Twenty-four companies that describe themselves as AI companies announced a funding round between September 21 and September 26, 2026. Every one of them was covered: each round in this cohort is confirmed at the company's own announcement and at one outlet that reported it.
Measured against the Machine Relations Index release dated September 26, 2026, seventeen of the twenty-four appear in no answer at all. Their domains are cited in none of the 132,514 source events recorded across 16,925 eligible answer runs since May 10, 2026. Seven appear. One of the twenty-four carries a confidence grade in the release.
Those companies raised $6.05 billion between them in six days. The seventeen with no recorded citation raised $1.25 billion of it.
The cohort, and why it is a cohort #
A cohort is only worth measuring if its membership was fixed before the measurement. This one was: the roster of twenty-four rounds was compiled and published as a weekly funding record, one line per round, with the inclusion rule stated in advance — a round counts when the company's own announcement describes the company, not a feature, as AI or as built for AI, and when amount, round and lead investor are confirmed at the company's announcement or at the outlet that reported it. Fund closes, standalone debt facilities and valuation-only news are excluded. Eight companies that announced rounds that week were excluded because their own materials do not describe them as AI companies. The roster, its rule and its ownership are disclosed on the publisher's method page; the week itself is at Press Pool.
The measurement was then taken from the other side, against a release built without reference to the roster. The Machine Relations Index observes a fixed question set across categories on six production answer engines and records which source domains those answers cite. The release used here is mri_score_v2.0+2026-09-26+2b779408cfda, window May 10 to September 26, 2026, 133 observed dates, 16,925 eligible answer runs, 132,514 source events and 23,978 distinct source domains observed.
The question the pairing answers is narrow and worth stating exactly: on the day a company announces a funding round, is its own domain already a source that answer engines reach for?
How many of the twenty-four appear at all #
Seven of twenty-four domains appear in at least one eligible answer run. Seventeen appear in none.
| Appearance in the release | Companies | Capital announced |
|---|---|---|
| At least one cited answer run | 7 | $4.79 billion |
| No cited answer run | 17 | $1.25 billion |
| Total | 24 | $6.05 billion |
One round in the cohort was announced in euros without a dollar figure and is excluded from the capital column; twenty-three of the twenty-four are dollar-denominated.
Seventeen of twenty-four is the headline, and it is a floor rather than a ceiling on the problem. The seven that do appear are not, on this evidence, visible either.
The seven that appear, in full #
| Company | Round | Announced | Cited runs | Days cited of 133 | Confidence |
|---|---|---|---|---|---|
| Firecrawl | $75M Series B | Sep 22 | 46 | 13 | C |
| Snorkel AI | $350M Series E | Sep 22 | 14 | 11 | collecting |
| Heidi | $340M Series C plus growth | Sep 22 | 11 | 7 | collecting |
| Cyera | $400M Series G extension | Sep 22 | 9 | 5 | collecting |
| Nscale | $3.36B pre-IPO convertible notes | Sep 25 | 4 | 4 | collecting |
| Verda | $189M Series B plus additional investment | Sep 22 | 2 | 1 | collecting |
| Ema | $77M Series B | Sep 23 | 1 | 1 | collecting |
Eighty-seven cited answer runs across the whole cohort. Forty-six of the eighty-seven belong to one company. Firecrawl's 46 runs are 0.27 percent of the 16,925 runs observed, and rank its domain 274th of the 23,978 domains the release observed at all — the only member of the cohort with a rank or a grade. It was observed on five of the six production engines, which is a floor on its breadth rather than a measurement of it, because per-engine capture is not uniform across the six.
The remaining six sit between one and fourteen runs, on one to eleven days out of 133. Nscale announced the largest round of the week, $3.36 billion in pre-IPO convertible financing, reported by TechCrunch, and its domain was cited in four answers on four days in a 133-day window.
Does the size of the round predict citation presence? #
Partly, and less than the rank correlation suggests. Across the twenty-three dollar-denominated rounds, the Spearman rank correlation between capital announced and cited answer runs is 0.60. That is a real positive association, and it is mostly one fact restated: sixteen of those twenty-three rounds are tied at zero citations, and they are concentrated among the smaller rounds.
