A citation-rate leaderboard is not a market-share table. In Machine Relations measurement, a domain's citation rate usually means the share of answer runs in which that domain appeared at least once. Because one answer can cite several domains, marginal citation rates can add above 100% without any mathematical conflict.
Citation rate leaderboards measure nonexclusive AI citation outcomes #
A marginal citation rate answers one question: in how many observed runs did this source appear? It does not answer how many total citation slots the source occupied, which sources appeared beside it, or whether it displaced another source. That distinction matters for AI visibility, share of citation, and any buyer trying to compare source presence across engines.
The September 17 Machine Relations Index v2 release observed the AI Visibility & GEO category across 698 category runs. In that release, YouTube appeared in 179 of 698 runs, or 25.64%, while Reddit appeared in 154 of 698 runs, or 22.06%. Those numbers are marginal run rates: each numerator counts runs where the domain appeared at least once, and the denominator is the same 698 observed category runs. They are not mutually exclusive outcomes and they are not shares of all citation slots.
That is why the public leaderboard should be read as a source-presence table, not as a winner-take-all market map. The correct sentence is not "YouTube owns 25.64% of the citation market." The correct sentence is "YouTube appeared in 25.64% of observed AI Visibility & GEO answer runs in this release window." The first sentence silently changes the unit. The second preserves it.
This note extends the denominator discipline in Can an overall AI citation leaderboard use one category denominator? and the uncertainty discipline in Citation-rate uncertainty needs run-date clustering. Its narrower job is the nonexclusive-outcome problem: several domains can be cited in the same answer, so their marginal run rates are allowed to overlap.
A hypothetical answer-by-domain matrix shows why citation rates can sum above 100% #
One AI answer can create positive outcomes for multiple domains at the same time. The table below is hypothetical; it is not the September 17 dataset and it should not be treated as observed co-citation evidence. It is a compact illustration of the arithmetic behind marginal rates, not a claim about the observed September 17 co-citation pattern.
| Hypothetical answer run | YouTube cited? | Reddit cited? | Analyst site cited? | Total citation slots in the answer |
|---|---|---|---|---|
| Run 1 | Yes | Yes | No | 2 |
| Run 2 | Yes | No | Yes | 2 |
| Run 3 | No | Yes | Yes | 2 |
| Run 4 | Yes | Yes | Yes | 3 |
| Run 5 | No | No | Yes | 1 |
Run-level deduplication means a domain is counted once per answer run even if the answer links to that domain more than once. In this hypothetical five-run matrix, YouTube appears in runs 1, 2, and 4. Its marginal citation rate is 3/5, or 60%. Reddit appears in runs 1, 3, and 4. Its marginal citation rate is also 3/5, or 60%. The analyst site appears in runs 2, 3, 4, and 5. Its marginal citation rate is 4/5, or 80%.
Those three marginal rates sum to 200%. Nothing is wrong. The sum exceeds 100% because the categories are not exclusive. Run 1 has both YouTube and Reddit; run 4 has all three domains. A market-share table would require every unit to be assigned to one and only one row. A marginal citation-rate table has no such rule.
This is the same measurement family as collecting is not a zero citation rate and publication eligibility is not citation change: the label on the metric determines what inference is allowed. A clean denominator prevents the leaderboard from making a stronger competitive claim than the data supports.
Run-level deduplication is different from citation-slot share #
Citation rate and citation-slot share have different denominators. Citation rate divides domain-present runs by observed runs. Citation-slot share divides domain citation slots by total citation slots. Those denominators answer different questions, and combining them produces false precision.
The hypothetical matrix makes the distinction visible. Across five answer runs, there are ten citation slots: two in run 1, two in run 2, two in run 3, three in run 4, and one in run 5. If the only question is marginal citation rate, YouTube is 3/5, Reddit is 3/5, and the analyst site is 4/5. If the question is citation-slot share, then each domain's slot count must be divided by ten slots, not five runs.
| Metric | Numerator | Denominator | What it answers |
|---|---|---|---|
| Marginal citation rate | Runs where the domain appears at least once | Observed answer runs | "How often did this domain show up at all?" |
| Citation-slot share | Citation slots occupied by the domain | All citation slots in the measured set | "How much of the citation inventory did this domain occupy?" |
| Pairwise overlap | Runs where two named domains appear together | Observed answer runs, or runs containing either domain | "How often did these domains co-occur?" |
The September 17 MRI release supports the first row for the public AI Visibility & GEO category: YouTube 179/698 and Reddit 154/698. It does not expose a complete paired-overlap table for that public category. Therefore the release does not support claims such as "YouTube and Reddit co-occurred in X% of answers" or "Reddit took share from YouTube." Those claims would require the underlying answer-by-domain matrix or an explicitly published overlap table.
The separate synthetic visibility panel reports different units: 183 owned citation slots out of 562 citation slots and 31 prompt presences out of 35 prompts. That panel can help explain why units matter, but it must not be merged with the public MRI category denominator. The panel's denominator is not 698 observed category runs, and its owned-slot count is not the same kind of object as a public domain marginal count.
Marginals cannot recover intersections without more data #
A set of marginal citation rates cannot reveal the intersections between domains. If all we know is that YouTube appeared in 179 runs and Reddit appeared in 154 runs out of 698, we do not know how many runs cited both. The overlap could be low, high, or somewhere between, subject to the arithmetic limits of the observed universe.
The minimum possible overlap is not automatically zero when the sum is below the denominator, and the maximum possible overlap is constrained by the smaller marginal count. For YouTube and Reddit in the September 17 category slice, the overlap cannot exceed 154 runs because Reddit appears in 154 runs total. But the public marginal counts alone do not tell us the actual overlap. They also do not tell us which other domains were cited in the same answers, how many slots those answers contained, or whether each domain appeared once or multiple times inside an answer.
