How the Machine Relations Index Measures AI Citations

The Machine Relations Index measures which sources AI answer engines cite, as a rate over a fixed set of monitored answer runs, and it holds every published figure to a written update standard: each figure carries its release and window, each methodology change is logged here with its date and its effect, and each update reaches every page that carries the figure. This page is the Index's methodology reference and its update log. When an AI citation figure moves, this is where the reason, the size of the change and the date are recorded.

How the Index measures #

The Index monitors a fixed registry of buyer prompts across six answer engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity. Each prompt run on each engine is one monitored answer run, and each run records the sources the engine cited. Two of those six, the Google answer surfaces, each stopped returning an observable answer for a stretch of the window and have returned answers again since late September; a third, Gemini, answers intermittently, and the release of 2026-10-07 publishes it as quiet. The dates, the size and the effect are recorded under Known data issues below, and the citations each engine contributed outside those stretches are real observations that remain in the Index.

Every published per-domain figure is a direct observation over those runs:

Citation rate: the share of monitored answer runs in which the domain was cited, on one denominator for every domain in the release.
Cited runs: the count behind that rate.
Days cited: how many distinct days in the window the domain was cited on.
Observed engines: which of the six engines cited it.
Confidence: a grade of A, B or C, published only above the evidence floor.

The evidence floor. The Index publishes a rate only where it has seen at least 10 observed runs across at least 7 distinct dates. Below that line a domain is marked collecting rather than scored. Collecting is not a zero rate; it is the instrument declining to publish a number the evidence cannot yet support.

Release identifiers. Every release carries an identifier of the form mri_score_<methodology version>+<date>+<hash> and a stated window, so any published figure can be traced to the run that produced it. Releases before the blind-engine-day correction below carry mri_score_v2.0; releases with it carry mri_score_v2.1.

Known data issues #

The Index keeps a dated list of known issues affecting its own reported data, on the model of Google's Search Console data anomalies log cited below. Each entry states the dates it covers, its size against a named release's denominator, and what it does and does not affect.

Answer runs from a non-answering engine stayed in the denominator, so affected citation rates are floors #

The defect. A monitored answer run is counted in a citation rate's denominator if the run completed, and a run is counted in the numerator only if the engine named a source in it. Those are two different populations, and nothing in the scoring path reconciled them. So on a day when an engine ran normally but returned no citation anywhere in the corpus, its runs entered every affected denominator and could not reach any numerator. The rate that comes out is arithmetically lower than the rate the evidence supports. Every affected figure understates. Correcting this raises the Index's own published numbers, including our own domains', which is why the mechanism and its size were published here ahead of the corrected release rather than alongside it.

It is wider than the two Google lanes. The two Google answer surfaces are the visible cases. Gemini is a third: it answered intermittently rather than going dark continuously, which is why sweeps looking for multi-week outages walked past it. ChatGPT and Perplexity have no such days at all, and Claude has one, which the rule below deliberately declines to exclude.

The rule, and the floor it carries. A blind engine-day is an engine-day with zero eligible citation events corpus-wide and at least five successful eligible runs for that engine that day. Blind engine-days leave the denominator; the numerator is not touched.

The five-run floor is the part worth stating, because the rule without it is wrong at the tail. A single run returning no citation is an ordinary result, not evidence that an engine was not answering. Claude — an engine with no collection problem in this window — has exactly one day that the unfloored rule would call blind, carrying exactly one run. The floor removes that false positive while keeping the outages: on the window ending 2026-09-18, the unfloored rule excluded 88 engine-days and 866 runs and the five-run rule excludes 70 engine-days and 829 runs, keeping 95.7% of the exposure. A higher floor is not more conservative; at eight runs the rule re-admits 83 runs from the stretch where Google AI Overviews logged 238 completed runs across 33 days and zero citations, which is a confirmed outage being counted as real data.

Size, measured on the window 2026-05-10 to 2026-10-06. 948 of 21,813 eligible runs are blind-engine-day runs, 4.35% of the denominator, across 81 engine-days: Gemini 22 days and 495 runs, Google AI Mode 25 days and 191 runs, Google AI Overviews 34 days and 262 runs, and none for ChatGPT, Claude or Perplexity. The last blind engine-day is 2026-09-24; no engine-day since qualifies, so the 948 runs are the same population first measured on the window to 2026-09-25, where they were 5.6% of 16,788. Eligible citation events are 168,406 before and after, which is the numerator being untouched, and days observed are 143 on both sides.

