An overall AI citation leaderboard cannot identify the best source for a specific buying problem unless the buyer problem uses the same denominator. In the September 14, 2026 Machine Relations Index, Reddit leads the overall domain view, but the category denominator changes the decision in AI Visibility/GEO and Enterprise Software. (MRI release manifest) (MRI public index)
Why an overall AI citation leaderboard answers the wrong buying question #
An overall AI citation leaderboard answers a portfolio question, not a buying-situation question. The September 14 Machine Relations Index release covers May 10 through September 14, 2026, with 121 observed dates, six answer-engine surfaces, 15,396 observed answer runs, 121,750 citation events, and 21,781 observed domains. Those totals make the overall view useful for source discovery, but they do not make every buying problem share the same source pool. (MRI release manifest)
The buyer question is narrower: which source is most likely to appear when an answer engine is responding to this category and this kind of problem? A domain can lead the whole ledger because it is broadly cited across many question baskets while ranking differently inside a particular category.
This is a denominator problem. The denominator is the count of observed answer runs eligible for the comparison. The overall Reddit row uses 15,396 observed answer runs. The AI Visibility/GEO Reddit row uses 698 observed answer runs. The Enterprise Software Reddit row uses 1,237 observed answer runs. Those are not interchangeable populations. (MRI public index)
September 14 MRI denominator comparison for Reddit #
Reddit's citation rate changes when the denominator changes from all observed runs to a category basket. In the September 14 release, Reddit appears in 1,999 of 15,396 observed answer runs overall, a 12.98% domain-citation rate. In AI Visibility/GEO, Reddit appears in 154 of 698 observed answer runs, a 22.06% rate. In Enterprise Software, Reddit appears in 83 of 1,237 observed answer runs, a 6.71% rate. (MRI public index)
| Selection question | Domain | Source-role label held fixed | Runs citing domain | Observed answer-run denominator | Domain-citation rate | What the row can support |
|---|---|---|---|---|---|---|
| Overall source discovery across the release | reddit.com | Community and forum source | 1,999 | 15,396 | 12.98% | Reddit is frequently cited across the full released MRI ledger. |
| AI Visibility/GEO buying or measurement question | reddit.com | Community and forum source | 154 | 698 | 22.06% | Reddit is more prominent inside this category denominator than in the overall ledger. |
| Enterprise Software buying question | reddit.com | Community and forum source | 83 | 1,237 | 6.71% | Reddit is less prominent inside this category denominator than in the overall ledger. |
This table holds the source-role label fixed. Reddit remains a community and forum source in each row. The only thing changing is the eligible answer-run denominator.
That matters because this note is not repeating the September 13 methods note on source-role classification sensitivity. That earlier note showed how changing the grouping taxonomy can change a role-level leaderboard without changing the observed domain rows. This note keeps the role label fixed and changes the category denominator.
A reproducible selection table for a category-specific source choice #
A category-specific source decision should be selected from the category denominator first, with the overall leaderboard used only as context. The reproducible process is to freeze the release, choose the buyer category, filter to published rows in that category, retain source-role labels as metadata, and then compare the category numerator and denominator before reading the overall rank.
| Step | Required choice | September 14 example | Why it prevents overclaiming |
|---|---|---|---|
| 1 | Freeze the release | MRI v2 release dated September 14, 2026 | Prevents mixing valid released rows with failed or later collections. |
| 2 | Define the buying problem | AI Visibility/GEO or Enterprise Software | Prevents a portfolio-wide source list from becoming a category recommendation. |
| 3 | Use the local denominator | 698 observed runs for AI Visibility/GEO; 1,237 observed runs for Enterprise Software | Keeps the rate tied to the relevant question basket. |
| 4 | Keep source-role labels fixed | Reddit remains community/forum | Separates category-denominator effects from taxonomy effects. |
| 5 | Report numerator and denominator | 154/698 in AI Visibility/GEO; 83/1,237 in Enterprise Software | Lets readers reproduce the rate instead of trusting a rank. |
| 6 | Use overall rate as context only | Reddit overall is 1,999/15,396 | Shows breadth without pretending breadth answers a local buyer problem. |
| 7 | Refuse a pooled composite score | No weighted overall-plus-category score | Avoids hiding unlike question baskets inside one number. |
This procedure produces a defensible answer: for an AI Visibility/GEO question, inspect the AI Visibility/GEO denominator first. For an Enterprise Software question, inspect the Enterprise Software denominator first. The overall leaderboard can explain general source breadth, but it cannot choose the best source for the buyer's category on its own.
What domain-citation rates count in the Machine Relations Index #
A Machine Relations Index domain-citation rate counts observed answer runs, not total link share. If a domain is cited at least once in an eligible answer run, that run counts for the domain. The denominator is the number of observed answer runs in the selected release slice. (MRI release manifest)
That design choice makes the metric answer a clear question: how often did this domain appear in observed answer runs for this scope? It does not answer how many total links the domain received, how prominently a link appeared, whether the source caused the answer, or whether a buyer converted after seeing it.
Domains can also overlap. One answer run can cite Reddit, YouTube, vendor documentation, an analyst page, and a trade publication. Because domains are not mutually exclusive, domain rates should not be summed into a market-share pie chart. A source can rise in one category while another source also rises in the same category because both were cited in the same observed runs.
