Ranked Without a Grade: Half of the AI Citation Index's Top-Ten Positions Belong to Ungraded Domains

In the September 18, 2026 release of the Machine Relations Index, 512 domains hold a top-ten position in at least one published segment. 257 of them — almost exactly half — carry no confidence grade at all. In every place the release reports a grade for them, it reports collecting.

procuredesk.com is one of them. Its profile records a 0.13% overall citation rate across 21 of 15,782 monitored answer runs, confidence Collecting, and, three lines further down, a standing of #1 of 264 on Enterprise Software top lists at a 16.67% rate inside that segment.

Both readings are correct, and they are not in tension. They are the output of two different tests applied to two different objects, and the gap between them is where most of the web's actual AI-citation standing is currently sitting — visible in the data, invisible to the domains that hold it.

What was measured #

Every figure here comes from the public release at machinerelations.ai/data/machine-relations-index.json, contract machine_relations_index_public_view_v2.0, methodology mri_score_v2.0, artifact 8fa38e54dd0a, generated September 18, 2026. The observation window runs May 10 to September 18, 2026 — 125 days, 15,782 observed answer runs, 912 eligible queries, 124,397 source events, 22,179 cited source domains, six engines.

The six are ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity. Each publishes its own account of how it surfaces and links sources: OpenAI documents a search crawler distinct from its training crawler (OpenAI bots), Google describes AI Mode as its own search experience (Google, AI Mode) and documents its crawler fleet and AI-surface guidance separately (common crawlers, AI features and your site), Anthropic documents citations as part of Claude's web search (Anthropic, web search tool), Google documents grounding with Search for Gemini (Gemini API), and Perplexity documents its own retrieval surface (Perplexity). Each reaches the web under its own user agent, governed by the Robots Exclusion Protocol as standardised in RFC 9309.

A segment is one subject category paired with one buyer question shape — for example Enterprise Software paired with top lists. This release carries 157 strata, of which 151 are collectable, 85 are published and 66 are still collecting.

Two thresholds, two objects #

The release states its evidence floor in its own methodology: 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. The floor is {min_observations: 10, min_run_dates: 7}, and the thing it governs is the stratum.

A domain's confidence grade is a separate judgement, made on that domain's own accumulated evidence across the whole window. In this release those two tests draw their lines in very different places:

30 is not a published rule; it is the observed boundary of this release. No graded domain in it was cited in fewer than 30 answer runs, and 25 domains sit exactly at that floor — the lowest of them on 8 distinct dates. Median grade C sits at 43 cited runs across 18 dates. Every domain below that line reads collecting, and there are 21,679 of them against 500 graded (14 A, 56 B, 430 C).

So a segment can clear its floor, publish, and rank all of its cited domains — while most of those domains individually sit far below where domain-level grading begins. The rank is a position inside a population that has been measured enough to report. The grade is a statement about how much evidence stands behind one domain's own rate. Treating the absence of the second as the absence of the first is the error, and it is the same error the Index guards against elsewhere: a collecting label is not a zero.

This is ordinary measurement practice rather than an Index quirk. Statistical agencies publish minimum-observation and reliability rules that gate whether an estimate is shown, separately from the estimate's rank or position — see the NIST/SEMATECH handbook on sample size and its engineering statistics handbook, the Census Bureau on sample size and data quality and design and methodology, AAPOR's standards, the BLS Handbook of Methods, and NIST on measurement uncertainty.

Where the ranks actually are #

For each of the 22,179 domains, take its best rank across every published segment it appears in:

Read across the top three rows: 512 domains hold a top-ten position, and 257 of them are ungraded. Widen to the top twenty-five and it is 1,363 domains, 980 of them ungraded — 71.9%. The further down the table you go, the more completely the positions belong to domains with no grade, which is what you would expect; what is not obvious is how far up that holds. Even the number-one slots are not a graded monopoly: 45 domains lead a published segment and three of them carry no grade.

The 500 graded domains behave in the mirror image. 255 of them hold a top-ten position, 42 hold a number one, and just 9 have no top-100 position at all — 8 ranked 101st or lower, and one cited only inside segments that have not yet published, so it has nowhere to be ranked. Grading tracks position strongly — it just does not exhaust it.

At the other end, 1,475 domains have no published-segment rank at all. They were cited only inside segments that have not yet cleared the floor. For those domains the Index currently reports presence and nothing else, and the honest reading is that their standing is not yet measured rather than that it is poor.