What moves, and by how much. Of 106 published category and question-shape strata, 74 carry blind runs and move; 32 do not move at all. Median denominator relief among those 74 is 6.64% and the maximum is 18.75%; median published-rate uplift is 7.11% and the maximum is 23.08%, on enterprise software top-list questions, whose denominator falls from 144 runs to 117. Every overall per-domain rate rises by the same factor, 4.54% on this window. Cited-run counts, days-cited counts, observed-engine lists and the numerator behind every rate are unchanged.

No rank moves. The relief is uneven across strata, but every rank the Index publishes compares domains inside one shared denominator: the overall rank, the source-role rank, the evidence-qualified rank, each category's top domains and each stratum's ranking. Rescaling a shared denominator raises every rate in that comparison by the same factor, so no order changes. Measured on this window, no ordinal rank changes anywhere, the top 25 is the same set in the same order, and no confidence grade moves. The one field that shifts is the peer-context position of 243 graded domains, none in the top 25, where rows tied at four-decimal rounding regroup after the rescale. An earlier version of this entry said some ranks would move; the measurement shows they do not.

Zero strata stop publishing. A correction that thins a denominator can push a stratum below the Index's evidence floor of 10 observed runs across 7 distinct dates, and a rate quietly disappearing must not look like a rate quietly improving. On this window, no published stratum falls below the floor after the correction, and no stratum's set of observed dates changes. The thinnest published stratum after the correction holds 86 runs across 7 dates. The count depends on the window; a stratum that falls below the floor is published as collecting, never as a rate.

Backfill only. This rule is a heuristic standing in for a fact the collector should record directly. It applies to history that predates a per-engine-day collection-health signal, and stops being consulted for new days once that signal exists. The collector has graded each lane's output nightly since 2026-09-25, but it keeps that grade for the latest night only, so the grade cannot yet stand in for the rule on past days. Each corrected release publishes a denominator_exclusion block naming the rule, the floor, the per-engine totals and every excluded engine-day, and the release cannot publish unless that block's totals reconcile with its denominator, because a rate whose denominator excludes observations cannot be checked without them.

Status. The correction takes effect in the first release identified mri_score_v2.1, which bumps the methodology version because it moves historical scores. Every mri_score_v2.0 release, including mri_score_v2.0+2026-10-07+95c592f6663c, carries the uncorrected denominator, so every affected published rate in it is a floor. The figures above are measured on the window stated; each corrected release carries its own totals in its denominator_exclusion block. The correction missed its original target, the 2026-09-30 release, and the update log records it.

The two Google answer surfaces returned no observable answer for a stretch of this window #

Google AI Overviews returned its last observable overview block on 2026-08-14, and Google AI Mode its last observable answer on 2026-09-13. After those dates the monitored runs for each lane continued to be issued and recorded as completed while the collection provider returned either no answer surface or a fault. The collection provider for both lanes has since been replaced, and both lanes have returned answers again since late September. Release mri_score_v2.0+2026-10-07+95c592f6663c records five of the six engines as currently returning answers and Gemini as quiet, with 12 of 125 collected calls returning an answer, 10%. Coverage across the window is still uneven: of 144 observed days, Gemini answered on 122, Google AI Mode on 102 and Google AI Overviews on 85.

Size: 513 runs of 16,475, and no citation among them. Measured on release mri_score_v2.0+2026-09-23+fdec6388001a, 513 monitored runs, 3.1% of that release's denominator, returned no observable answer surface. None of the 513 carries a citation, so no citation in the Index is affected, and no domain's cited-run count, days-cited count or observed-engine list changes. What the 513 affect is the denominator every citation rate is divided by. That denominator is uniform across every domain in the release, so on this issue alone no rank and no confidence grade moves and every published rate is understated by the same 3.1% relative. On that release forbes.com published at 675 cited runs of 16,475, a 4.10% citation rate; the same 675 cited runs on a denominator of 15,962 are 4.23%. In release mri_score_v2.0+2026-10-07+95c592f6663c forbes.com publishes at 931 cited runs of 22,344, a 4.17% rate, and that figure is a floor for the reason given in the entry above: with the 948 blind-engine-day runs out of the denominator, the same 931 cited runs of 21,396 are 4.35%.

How this entry relates to the one above. The two counts are produced by different rules over different windows and are not a difference to be subtracted. The 513 selects runs where the provider returned no answer surface on two lanes over the window to 2026-09-23. The 948 selects blind engine-days on three engines over the window to 2026-09-25, by the engine-day's citation output rather than by a provider fault. The two populations mostly overlap, and the blind-engine-day rule removes most of this one, not all of it. Re-run over the same window, this entry's selector picks 515 runs, two more than the 513 first counted. Of those, 445 fall on blind engine-days and leave the denominator with the correction. The other 70 stay in it: 35 are provider-fault runs on days when the same lane cited sources on other queries, and 35 fall on zero-citation days with fewer than five runs, below the rule's floor. The 948 is the figure to read for the corrected release.

