44% of the AI Citation Map Publishes No Rate Today — And That Is Not Zero

Today's Machine Relations Index release publishes citation rates for 85 of its 151 measurable segments. The other 66 publish no rate at all. Read the wrong way, that looks like proof AI engines cite nobody in those categories. It is the opposite: in one of them, a single domain appears in 33 of 101 observed answers and still shows no published rate.

That gap between unavailable and zero is the most expensive misreading available in AI visibility work right now, and as of today it is a misreading anyone can make in public, because the per-category pages went live this morning.

What moved in the last 24 hours

Release mri_score_v2.0+2026-09-18+8fa38e54dd0a, generated 2026-09-18, observation window 2026-05-10 through 2026-09-18, 125 days observed.

Measure	2026-09-17 release	2026-09-18 release	Change
Cited domains	22,026	22,179	+153
Source events	123,538	124,397	+859
Answer runs	15,678	15,782	+104
Segments publishing a rate	—	85 of 151 measurable	56.3%

Six engines reported healthy runs inside the window: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity.

The release floor is dates, not volume

A segment is one category paired with one question shape — "healthcare services / best tools," "AI infrastructure / best tools." Each keeps its own denominator. A segment publishes a leaderboard only after it clears two thresholds: 10 observed answer runs and 7 distinct observation dates.

The second threshold is the one that surprises people. It is a temporal test, not a volume test. It exists so a rate cannot be minted from a single busy day of sampling, because answer engines are not stable hour to hour and a one-day burst is a snapshot, not a rate.

The consequence is counterintuitive, and today the index shows it plainly.

Three reasons a rate is missing, and none of them is zero

The index headline reads "85 of 157." That 157 is the full stratum universe, and it is worth taking apart, because a missing rate means three different things and only one of them is about how much evidence has accumulated.

State	Segments	What it means
Publishing a rate	85	Cleared 10 runs across 7 distinct dates
Collecting, with observations	17	Real runs recorded, still under the floor
No observations yet	49	The prompt basket has not reached this pairing
Not measurable at all	6	Structural placeholder, never collectable
Total universe	157	151 measurable + 6 placeholders

The six at the bottom are the legacy news-topic bucket crossed with the six buyer question shapes. That bucket only ever receives news-shaped prompts, so those pairings cannot accumulate evidence by construction. The index keeps the rows so the arithmetic is honest rather than quietly dropping them, and flags them not_collectable in the public dataset. They are not a gap in the measurement; they are a gap that cannot exist.

That leaves 151 measurable segments, and the 66 without a rate split cleanly. Seventeen are genuinely mid-collection, concentrated in three categories: healthcare services, HR and talent, and iGaming and betting. Forty-nine have zero observed runs — nine categories, including Industrial, Local Services, Logistics & Freight, Professional Services, Sales, GTM & CRM, VC & Private Equity and Physical Consumer Products, where the prompt basket has simply not run buyer questions yet.

Martech & Advertising is the instructive one. Its news-driven segment publishes on 602 observed runs across 53 dates. All six of its buyer-question segments have zero runs. Same category, same release: settled on one question shape, unsampled on six others. A category read that averages across those is not a read at all.

A collecting segment with more evidence than a published one

	Healthcare services / best tools	AI infrastructure / best tools
State	Collecting	Published
Observed answer runs	101	77
Distinct run dates	6	7
Observed domains	199	204
Top source by cited runs	Omnimd.com — 33 runs	Medium — 36 runs
Published citation rate	None	46.75%

Healthcare services / best tools has more observed answer runs than AI infrastructure / best tools. It has been collecting since 2026-09-03. It is one observation date short of the floor, so it publishes no rate, no rank and no winner.

Meanwhile the domain leading it — Omnimd.com, classified as an editorial publication — sits in 33 of 101 observed answers. In raw cited-run terms that is more appearances than YouTube earns in the published AI infrastructure board, where 22 cited runs is good enough for a 28.57% rate and second place behind Medium. Beckers Hospital Review is in 12, Forbes in 7, CureMD in 19 and PracticeSuite in 17.

Engines are citing in that category, and the index can already name 199 distinct domains they cited across 101 observed runs. What is unfinished is the rate, not the citing. That is a statement about this segment's observation set, and it is not evidence about any particular brand's standing — a brand absent from those 199 is unmeasured here, not ranked last.

What the published boards look like when they do mature

For contrast, the matured AI infrastructure board is dense with sources a traditional media plan would never buy: BentoML at 22.08%, Firecrawl at 19.48%, arXiv and Deploybase at 18.18%, Spheron at 16.88%, ZenML at 16.88%, Hugging Face at 15.58%, Redis at 12.99%, LinkedIn at 12.99%, GitHub and Reddit in the top twenty.

That is what a finished measurement looks like: one editorial publication on top, then a long run of practitioner documentation, repositories and vendor engineering writing. Across the whole index the same pattern holds — Reddit leads all 22,179 domains at 12.75%, and Medium is the highest-ranked of 1,222 classified editorial publications at 5.11%, ahead of Forbes at 4.11%, with TechRadar third among editorial publications on 448 cited runs across four engines.

The operating rule

Four things follow, and they are worth holding before anyone briefs a client or a board on a category read.

A missing rate is a sample-size statement about us, not a behaviour statement about engines. The index says so on the page: missing mature rates and ranks are unavailable, not zero. Treat a segment without a rate as an open question, never as a cleared field.

Ask which kind of missing it is. Seventeen segments are mid-collection with real observations you can already reason about. Forty-nine have no observations at all, which tells you nothing about the category and everything about the sampling frame. Six can never be measured. Those are three different answers to the same blank cell.

Cited-run counts in a collecting segment are still real observations. They are not a rate, and they should never be quoted as one. But they tell you which domains keep turning up, which is enough to shape a first hypothesis about where the citations in that category actually live.

Check the date count before the rate. Two segments with identical run counts can be a published leaderboard and a blank page depending on how those runs spread across days. The date count, printed on every segment page, is the field that decides it.

Why this sits inside Machine Relations

Machine Relations treats an answer engine as a source-selection system that can be observed, not a black box to be guessed at. The discipline's whole claim is that the selection is measurable and therefore earnable.

That claim only survives if the measurement refuses to fill its own gaps, and if it is precise about which gap it is looking at. A scoreboard that published a rate off six days of data would be more satisfying and less true, and every downstream decision made on it would inherit the error. One that let "no observations yet" and "still collecting" share a label would hide the difference between a category we have not asked about and a category we are halfway through measuring. Publishing no rate on 66 of 151 measurable segments, and saying which 49 have never been sampled, is the index being honest about what it does not yet know — and an instrument that declines to guess is the one worth citing when it does speak.

FAQ

Does a segment with no rate mean AI engines ignore that category? No. It means one of three things: the observation set has not yet spread across enough distinct dates, no runs have been recorded for that category and question shape at all, or the pairing is structurally not collectable. Healthcare services / best tools has 101 observed answer runs and 199 observed domains while still publishing no rate.

What are the exact thresholds? Ten observed answer runs and seven distinct observation dates, per segment, in the current release.

Why does the index say 157 segments when only 151 are measurable? The universe keeps six placeholder rows — the legacy news-topic bucket paired with the six buyer question shapes — that can never accumulate evidence. They are flagged in the public dataset so the totals reconcile instead of silently shrinking.

Can cited-run counts be quoted publicly? As observation counts, yes, with the run and date denominators attached. As a citation rate or a rank, no — that is the specific thing the floor exists to prevent.

How often does this change? The index rebuilds daily. A segment one date short today can publish tomorrow, which is why the release id belongs in any figure that leaves your building.