What the 257 look like #

The domains holding a top-ten position without a grade are not marginal, and they are not a rounding error against the floor:

240 of the 257 clear ten cited runs and 147 clear seven distinct dates. They sit in the corridor between a stratum's floor and where domain grading starts — enough evidence for a segment to report their rate and rank, not yet enough for the release to grade the domain itself.

They are also not single-engine flukes. Of the 257, 132 were cited by four or more of the six engines and 23 by all six; only 10 were cited by a single engine. A domain cited by six engines has cleared six retrieval systems that share very little machinery.

Their assigned source roles are lopsided in one direction: 237 of the 257 are classified as other observed source — the residual class, the sites the release could not place in a named role. Of the rest, 8 are vendor-owned, 6 editorial publications, 3 academic or government, 3 market and company databases. These are, overwhelmingly, sites with no institutional label attached to them, holding positions in segments where institutional labels are what everyone assumes wins.

Here are nine of them, each checkable on its own profile page:

The segments these positions sit in are not thin. The median one had 219 distinct domains cited in it over the window, with an interquartile range of 186 to 272. Leading a field of 264 is not the same achievement as leading a field of 5, and the Index states the denominator on every row for exactly that reason.

Which classes hold positions at all #

Holding any top-ten position in any published segment, by the source role the release assigned:

Two rows are worth pausing on. Vendor-owned sources hold more top-ten positions than editorial publications do — 54 against 52 — from a class a third smaller (823 domains against 1,222), so their per-domain rate is half again higher. A vendor's own site is not structurally disadvantaged in these segments. And the nine domains in the wire and press-release class hold zero top-ten positions in any published segment in this release, which is consistent with the same class holding 0.20% of all citation in the concentration curve for this release.

The positions are also spread evenly across question shapes rather than clustering in one kind of answer. Of the 257: problem-first research 50, how buyers choose 48, top lists 45, comparisons 42, is it worth it 39, best tools 33. Across categories the top five are Emergent Prosumer (30), Enterprise Software (29), Consumer Products (27), Education & Training (26) and Fintech (24), covering thirteen categories in all.

How to look up any domain #

Every one of the 22,179 domains in this release has a public page, and it is the same page whether the domain is graded or not.

• Profile: machinerelations.ai/index/domains/<domain> — overall citation rate, cited runs, distinct dates cited, how many of the six engines cited it, confidence grade, and a row for every segment it has been observed in, each with the segment's state, the domain's rate inside it, and its standing as a rank out of the segment's total.
• Segment leaderboard: machinerelations.ai/index/categories/<category>/<shape> — the ranked field the standing is drawn from.
• Badge: machinerelations.ai/index/domains/<domain>/badge.svg — the current release's reading as an embeddable SVG, stamped with the release id, linking back to the profile. It renders for collecting domains too, and returns 404 if the domain leaves the released population.
• Machine-readable: the full release JSON, and the Index as the entry point.

Four things are worth reading in order on a profile. Which segments show state published, because only those carry a standing. The standing itself, as a rank and a denominator. The distinct-dates figure, because that is the axis most often short when a rate looks strong. And how many of the six engines have cited it, because a position held on one engine and a position held on five are different kinds of position even at the same rank.

What this does not mean #

A rank without a grade is a measured position with less evidence behind it than a graded one, and it should be read that way. The Index publishes the grade precisely so the difference stays visible; the point here is that the absence of a grade is not the absence of a position, not that the two are equivalent.

Nothing here is causal. The release describes what six engines cited across a fixed basket of buying and research questions over 125 days. It does not say why a domain holds a position, and it does not say that any change would move one. The next release is the only test of that.

The population is cited domains, not eligible ones. A site cited zero times in the window is not in the file, so the 22,179 is not a denominator of "sites that could have been cited". And a citation is not proof that the cited page supported the claim beside it — that is measured separately in Answer-Source Fidelity, and none of its results follow from the ranks above.

Release: mri_score_v2.0, contract machine_relations_index_public_view_v2.0, artifact 8fa38e54dd0a, generated 2026-09-18. Window: 2026-05-10 to 2026-09-18, 125 days observed. Population: 22,179 cited domains, 15,782 observed answer runs, 912 eligible queries, 85 published segments of 151 collectable. Evidence floor: 10 observed runs across 7 distinct run dates, applied to a stratum. See also the concentration curve of this release and the Machine Relations Index Report for September 18, 2026.