How long it went unrecorded. Neither lane raised an alarm when it stopped returning answers, because the Index had no check that reads a lane's own output volume rather than its run count. A lane returning zero is indistinguishable from a lane nobody is watching. That check now runs per lane: the collector has graded each lane's own output every night since 2026-09-25, and from release mri_score_v2.0+2026-10-07+95c592f6663c the release's coverage statement reads that grade, which is why Gemini is published as quiet rather than healthy. The same absence is what let blind engine-days sit in the denominator: the roster's per-engine day counts are computed from what each engine cited on, while the score denominator counts what each engine ran, and the two populations were never reconciled. It is recorded here because the standard below applies to the instrument as well as to the figures it produces.

Per-domain figures and class figures #

The Index also assigns every cited domain a source_role: editorial publication, wire and press-release distribution, community platform, vendor-owned, and five others. Class share and class rank are computed across those roles.

The two kinds of figure behave differently. Per-domain citation rate, cited runs, days cited, observed engines and confidence are observations: a domain either appeared as a source in a run or it did not, and no classification decision touches them. Class share and class rank depend on how domains are classified, so they are the figures a methodology update to classification can move. That is why the Index treats the per-domain rate as its primary measure and publishes class figures with a note on how to read them.

How the editorial class is built today. In the current release, editorial classification combines source-type evidence with AuthorityTech's placement catalog. 96 of the 1,231 public editorial-class members, 7.8% of the class denominator, are admitted through the catalog; the rest are admitted on source-type evidence. The class currently includes six platform domains alongside edited newsrooms: medium.com, amazon.com, prnewswire.com, apple.com, wordpress.com and blogspot.com. Together they hold 1,358 of the class's 13,525 cited runs, 10.04% of the class. Hosted open-publishing products such as substack.com, beehiiv.com and dev.to are filed as community platforms.

Where the Index is going: catalog-free classification #

The Index is separating three questions that answer different things, so that none of them reads another:

Classification answers what kind of source a domain is, from evidence about the domain alone. The placement catalog is removed from classification.
Orderability answers whether AuthorityTech can place content on a domain, from the catalog alone, and is published as its own field on each domain profile.
AuthorityTech's publications page remains the commercial surface listing what is available to place.

With those separated, every editorial class rank is produced by a classifier that has never seen an order book, and a reader can see orderability as a disclosed field rather than infer it. It also makes the editorial class larger and more complete: widely cited publications such as healthline.com, cited in 298 of 16,307 monitored answer runs, along with bankrate.com, webmd.com, tomsguide.com, kiplinger.com, everydayhealth.com and investopedia.com, enter the class on source-type evidence.

The sequencing is deliberate. Source-identity coverage is completed first, the catalog is removed from classification second, and the orderability field is published third. Removing the catalog before identity coverage is complete would drop genuine publications such as forbes.com, techcrunch.com, time.com, arstechnica.com, inc.com and fastcompany.com out of the class and make it narrower rather than more accurate. Until that release ships, read the per-domain rate first and the class rank second.

The update standard the Index holds #

Machine Relations is the discipline of earning AI-engine citations, and an index that informs that work has to be auditable after publication as well as at it. The Index applies five rules to every published figure.

Every published figure carries its release identifier and observation window. A number without a traceable artifact behind it cannot be checked later, because nobody can tell which run produced it.
An update runs to every page carrying the affected figure, not only to the page where the change was made. The unit of an update is the claim, not the document. A claim has more surface forms than the page that first carried it, so updates are found by the claim's shape, not only by its wording.
The update is logged with its date and its size. A reader can see what moved, when, and by how much.
What an update does not affect is stated as explicitly as what it does. Per-domain observations that a classification change does not touch are named as unaffected, so sound measurement is not discarded along with the figure that moved.
The update log is public and dated. This page is it, and it is appended to rather than replaced.

What measurement institutions do #

Established measurement institutions have long settled what a publisher owes a number after it is published, and the Index's standard is drawn from them.

Media measurement. The Media Rating Council's Minimum Standards for Media Rating Research, the assessment basis for accrediting audience measurement services, states it as Disclosure Standard B.2 in the December 2011 revision: "Each rating report shall point out changes in, or deviations from, the standard operating procedures of the rating service which may exert a significant effect on the reported results. This notification shall indicate the estimated magnitude of the effect. The notice shall go to subscribers in advance as well as being prominently displayed in the report itself."

