. A brand cited many times across many prompts has a rate that moves slowly, because each individual answer is a small share of the total. A brand cited twice has a rate that halves when one answer changes, and no rewrite caused that.

The practical consequence is that the first thing a measurement programme should establish is its own noise floor, which is what running the baseline twice a week apart is for. If two identical runs a week apart differ by eight points before anything was changed, then an eight-point gain in month two is not a result.

Rotation makes this worse in a specific and predictable way. Trakkr's "The Half-Life of AI Citations" found 73.5% of citation URLs appeared exactly once and did not return, so the set of pages behind a stable brand-level number is churning underneath it. A page dropping out is therefore the normal case rather than a failure, and only a sustained pattern across several runs is worth acting on.

What does count as a real move is direction held across three consecutive runs on an unchanged question set, or a change in which companies are named rather than how often. Our guide to measuring AI citation decay works through the arithmetic, including how to size your own variance before reading a trend and why the published citation half-life figures disagree with each other, which is the reason none of them works as a benchmark.

What should you compare your citation rate against instead?

Compare your citation rate against your own previous runs and against the competitors who appear in the same answers, and stop there. Both comparisons use the same prompt set, which is the only way to hold the denominator still, and both are available from the log you are already keeping.

The time comparison is the one that tells you whether the work is working. Same questions, same engines, same counting rule, read across at least three runs. Anything else introduces a change in measurement at the same time as a change in the site, which makes the result uninterpretable in either direction.

The competitive comparison is the one that tells you whether the rate is good. If you appear in 4 of 12 shortlist answers and one competitor appears in 11, the gap is specific and diagnosable: look at what the answers cited when they named them. If every company in the category appears in roughly a third of answers, a third is what good looks like in that category, and no industry benchmark was needed to learn it.

A third comparison is occasionally worth running and often misread: your rate per engine. The engines retrieve differently, and a company can be strong in ChatGPT and absent from Google AI Overviews because one leans on vendor documentation and the other on pages that already rank. Our guide to which AI engine to optimise for first covers how to decide where that gap matters.

What is not worth doing is building a composite visibility score out of several of these. A single number that blends mention rate, position and sentiment moves for reasons nobody can trace, which defeats the purpose of measuring at all.

Which AI visibility numbers are worth reporting to leadership?

Report four numbers to leadership and keep them stable: how many questions in the set name you, how that compares with the previous run, who else is named, and which of your pages the answers cited. Those four fit on a slide and each one is traceable back to a logged run, which is the argument for automating the runs that produce them rather than assembling the figures by hand each month.

Deliberately leave the composite score out. Leadership will ask what good looks like, and the honest answer is that the comparison is against the last run and against named competitors rather than against an industry figure. Giving a benchmark you cannot defend buys one quiet quarter and an awkward conversation later, because the first person to look up a different benchmark will find one that disagrees.

Where an external figure genuinely helps is in explaining variance rather than in setting a target. Telling a board that the median brand's week-over-week citation count swings by 51.8%, as Trakkr's "The Half-Life of AI Citations" measured, is what buys permission to report trends over quarters instead of months. The ConvertMate GEO Benchmark 2026 finding that 83% of AI citations come from pages outside Google's organic top 10 serves the same purpose for the question of why rankings do not predict this.

Pair the four numbers with one sentence on what changed on the site before them. A report that shows movement without naming the change behind it cannot be used to decide what to do next, which is the only reason to produce it.

If you do not yet have the first run to report, a free visibility assessment produces exactly those four numbers across ChatGPT, Claude, Perplexity and Google AI Overviews, with the prompt log, which is the baseline every later comparison needs.

What else do people ask about AI citation rates?

Is a 20% citation rate good?

It depends entirely on which prompts are in the denominator, which is why MaxAEO's 2026 benchmark study found a median of 31% and a top decile of 74% on one fixed prompt panel alone. On a set of narrow questions your product directly answers, 20% is low. On a set of broad category questions where review sites dominate the retrieval, 20% is strong. The figure becomes meaningful once you know your own previous rate on the same set and the rate of the competitors appearing beside you.

What is the difference between citation rate and share of voice?

Citation rate is the share of prompts in your set where you appear. Share of voice is your share of all the companies named across those same answers, so it moves when a competitor gains even if your own appearances are unchanged. Both come off the same run log, and reporting them together is what distinguishes losing ground from a category getting more crowded.

Why did our citation rate drop when we changed tools?

Almost certainly because the new tool counts differently or uses a different prompt set, not because anything happened to your visibility. Tools vary in whether they count unlinked brand mentions, whether they weight position in the answer, which engines they query and how often. A tool change resets the series, so the sensible move is to run both for a period or to treat the switch as a new baseline.

How many runs before a trend is real?

Three consecutive runs moving in the same direction on an unchanged question set is the usual threshold, and the reason is the size of the underlying noise. The Trakkr citation decay study measured a typical brand's week-over-week citation count swinging by 51.8% across 857,000 daily reports, so two points of data cannot distinguish a trend from the normal churn. Sizing your own variance with two baseline runs a week apart is what makes the threshold defensible for your set.

Should we track citation rate per page or per brand?

Both, because they fail differently and the gap between them is diagnostic. Profound's 2026 study measured per-page citation share and found a median half-life of 11 days from peak, while Trakkr measured brand-level counts and found a 31-day median. A brand holding its count while its URLs churn is working as intended; a brand losing its count while individual pages hold is usually a coverage problem rather than a decay problem.

Related guides

The Assessment

Get the only benchmark that transfers: your own first run.

A free visibility assessment runs your buyer questions across ChatGPT, Claude, Perplexity and Google AI Overviews and reports how many name you, who is named instead, and which pages the answers cited, with the prompt log included.

Request a free visibility assessment →
SIGNALS · A BlackSig Systems company