Report the number of answers behind every engine-level reading of AI search visibility. In the October 19, 2026 Machine Relations Index release, 68.2 percent of the 31,998 cited domains appear on one engine only. That share is 100 percent for a domain cited in a single run, 21.1 percent at six to nine cited runs, 6.7 percent at 10 to 19, and 0 percent at 50 or more. A one-engine pattern is mostly what a thin sample looks like.
This page reads the Index's engine breadth field against the number of runs that cited each domain. It is a finding about measurement, not about which engine is better or which brand to prefer.
What the Index counts #
The release is mri_score_v2.1, dated 2026-10-19, covering 27,776 answer runs from May 10 to October 19, 2026 across six engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity. The unit is a root domain. Two fields matter here.
- Runs cited is the number of answer runs in the window whose cited sources included the domain, counted once per run however many times the domain appears in it.
- Engine breadth is the set of engines that cited the domain at least once in the window, and its count runs from one to six.
Engine breadth records presence, not rate. A domain cited once on Perplexity and a domain cited 500 times on Perplexity both read as one engine if no other engine cited them. The public artifact carries a per-domain engine list and a per-domain run count, and nothing else per engine.
Single-engine share by number of cited runs #
| Runs cited by the domain | Domains | On one engine only | Share on one engine | Share on three or more engines |
|---|---|---|---|---|
| 1 | 14,422 | 14,422 | 100.0% | 0.0% |
| 2 | 5,361 | 3,848 | 71.8% | 0.0% |
| 3 | 2,913 | 1,616 | 55.5% | 8.1% |
| 4 | 1,804 | 794 | 44.0% | 18.1% |
| 5 | 1,266 | 442 | 34.9% | 28.9% |
| 6 to 9 | 2,667 | 563 | 21.1% | 44.5% |
| 10 to 19 | 2,042 | 137 | 6.7% | 70.4% |
| 20 to 49 | 1,107 | 13 | 1.2% | 92.9% |
| 50 or more | 416 | 0 | 0.0% | 98.8% |
| All cited domains | 31,998 | 21,835 | 68.2% | 15.6% |
The first row is arithmetic. A domain cited in one run can appear on at most one engine, so all 14,422 of them, 45.1 percent of every cited domain, are one-engine domains by construction. Those domains account for 66.0 percent of the 21,835 one-engine domains in the release.
The rest of the column is a slope, not a step. There is no run count at which single-engine readings switch off. The share is 71.8 percent at two cited runs, 34.9 percent at five, 6.7 percent at 10 to 19, and zero among the 416 domains cited in 50 or more runs.
Why a thin sample reads as a one-engine brand #
Each extra cited run is another chance for a domain to appear on a second engine, so the single-engine share must fall as runs accumulate. The table shows how fast it falls in this release. Domains the Index has seen cited in two runs sit on one engine 71.8 percent of the time, and domains cited in six to nine runs 21.1 percent of the time.
The release assigns a confidence grade of A, B or C to 879 of the 31,998 domains. The other 31,119 remain in the collecting state. Most of the domains behind the 68.2 percent are therefore domains with very little evidence behind them, and the headline share is dominated by the long tail rather than by any consistent engine preference.
Single-prompt tests that report the share of brands cited by one engine only have the same structure. A prompt run once per engine gives each brand at most one run per engine, so a one-engine result rests on one answer each, which is the left edge of this table. Even a rerun of the same question cites a different mix: in a later Index panel of 980 questions and 22,620 answer runs (July 6 to October 20, 2026), two random halves of one question's runs shared 46.1 percent of their cited-source volume (how rewording and rerunning a prompt changes the sources AI engines cite).
What the table does not show #
- Rates. Presence on an engine is not the same as an equal citation rate on it. The public artifact has no per-engine citation rate for a domain, so this page cannot say whether engines agree on how often a frequently cited domain is cited. Within a release, only presence can be compared across engines.
- Whether the slope is sampling or preference. Part of the fall is arithmetic. How much of the remaining single-engine share reflects genuine engine-specific sourcing is not established here, because the artifact does not publish per-engine run counts for each domain.
- Which engine holds the lone citation. Among the 150 domains cited in 10 or more runs on one engine only, Perplexity holds 84 and ChatGPT 60; Claude holds 4 and Gemini 2. Neither Google surface holds any. The cause is not established. Engines differ in how many sources they cite per answer and in how many days each was observed (see per-engine observation coverage).
- Brands. The unit is a root domain. A brand with several domains, or one cited only through another publication's page, is counted under those domains.
The reporting rule this implies #
An engine-level AI share of voice or visibility figure carries one more disclosure beside the engine set, query set, counting rule, date window and weighting: the number of answer runs behind it. The table gives a practical reading. A domain labelled as visible on one engine only, when it has been cited in a handful of runs, is reported as a thin sample rather than as an engine pattern. The same label on a domain cited in 20 or more runs is a pattern worth examining, since only 1.2 percent of such domains carry it in this release.
This is a reporting rule, not a count to target. The slope above is specific to this release, its question set and its six engines, and it will move as the window grows.
For the metric definitions this page leaves aside, see AI share of voice, which counts qualifying brand mentions, and Share of Citation, which counts cited-source slots under its own denominator. The Machine Relations Index publishes the release this page reads.
Method note. Every figure can be recomputed from the public release artifact at machinerelations.ai/data/machine-relations-index.json, release mri_score_v2.1+2026-10-19+da0ee4378219; the release manifest at machinerelations.ai/data/mri-release-manifest.json carries the artifact hash. For each entry in the domains array, read mri_score_v2.overall.runs_cited and mri_score_v2.overall.engine_breadth.engine_count. Group domains by runs_cited into the bands in the table and count those with an engine_count of 1, and those of 3 or more, in each band. The confidence counts come from mri_score_v2.overall.confidence. The lone-engine split reads the single entry of engine_breadth.engines for domains with runs_cited of 10 or more and an engine_count of 1.
Frequently asked questions #
How many engines should an AI share of voice report cover? #
All the engines the reader's buyers use, each reported with its own run count. The Index observes six, and a report that covers fewer should name which engines it leaves out. The finding above concerns how to read a per-engine result, not how many engines to include.
Does a domain cited on one engine only have a visibility problem on the others? #
This page cannot say. In the 2026-10-19 release, 14,422 domains are one-engine domains because they were cited in a single run, and a further 7,413 were cited in two or more runs. The Index records where a domain was cited, not where it was absent from an answer that could have cited it.
What run count makes an engine-level reading stable? #
The Index does not set one for engine readings. Its evidence floor, which decides when a segment is published, is 10 observed runs across at least 7 run dates. For engine breadth, 6.7 percent of domains are still single-engine at 10 to 19 cited runs, 1.2 percent at 20 to 49 and 0 percent at 50 or more, which describes this release and is not a threshold.
Where can the engine-breadth figures be checked? #
In the public release artifact and manifest linked in the method note above. The fields are runs_cited and engine_breadth, both per domain.