# Six AI Engines Observed 74 to 133 Days in One Release

All six answer engines in the September 26, 2026 Index release are healthy, and their observed days in the same window range from 133 down to 74.

Canonical URL: https://machinerelations.ai/research/per-engine-observation-coverage-ai-citation-index-2026
Published: 2026-09-27
Research type: Index Analysis
Tags: ai-search, citations, measurement, methodology, answer-engines

## Source Body

The Machine Relations Index release dated September 26, 2026 reports all six answer engines healthy and the release not degraded. Over the same window, the six engines were not observed on the same number of days. Perplexity was observed on 133 of 133 days. Google AI Overviews was observed on 74.

That is the whole finding, and it has a consequence for every cross-engine number anyone computes from a six-engine dataset. Current engine health and historical engine coverage are two different facts. A release can be entirely healthy today and still carry a window in which one engine contributed 1.80 times as many observation days as another.

## The roster, in full

The per-engine day counts are published in the release manifest at [mri-release-manifest.json](https://machinerelations.ai/data/mri-release-manifest.json), under `engine_roster.days_observed`. Release `mri_score_v2.0+2026-09-26+2b779408cfda`, window May 10 through September 26, 2026, 133 observed days, 16,925 eligible answer runs, 132,514 citation events, 23,978 distinct cited domains.

| Engine | Days observed | Share of the 133-day window | Status in this release |
|---|---:|---:|---|
| Perplexity | 133 | 100.0% | healthy |
| Claude | 132 | 99.2% | healthy |
| ChatGPT | 129 | 97.0% | healthy |
| Gemini | 111 | 83.5% | healthy |
| Google AI Mode | 91 | 68.4% | healthy |
| Google AI Overviews | 74 | 55.6% | healthy |

Six engines across 133 days is 798 possible engine-days. The release observed 670 of them, or 84.0%. Of the 128 engine-days not observed, 101 belong to the two Google answer surfaces. One quarter of the roster accounts for 78.9% of the shortfall.

## Why the two Google surfaces are the ones that go dark

The mechanism is documented by Google and it is not a failure. AI Overviews appear, in Google's own description, when its systems determine that generative AI can be especially helpful for a given query, and the feature is available in a named list of countries and languages rather than everywhere ([Google Search Help](https://support.google.com/websearch/answer/13572151)). A surface that appears conditionally on the query cannot be observed on a day when no query in the basket triggers it. Google has described the same conditional behaviour since the feature's general rollout ([The Keyword](https://blog.google/products/search/generative-ai-google-search-may-2024/)), and its guidance to site owners treats AI features as an appearance that may or may not be present for a given search ([Google Search Central](https://developers.google.com/search/docs/appearance/ai-features)).

The other four engines are API surfaces with an explicit retrieval step. Perplexity, OpenAI and Anthropic each expose web search as a documented capability whose result set can be empty for a given call ([Perplexity](https://docs.perplexity.ai/home), [OpenAI](https://platform.openai.com/docs/guides/tools-web-search), [Anthropic](https://docs.claude.com/en/docs/agents-and-tools/tool-use/web-search-tool)). Retrieve-then-generate architectures make that step separable by design, which is why an answer can exist with no cited source and a day can exist with no answer at all ([Lewis et al., 2020](https://arxiv.org/abs/2005.11401)).

So the coverage column is not a quality ranking of the engines. It is a property of the measurement, and it belongs next to any number built on top of it.

## The consequence, measured rather than asserted

Divide 670 observed engine-days by engine and the implied weights fall out. In any cross-engine aggregate that treats every observed engine-day as one unit, this release weights Perplexity at 19.85%, Claude at 19.70%, ChatGPT at 19.25%, Gemini at 16.57%, Google AI Mode at 13.58% and Google AI Overviews at 11.04%.

That is not a rounding difference. It is visible in the public dataset. The release records, for each of the 23,978 cited domains, which engines cited it. Count how many domains each engine reached, then divide by the days that engine was observed:

| Engine | Domains it cited | Days observed | Domains per observed day |
|---|---:|---:|---:|
| Perplexity | 10,751 | 133 | 80.8 |
| Google AI Mode | 7,464 | 91 | 82.0 |
| Gemini | 7,278 | 111 | 65.6 |
| Claude | 5,585 | 132 | 42.3 |
| ChatGPT | 5,399 | 129 | 41.9 |
| Google AI Overviews | 2,337 | 74 | 31.6 |

Read the raw column and Perplexity reaches 44% more domains than Google AI Mode. Normalise by observed days and the two are indistinguishable, with Google AI Mode marginally ahead. The ranking of the two broadest engines in this release is decided by coverage, not by behaviour.

Google AI Overviews is the narrowest engine on both readings, which is the useful negative result: coverage explains the top of this table and does not explain the bottom. An engine can be both under-observed and genuinely narrow, and only the normalised column separates the two cases.

## What this does not distort

Coverage asymmetry of this size would matter most where a domain's record depends on a single engine, because an engine observed on 74 days cannot place a domain on a 75th. In this release that ceiling binds nothing. Of the 23,978 cited domains, 16,313 — 68.0% — appear in exactly one engine's answers, and every one of them is thin: the most-cited single-engine domain appears on 14 days, and the most-cited domain reached only by Google AI Overviews appears on two. No domain in this release is cited on more than 74 days by two or fewer engines.

The graded set moves in the opposite direction. Of the 521 domains carrying a confidence grade of A, B or C, 240 appear in all six engines, six appear in as few as two, and none appears in only one. The release's evidence floor of 10 observed runs across seven distinct run dates has, in practice, selected for domains broad enough that no single engine's coverage decides their record.