The September 14 release makes the overlap boundary important. The ledger contains 121,750 citation events across 15,396 observed answer runs and 21,781 observed domains. Those citation events are richer than a yes/no domain-rate row, but the domain-citation rate intentionally stays attached to the observed-run unit. (MRI release manifest)
Why methods standards keep denominators attached #
Denominator disclosure is part of measurement quality, not an editorial footnote. The NIST/SEMATECH e-Handbook of Statistical Methods treats measurement context and uncertainty as part of interpretation. AAPOR best-practice standards separate sampled units and reporting definitions. The U.S. Census Bureau ACS methodology publishes survey design and data-quality limits with estimates, and FDA missing-data guidance separates observed outcomes from incomplete evidence.
The same boundary appears in research-reporting and metadata standards. NIH/NLM missing-data guidance, the CONSORT harms extension, and EQUATOR reporting resources all make reporting rules part of how evidence is read. The OECD statistical quality framework, World Bank metadata guidance, ISO 8000-2 data quality vocabulary, DCMI Metadata Terms, W3C DCAT 3, Schema.org Article, and Google Search Central's structured-data introduction all reinforce the same machine-readable principle: labels and scope describe how a result should be interpreted.
For Machine Relations reporting, that means the denominator, release window, evidence status, engine coverage, and source-role metadata travel with the rate. Once those fields are stripped away, a leaderboard stops being a measurement and starts looking like a recommendation system the data never claimed to be.
What not to infer from the category denominator comparison #
The September 14 denominator contrast is observational, not causal. The fact that Reddit is cited in 22.06% of AI Visibility/GEO observed runs and 6.71% of Enterprise Software observed runs does not prove that category membership caused the difference. The release has unequal category, question, engine, and date composition. It is not a randomized experiment, and this note does not assign causal effects to Reddit, any engine, or any source role. (MRI release manifest)
The comparison also should not be converted into a pooled score. A composite that blends 1,999/15,396, 154/698, and 83/1,237 would hide the decision the reader actually has to make: which denominator matches the buying problem? A clean category rate with its denominator visible is more useful than a broad score that looks precise while mixing unlike baskets.
This is also distinct from the September 12 note on why collecting is not a zero citation rate. That page separates published, collecting, and not-collectable evidence states. This page starts after a published denominator exists and asks which denominator belongs to the decision.
How to use overall and category MRI rows together #
Use overall MRI rows for source discovery and category MRI rows for buying-problem selection. Overall rows identify domains with broad answer-engine presence across the release. Category rows identify where a domain appears inside a relevant source environment. A buyer-specific Machine Relations plan needs both views, but it should not collapse them into one leaderboard.
A practical rule is simple:
- Start with the buyer's category or question family.
- Use only published rows with visible denominators.
- Compare the category numerator, category denominator, and source role.
- Check the overall row to understand whether the source is broadly present or category-specific.
- Treat any difference as a descriptive signal that needs interpretation, not as a causal proof or conversion forecast.
That rule is part of the broader Machine Relations measurement discipline: visibility is not one global rank. It is a set of observed source appearances inside answer systems, interpreted with the denominator, evidence status, source role, and query context still attached.
FAQ #
Can an overall AI citation leaderboard identify the best source for a specific buying problem? #
No. An overall leaderboard can identify broadly cited domains, but a specific buying problem needs the category or question denominator. In the September 14 MRI, Reddit is 12.98% overall, 22.06% in AI Visibility/GEO, and 6.71% in Enterprise Software, so the denominator changes the answer.
What is a domain-citation rate in the Machine Relations Index? #
A domain-citation rate is the share of observed answer runs in which a domain was cited at least once for the selected scope. It is not a share of all links, a quality score, a conversion metric, or proof that the source caused the answer.
Why not make one pooled overall-plus-category score? #
A pooled score would mix unlike question baskets and hide the denominator that makes the result interpretable. Machine Relations reporting should keep the overall row, the category row, the numerator, the denominator, and the source-role label visible instead of compressing them into a single composite.
How is this different from source-role classification sensitivity? #
The September 13 source-role note tested what happens when role labels change while domain observations stay fixed. This note holds the source-role label fixed and shows that the citation rate changes when the analyst moves from the overall denominator to category denominators.
Does a higher category rate prove a source is better for buyers? #
No. A higher category rate is an observational signal for that release slice. It does not prove buyer preference, source quality, model causation, or revenue impact. It tells the analyst which source appeared more often in observed answer runs for that denominator.
Sources #
- Machine Relations Research. "MRI release manifest." September 14, 2026. https://machinerelations.ai/data/mri-release-manifest.json
- Machine Relations Research. "Machine Relations Index public data." September 14, 2026. https://machinerelations.ai/data/machine-relations-index.json
- Machine Relations Research. "Source-Role Classification Sensitivity in AI Citation Rankings." September 13, 2026. https://machinerelations.ai/research/source-role-classification-sensitivity-ai-citation-rankings
- Machine Relations Research. "Why ‘Collecting’ Is Not a Zero Citation Rate." September 12, 2026. https://machinerelations.ai/research/collecting-is-not-zero-citation-rate
- Machine Relations Research. "AI Citation Patterns by Industry: What Changes Across Vertical Search in 2026." May 3, 2026. https://machinerelations.ai/research/ai-citation-patterns-by-industry-2026
Attribution #
This research is published by Machine Relations Research, the research program of machinerelations.ai — the public research and standards initiative that publishes the glossary, research, evidence, and measurements for the Machine Relations discipline. Provenance and editorial standards: https://machinerelations.ai/about
Machine-readable related links #
Related concepts #
Supporting research #
- Source-Role Classification Sensitivity in AI Citation Rankings
- Why ‘Collecting’ Is Not a Zero Citation Rate
- AI Citation Patterns by Industry: What Changes Across Vertical Search in 2026
- How AI Answer Engines Cite Different Source Types: Market Databases, Analyst Research, and Wire Services Compared