Machine Relations Index method

This is the open standard for measuring Machine Relations.

Version v2.1. Author: Jaxon Parrott. License: CC BY 4.0. Credit: Machine Relations Index, Jaxon Parrott — machinerelations.ai.

Concept DOI: 10.5281/zenodo.23652423.

The Machine Relations Index (MRI) is the neutral public benchmark for measuring which root domains AI answer engines cite. This methodology page is generated from the current public MRI artifact, so its counts, window, and methodology version stay aligned with the JSON dataset at /data/machine-relations-index.json.

Last generated: 2026-11-15

Current Public Benchmark #

Public data version: public view 2.0
Methodology version: v2.1
Measurement window: 2026-05-10 to 2026-11-15
Cited root domains: 37,736
Source events: 267,805
Monitored prompts: 1,255
Answer runs: 39,811
Answer engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, Perplexity
Stratum universe: 157 (category x question shape)
Measurable segments: 151
Segments that cannot accumulate evidence: 6 (excluded from every rate)
Validated source segments: 151 of 151 measurable
Segments still collecting observations: 0

What MRI Measures #

MRI measures observed citation behavior. A domain enters the public index when at least one monitored answer engine cites that root domain in the measurement window. The index does not measure SEO rankings, backlink scores, web traffic, paid promotion, or brand popularity.

The atomic unit is a source event: an answer engine cited a root domain while responding to a monitored prompt. Those source events are aggregated into domain-level source authority, source-role summaries, and source-segment citation rates.

What each engine's figures measure #

Models across the window. Whole-window engine figures pool the models used during that window. The resolved model is latest-success metadata, not attribution for every historical run. Model periods in the release manifest and engine responses count successful, public-eligible returned runs before engine-day denominator exclusions, not category-specific runs or citation counts. A null model means not recorded; missing or null periods mean historical attribution is unavailable. First and last dates do not imply uninterrupted use. A model change alone is not a collection regime.

Citation coverage and collection. Roster status includes citation yield. Degraded or quiet does not alone prove failed collection: responses can return without citations.

ChatGPT runs. They go through OpenAI's API with web search switched on. The model can change during the window; the release's model periods preserve that history, including runs whose model was not recorded. That is the retrieval layer apps and research agents built on OpenAI's API use. The signed-in ChatGPT app answers with its own models and search pipeline and cites differently: a preprint submitted on 2026-10-11 and not yet peer reviewed found that Google Search and OpenAI's API on comparable models together recovered about half of the hostnames ChatGPT Web cited (48.7 and 50.3 percent on two model settings) and 26 to 28 percent of its exact URLs (arXiv 2610.15382v1). Claude, Gemini and Perplexity runs are likewise API runs with web search; their historical models are recorded in the release's model periods when known.

Method note, 2026-09-25. From that date, Google AI Mode and Google AI Overviews runs come through a different collection path than earlier runs, with no overlapping days. Google AI Overviews answers collected on the two paths differ in how many distinct domains each cites, so its per-answer figures are read within one side of 2026-09-25, not as a trend across it. Google AI Mode changed earlier: its answers began citing far fewer distinct domains between 2026-08-10 and 2026-08-30, on the earlier path, and a same-hour comparison of the two paths on 2026-11-05 found both citing few distinct domains per answer. Google AI Mode figures from 2026-08-31 on are one collection regime, reported on their own, and are not compared or trended across the August change.

MRI Score #

The Machine Relations Index v2 reports source-segment citation rates — how often AI answer engines cite each source domain — published only once a segment clears the evidence floor of at least 10 observations across at least 7 distinct run dates, with each domain graded into confidence tiers A, B, C or collecting by how much evidence stands behind it.