Search measurement. Google maintains a public data anomalies log for Search Console, recording known issues that affect reported data. Its entries include a logging error that prevented Search Console from accurately reporting impressions from May 13, 2025 until April 27, 2026, disclosed in a dated public list.

Scientific publishing. The ICMJE recommendations carry a section on corrections and version control. Nature Portfolio's correction and retraction policy defines distinct amendment types and links each one to the original article. Crossmark lets a reader holding a document check whether it is the current version, Retraction Watch maintains a public database, and arXiv keeps every prior version addressable.

Official statistics. The Bureau of Economic Analysis publishes each quarter's GDP in successive vintages, an advance estimate, then a second estimate, then later revisions, and maintains revision information and a vintage history alongside the current figure. The JCGM guides published through the BIPM treat a measurement result as incomplete without a stated uncertainty.

Digital advertising. IAB Tech Lab maintains versioned measurement specifications so that a change in what is counted is visible as a change, rather than appearing as a change in the thing being counted.

The common rule is: publish what changed, say how big the change was, and keep the record addressable. AI citation measurement is roughly two years old and inherits none of this by default. The Index adopts it as its own standard.

Update log #

Date	Release	Methodology update	Effect
2026-09-20	`mri_score_v2.0+2026-09-20+47973f373a20`	Publication figures moved to the answer-run rate on one denominator, with class figures read separately from per-domain figures. Wire and press-release class share is computed with prnewswire.com counted in the wire class. A reading note is placed beside every editorial-class figure.	Wire and press-release class share published at 0.37%, 414 of 112,518 classified cited-domain observations, the smallest classified citation pool in the release. Per-domain rates, cited runs, days cited, observed engines and confidence unchanged.
2026-09-21	`mri_score_v2.0+2026-09-21+d73e239e7c5b`	Cross-engine overlap figures conditioned on evidence depth. A domain cited once is single-engine by arithmetic, so single-engine share is reported by citation band.	Single-engine share: at most 67.4% across the whole index, at most 1.5% among domains cited 10+ times, 0.0% among the 508 graded domains, whose pairwise engine overlap runs 57.3% to 90.6%. The three shares are upper bounds rather than measurements, because this release sits inside the two Google collection outages recorded above and missing days push domains toward fewer engines. Per-engine exclusive pools computed at 10+ citations: Perplexity 18. Google AI Mode reads 0 on the same column, which is also what a collection outage produces, so it is not reported here as an engine property. Per-query overlap measurements are unaffected. Restated 2026-09-25 to match the study that carries these figures.
2026-09-22	`mri_score_v2.0+2026-09-21+d73e239e7c5b`	Evidence conditioning extended to exclusivity counts and to engine-count gradients. A domain's engine count cannot exceed its citation count, so a gradient of citation rate against engine count is reported as a description of the index rather than as a causal effect.	Head overlap, which is frequency-ranked, is the lead measure for per-engine comparisons; exclusivity counts are reported as tail composition.
2026-09-22	`mri_score_v2.0+2026-09-22+4aa6d302f7ed`	Editorial classification disclosed: the current release combines source-type evidence with AuthorityTech's placement catalog. Catalog-free classification and a published orderability field set as the Index's direction.	96 of the 1,231 public editorial-class members, 7.8% of the class denominator, are admitted through the catalog. Per-domain rates, cited runs, days cited, observed engines and confidence unchanged.
2026-09-23	current	Publication pages across the estate aligned to the current methodology, each linking to this log. Class-figure notes carry the heading "How to read the editorial-class rank."	No figure changed.
2026-09-24	`mri_score_v2.0+2026-09-23+fdec6388001a`	Known data issues log opened. Google AI Overviews recorded as returning no observable overview block after 2026-08-14 and Google AI Mode no observable answer after 2026-09-13; collection provider for both lanes being replaced, and an output-volume check added per lane.	513 of the 16,475 monitored runs in the current release carry no observable answer surface and no citation, 3.1% of the denominator. No citation, cited-run count, days-cited count, observed-engine list, rank or confidence grade is affected. Every published rate is understated by the same 3.1% relative until the 513 runs leave the denominator in the next release.
2026-09-27	current release uncorrected; correction targeted at 2026-09-30	Blind engine-days recorded as staying in the citation-rate denominator: an engine-day with zero eligible citations corpus-wide and five or more completed runs contributed runs that could not reach a numerator, so affected rates are floors. Exclusion rule built with a five-run floor, numerator untouched, backfill only, and a `denominator_exclusion` block published with each corrected release.	On the window 2026-05-10 to 2026-09-25, 948 of 16,788 eligible runs are blind-engine-day runs, 5.6% of the denominator, across 81 engine-days on Gemini, Google AI Mode and Google AI Overviews. Eligible citation events unchanged at 131,670. 63 of 95 published strata move: median published-rate uplift 6.84%, maximum 31.03%. Zero strata fall below the evidence floor. Correcting this raises the Index's own figures. The current release still carries the uncorrected denominator; release figures and restated per-domain rates publish here with the corrected release.
2026-10-07	`mri_score_v2.0+2026-10-07+95c592f6663c`	The coverage statement reads the collector's nightly per-lane grade, so an engine that is reachable but returning few answers is published as quiet rather than healthy.	Gemini published as quiet: 12 of 125 collected calls returned an answer, 10%. Only the coverage statement and roster status change; no citation, rate, rank or confidence grade moves.
2026-10-07	corrected from the first `mri_score_v2.1` release	Blind-engine-day correction rebuilt on the current scoring code and set to ship as methodology `mri_score_v2.1`; it missed the 2026-09-30 target in the row above. Two statements on this page restated from measurement: the correction moves no rank, and it removes 445 of the 2026-09-24 entry's runs from the denominator, not all of them.	On the window 2026-05-10 to 2026-10-06, 948 of 21,813 eligible runs, 4.35% of the denominator, across the same 81 engine-days. 74 of 106 published strata move: median published-rate uplift 7.11%, maximum 23.08%. Every overall per-domain rate rises 4.54%. Zero strata fall below the evidence floor. No ordinal rank, top-25 position or confidence grade moves. 70 of the 2026-09-24 entry's runs stay in the denominator. Each corrected release publishes its own totals in its `denominator_exclusion` block.