That is the central reason a leaderboard should not be narrated as a displacement story. A higher marginal rate means a domain appeared in more answer runs. It does not prove it captured exclusive market share, displaced a competitor, or dominated every citation slot in the answer. For that inference, the evidence package needs either slot-level data, pairwise co-citation data, or both.
This also explains why the Machine Relations research library separates measurement questions. What is share of citation? defines the slot-share concept. Multi-engine AI citation overlap data addresses overlap as its own object. Cross-engine citation agreement treats source consensus separately from raw source presence. Each metric needs the unit that matches the inference.
How to read the September 17 AI Visibility & GEO category leaderboard #
The September 17 numbers support a marginal source-presence reading and no stronger overlap claim. The release window spans May 10 through September 17, covers 124 observed dates, and includes six engines. The full MRI release contains 15,678 overall runs, but the AI Visibility & GEO category denominator used here is narrower: 698 observed category runs.
A careful interpretation therefore has four parts:
- State the category denominator. The relevant denominator for the cited YouTube and Reddit figures is 698 observed AI Visibility & GEO category runs, not the full 15,678-run release.
- Name the metric as marginal. YouTube's 179/698 and Reddit's 154/698 are domain-present run counts.
- Avoid additive-share language. Their rates can coexist because an answer can cite both domains.
- Reserve overlap claims for overlap data. A public marginal table does not reveal the pairwise intersection between domains.
This reading is especially important for teams comparing GEO, AEO, and broader Machine Relations programs. A dashboard can be directionally useful and still be mathematically easy to over-narrate. The discipline is not to flatten every number into a market-share story. The discipline is to preserve the denominator until the decision being made matches the metric being used.
The adjacent demand signals for this note are also deliberately modest. Machine Relations Search Console data for August 16 through September 13 contained 957 rows, 18,648 impressions, and 2 clicks. The page /research/alternative-to-brightedge had 1,374 impressions for "brightedge alternatives," and crawlers made 324 assistant-class requests to B2B AI vendor research over 24 observed days through September 16. Those signals show that vendor-comparison and measurement readers are present; they do not prove human demand for this exact title.
Machine Relations needs denominator-literate measurement #
Machine Relations treats citation measurement as a decision system, not a trophy table. A leaderboard can identify where sources are showing up, but it only becomes operational when the metric's unit is preserved. A marginal run rate tells a team whether a source is present often enough to matter. A slot-share metric tells a team how much citation inventory a source occupies. An overlap table tells a team whether sources travel together.
Those distinctions change the action. If YouTube has a high marginal rate, the question is not automatically "Who did YouTube beat?" The better question is, "What kinds of answers use YouTube as supporting evidence, and what other sources appear in those same answers?" If Reddit has a high marginal rate, the question is not automatically "What share of the market does Reddit own?" The better question is, "Which prompts and engines treat community evidence as citable?" Those are Machine Relations questions because they connect source selection to citation architecture, not just rank display.
For AuthorityTech and Machine Relations research, the guardrail is simple: publish the strongest claim the unit supports, then stop. A marginal citation-rate leaderboard can show that multiple domains are repeatedly present in answer runs. It cannot, by itself, reconstruct intersections, slot share, or displacement. That limitation is not a weakness in the data. It is the condition that keeps the interpretation honest.
Methods references for denominator discipline #
The same denominator problem appears outside AI citation measurement. Google Surveys explains that multiple-answer percentages can exceed 100% when one respondent selects more than one answer; SurveyMonkey makes the same point for checkbox response percentages; Pew Research Center has tested how select-all-that-apply formats change endorsement rates. Statistical quality guidance from NIST treats proportions as denominator-bound estimates, and NCES standards separate analysis definitions from presentation. Public-sector data guides from the Australian Bureau of Statistics, the UK Government Analysis Function, and the Office for National Statistics all emphasize that table design and survey wording must preserve what is being counted. DOJ merger guidance is useful as the contrast case: market share is computed inside a defined relevant market, not by adding overlapping source-presence outcomes. Probability notes from Duke and Stanford describe the same formal distinction in another vocabulary: marginal, joint, and non-mutually exclusive events are different objects.
Reference links: Google Surveys multiple-answer analysis, SurveyMonkey percentages over 100%, Pew Research Center select-all methodology, NIST proportion confidence interval guidance, NCES statistical standards, Australian Bureau of Statistics form design standards, UK questionnaire design guidance, ONS crime statistics methodology guide, DOJ market-share and concentration guidance, and Duke joint, marginal, and conditional probability notes.
FAQ #
Can citation rates add up to more than 100%? #
Yes. Citation rates can add above 100% when they are marginal run rates. If one AI answer cites YouTube and Reddit, both domains receive a positive outcome for the same run. The rates are not mutually exclusive shares of a fixed market.
Is a citation-rate leaderboard the same as share of citation? #
No. A citation-rate leaderboard counts how often each domain appears in observed runs. Share of citation needs citation-slot counts and a citation-slot denominator. The two metrics can point in the same direction, but they are not interchangeable.
Can we infer YouTube and Reddit overlap from their marginal counts? #
No. The September 17 public release gives marginal counts for the AI Visibility & GEO category, including YouTube at 179/698 and Reddit at 154/698. It does not publish a complete paired-overlap table, so the exact intersection cannot be recovered from those marginals alone.
Why does this matter for AI visibility dashboards? #
It matters because a dashboard can be directionally right and still be misread. AI visibility measurement should keep run presence, slot share, and source overlap separate so teams do not convert a useful source-presence signal into an unsupported market-share claim.