So the honest statement is narrower than the finding first suggests. Per-engine coverage does not currently distort a published citation rate or a confidence grade in this release. It distorts engine-breadth counts, and it distorts any cross-engine comparison a reader assembles themselves from the domain records — which is the comparison most AI visibility reporting is built from. The same asymmetry has already been named as a confound on a cross-engine agreement figure rather than as a measured object ([Para Labs](https://paralabs.ai/blog/why-ai-visibility-scores-differ-between-tools)), and the behavioural differences between the two Google surfaces have been measured separately ([Paralax](https://paralax.ai/blog/google-ai-mode-ai-overviews-source-divergence-2026)).

## The reporting rule this implies

A cross-engine citation figure is a weighted average whose weights are per-engine observed days. Report them, or the reader cannot tell a behavioural difference from a coverage difference.

This is the established practice in measurement fields that have had longer to make the mistake. Survey standards require the disposition of the sample to be published alongside any rate computed from it, not summarised as a single number ([AAPOR Standard Definitions](https://aapor.org/standards-and-ethics/standard-definitions/)). Official statistics treat coverage and sample size as reportable quality dimensions rather than production detail ([Eurostat](https://ec.europa.eu/eurostat/web/quality/european-quality-standards), [UN Statistics Division](https://unstats.un.org/unsd/methodology/dataquality/), [Office for National Statistics](https://www.ons.gov.uk/methodology/methodologytopicsandstatisticalconcepts/qualityinofficialstatistics), [IMF Data Quality Assessment Framework](https://www.imf.org/external/np/sta/dsbb/2003/eng/dqaf.htm)). The American Community Survey publishes sample size next to data quality for exactly this reason ([U.S. Census Bureau](https://www.census.gov/programs-surveys/acs/methodology/sample-size-and-data-quality.html)), and the precision of any estimate scales with the observations behind it ([NIST/SEMATECH e-Handbook](https://www.itl.nist.gov/div898/handbook/prc/section1/prc14.htm)).

Three checks follow for anyone reading a cross-engine AI visibility number:

1. **Ask for the per-engine observed days, not the engine list.** A dataset that names six engines has told you its roster, not its weights. Where the dataset publishes a health status, check whether that status describes today or the window.
2. **Normalise before comparing engines.** Domains reached, citations recorded and breadth counts are all coverage-scaled quantities. A raw count difference of 44% survived normalisation as a difference of roughly nothing in this release.
3. **Treat a missing engine-day as missing, not as zero.** An unobserved day contributes no evidence in either direction. The distinction between unavailable and zero is the more expensive version of the same error, and it has its own record in the [Index research](https://machinerelations.ai/research/collecting-is-not-zero-citation-rate).

## Methods

Every figure here is read from two public artifacts of one release: the release manifest at `/data/mri-release-manifest.json` for `engine_roster.days_observed` and `coverage`, and the public dataset at `/data/machine-relations-index.json` for the per-domain `engine_breadth` records. The roster is published in the manifest rather than in the dataset file, so a reader who opens only the dataset will not find the day counts; the [Index page](https://machinerelations.ai/index) names the field and the manifest is where it lives.

Domains-reached counts are the number of distinct domains listing each engine in `mri_score_v2.overall.engine_breadth.engines`. Domains per observed day divides that count by the engine's `days_observed`. It is a ratio of two published quantities, not a modelled rate, and it carries no confidence interval: a domain can be cited many times on one day and once on many, and the breadth record does not distinguish them. Single-engine ceilings are computed from `days_cited` against each engine's window. The 84.0% engine-day figure is 670 observed engine-days over 798 possible, where possible is six engines across the release's 133 observed days.

No engine in this release is degraded, and nothing here is a claim that one is. The Index observes production surfaces whose availability is set by their providers, and the correct response to uneven availability is to publish it, which is what the manifest does. Where a stratum has not accumulated enough evidence to be interpretable, the [Machine Relations Index](https://machinerelations.ai/index) marks it collecting and prints no rate, and the same discipline applies one level up: an engine observed on 74 of 133 days is reported at 74 days, not averaged into a six-engine number that implies 133.

## 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

- [Machine Relations Index (MRI)](https://machinerelations.ai/glossary/machine-relations-index)
- [MRI Score](https://machinerelations.ai/glossary/mri-score)
- [Machine Relations (MR)](https://machinerelations.ai/glossary/machine-relations)
- [Citation Rate](https://machinerelations.ai/glossary/citation-rate)

### Supporting research

- [Only 508 of 22,377 AI-Cited Domains Are Cited Often Enough to Score](https://machinerelations.ai/research/ai-citation-evidence-floor-scoreable-domains-2026)
- [Do AI Citation Leaderboards Change Between Daily Releases? What the Machine Relations Index Re-Measures](https://machinerelations.ai/research/ai-citation-leaderboard-movement-between-releases-2026)
- [How Many Sources Does a Typical AI Answer Actually Cite? Published Counts Range From 2 to 22](https://machinerelations.ai/research/ai-citation-density-per-answer-denominator-2026)
- [How the Machine Relations Index Measures AI Citations: Methodology and Update Log](https://machinerelations.ai/research/ai-citation-index-correction-record-2026)

### Framework context

- [Machine Relations Index](https://machinerelations.ai/index)
- [Machine Relations Stack](https://machinerelations.ai/stack)
- [Evidence Base](https://machinerelations.ai/evidence)