Every domain an answer engine cites is measured on equal terms. For each subject category paired with a buyer question type, the model counts how many observed answer runs cited a domain and divides by the total observed runs for that segment to produce a citation rate. Whole-window engine figures pool the models used during that window. In the release manifest, engine roster.model periods splits runs total (successful, public-eligible returned runs, before denominator engine-day exclusions) by recorded model, with first and last observed dates; null means no model was recorded, and a date span does not imply uninterrupted use. resolved model is latest-success metadata, not whole-window attribution. History collected before per-engine-day collection health was recorded is excluded when an engine produced zero eligible citation events anywhere in the corpus that day and ran at least 5 successful eligible times; those days are marked zero eligible citations min 5 runs. From the day collection health is stamped at collection, a day is excluded only when that stamp says the engine was silent or dark, and it is marked collection health. Quiet days and days whose health was not measured stay in the denominator. Runs that cannot reach the numerator do not belong in the denominator. Rates are published only for segments that have accumulated enough observations across enough separate days to be stable; thinner segments are shown as still collecting rather than scored. Domains are ranked within their source type and across the full measured universe, and each domain carries a confidence grade reflecting the volume of evidence behind its rate.

The Machine Relations Index measures how often AI answer engines cite each source domain across a fixed basket of monitored buying and research prompts. It observes six engines — Perplexity, ChatGPT, Gemini, Claude, Google AI Mode, and Google AI Overviews — across a fixed basket of 24 measured subject categories plus a legacy news-topic bucket (25 taxonomy nodes in all). Findings are grouped into strata, where a stratum is one subject category paired with one question type. A stratum publishes a citation rate only after it clears the evidence floor of at least 10 observed runs across at least 7 distinct dates; below that line it is marked collecting rather than scored, so early or thin signal is never shown as settled. An engine-day is left out of the denominator when its runs cannot reach the numerator. History collected before per-engine-day collection health was recorded is excluded when that engine was asked at least five times and returned no eligible citation; each such day is marked with that backfill rule. From the day collection health is stamped at collection, a day is excluded only when that stamp says the engine was silent or dark. Quiet days and days whose health was not measured stay in the denominator. Every excluded engine-day is listed, with its run count and the rule that excluded it, in the denominator exclusion block of each release. Every domain an engine cites is measured on equal terms — membership in any external catalog is recorded only as a label and is never used to decide what gets scored. Google AI Overviews and Google AI Mode sometimes give a cited source as a Google redirect rather than the source itself. The Index records the publisher a redirect points to and leaves out one it cannot resolve; its observations from October 30 to November 11, 2026 were corrected the same way. The public Index reports rates and rankings; it excludes raw cited URLs, internal query identifiers, and provider payloads.

v1.1 was superseded by 2026-11-15 on 2026-07-05. The v1 composite score — engine breadth, query diversity, vertical spread, position quality, and temporal consistency combined into a single weighted number — is retired. v1 scores are not comparable to v2 citation rates; the two measure different things. The published v1-to-v2 crosswalk is the one-time transition receipt.

Evidence floor: 10 observed runs and 7 distinct run dates
Validated source segments expose rate and rank signals.
Segments below threshold remain visible as collection state but do not overclaim mature measurement.

Source Roles and Peer Context #

Every cited root domain is assigned a source role. Source roles make peer comparisons intelligible: analyst research should be compared with analyst research, market databases with market databases, editorial publications with editorial publications, vendor-owned sources with vendor-owned sources.

The public artifact exposes source-role counts and role-level citation totals. Domain profiles also include peer context, including benchmark-wide rank and source-role rank where the evidence qualifies.

Public Boundary #

The public dataset reports citation rates, rankings, and evidence counts. It excludes internal query identifiers, raw cited URLs, and answer-engine provider payloads. Practitioner catalog labels are downstream annotations only. They are not ranking filters in the neutral standard.

AuthorityTech publication intelligence is a practitioner projection derived from the MRI standard. It belongs downstream of the neutral dataset, not inside the dataset's citation provenance or sameAs identity.

Machine-Readable Receipts #

Human-readable standard: https://machinerelations.ai/index
Markdown standard: https://machinerelations.ai/machine-relations-index.md
Public JSON artifact: https://machinerelations.ai/data/machine-relations-index.json
V1 to v2 crosswalk: https://machinerelations.ai/data/mri-crosswalk-v1-v2.json

Adopt or implement the standard #

This is the open standard for measuring Machine Relations. Version v2.1. It is CC BY 4.0. The required credit is Machine Relations Index, Jaxon Parrott — machinerelations.ai.

Author: Jaxon Parrott. Publisher: Machine Relations Research.