Sources and method #

Index figures were read live from the public per-domain profiles of the releases named in the log, methodology mri_score_v2.0, window May 10 to September 20, 2026 for the September 20 release, 16,039 monitored answer runs across ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity. Every domain in the index has a public profile under machinerelations.ai/index/domains.

The blind-engine-day figures were produced by building the Index's scoring pipeline twice over the same window, 2026-05-10 to 2026-09-25, once on the unchanged pipeline and once with the exclusion rule applied, both as builds that write nothing, and diffing the two results: eligible runs, eligible citation events, days observed, published strata and each stratum's denominator and rate. The 948 runs and 81 engine-days reconcile to the per-engine day and run counts stated. The five-run sensitivity figures, 88 engine-days and 866 runs unfloored against 70 and 829 at the five-run floor, were computed on the window ending 2026-09-18 and are stated on that window. The live release identifier, window, per-engine day counts and coverage statement were read from the Index's own public release manifest on 2026-09-27, and the forbes.com figures from its live public profile in each release named.

The figures for the window ending 2026-10-06 come from the same two-build comparison, run on the scoring code that ships the correction. On the uncorrected side that build reproduces every published figure of release mri_score_v2.0+2026-10-06+2251ea11a4db. Ranks were compared on every published rank field: overall, source role, evidence-qualified, each category's top domains, each stratum's ranking and the peer-context position. The 2026-10-07 release identifier, window, coverage statement and per-engine day counts were read from the public release manifest on 2026-10-07. The forbes.com figure was read from the same release's public Index file; its corrected rate is 931 cited runs over the release denominator less the 948 blind-engine-day runs, every one of which falls inside that release's window.

The known-data-issue figures were read from the Index's own collection record using the current release's filters — successful, public-scoring-eligible runs over the window May 10 to September 23, 2026 — which reproduce the published run count of 16,475 exactly. The 5,054 and roughly 22,000 citation figures and the 730 and 1,288 answer-page counts come from that same read.

The 96-of-1,231 figure was produced by re-running the release's classification function over the editorial-class members with the placement catalog withheld and counting the members that leave the class. The healthline.com figure was read from its live public profile in release mri_score_v2.0+2026-09-22+4aa6d302f7ed.

External standards and policies were read at each publisher's own page on September 20, 2026, and each is linked above at the URL read: the Media Rating Council's Minimum Standards for Media Rating Research (December 2011 revision, quoted verbatim from the PDF published at mediaratingcouncil.org), Google's Search Console data anomalies log, the ICMJE recommendations, Nature Portfolio's correction and retraction policy, Crossref's Crossmark, the Retraction Watch database guide, arXiv's submission and versioning help, the Bureau of Economic Analysis GDP page carrying revision information and vintage history, the JCGM publications hosted by the BIPM, and IAB Tech Lab's standards index. No third-party roundup was used as evidence for what any of these organisations publishes.

Related research: how concentrated are AI citations.

How the Machine Relations Index Measures AI Citations: Methodology and Update Log