Among the seven companies that appear at all, size and presence disagree. Ordered by capital, the cohort's top five rounds — Nscale at $3.36 billion, Island and Cyera at $400 million each, Snorkel AI at $350 million and Heidi at $340 million — account for 38 of the cohort's 87 cited runs between them, and one of those five, Island, has none. The other 49 runs belong to companies further down the list. The eleventh-largest round in the cohort holds more than half the cohort's total citation presence.
Two rounds of nearly identical size, announced two days apart, sit on opposite sides of the line: Cyera's $400 million Series G extension, covered by SecurityWeek, appears in nine runs; Island's $400 million Series F, covered by CNBC, appears in none. Capital is not the variable that separates them.
What "collecting" means here, and what it does not #
Six of the seven companies that appear carry the label collecting rather than a confidence grade. That label is a statement about evidence, not about performance. The Index publishes a rate for a stratum only once it clears the release's evidence floor, which in this release is ten observed runs across seven observed dates; below that, a rate is not yet interpretable and the release says so rather than printing a number. The distinction between an unpublished rate, a measured zero and a category that cannot be collected at all is set out in Why 'Collecting' Is Not a Zero Citation Rate.
So the correct reading of the seven is not "six of them scored badly." It is that one of the twenty-four has accumulated enough observed evidence for the release to grade it, and the other six have not.
Why absence is measured against the release's own run universe #
The seventeen are a different case, and the claim about them has to be stated as narrowly as the evidence allows.
Seventeen domains have no citation event in the release. Not a low rate — no event. The denominator is the release's own observed run universe: 16,925 eligible answer runs on six engines over 133 dates, producing 132,514 source events across 23,978 domains. None of those events names any of the seventeen domains.
What that supports: within this question set, on these engines, over this window, these domains were not a source. What it does not support: that no engine would ever cite them, that they were never mentioned by name in an answer, or that a buyer typing their company name would fail to get an answer about them. A question set is a sample of demand, not all of it, and a domain absent from a sample can still be cited outside it. The measure here is source selection — whose page the answer is built from — and nothing else.
That boundary is the ordinary one for an observational record: an incomplete record is not an observed absence, and an observed absence is not an explanation. The same discipline is standard in measurement practice generally, where the denominator and the missingness rule are part of the result rather than formatting around it (NIST/SEMATECH e-Handbook).
What the window can and cannot show about the announcements #
The window ends September 26, 2026, the last day of the announcement week. The rounds were announced between September 21 and September 25. So for all but the final days, the measurement describes the standing citation presence each company brought into its announcement, not the effect the announcement had.
That is the useful thing about it. It is a baseline taken at the moment of maximum coverage: twenty-four companies, each with a fresh primary announcement and at least one outlet write-up, across TechCrunch, Tech.eu, SiliconANGLE, The Next Web, RuntimeWire and the wires, including PR Newswire. Several companies published detailed primary announcements of their own on the day, among them Enveda and Kontext. Seventeen of the twenty-four entered that week with no recorded presence in the answer layer at all.
What this window cannot show is whether the coverage changes that. A week of announcement coverage entering the crawl and retrieval layer takes longer than four days to appear in a source-selection measure, if it appears at all.
The re-read, registered in advance #
The cohort is fixed and the measure is reproducible, so the follow-up is worth committing to before the result is known. On the release dated October 26, 2026 — thirty days after this cohort's announcement week — the same twenty-four domains will be read again from the same field, mri_score_v2 overall runs_cited, against the same window start of May 10, 2026, and the delta published whatever it says.
Three outcomes are distinguishable in advance. If several of the seventeen move off zero, announcement coverage feeds source selection on a measurable lag, and the lag is the finding. If the seventeen stay at zero while the seven grow, presence compounds where it already exists and announcements do not create it. If nothing moves, the announcement week is not an input to source selection at this scale at all.
Registering the outcomes now is what makes the re-read evidence rather than a search for a story in a second data pull.
What a company can read from this #
One number is worth carrying out of this: of twenty-four companies that had just raised, and every one of which had press, one had accumulated enough presence in the answer layer to be graded.
The implication is about sequence rather than spend. Announcement coverage and answer-layer presence are being produced by different mechanisms, and the cohort shows the second one is not a fast by-product of the first. A company that needs to be a source when a buyer's assistant researches its category is looking at a different and slower kind of work than a funding announcement, and the day the round is announced is a late moment to start it.
For a company that wants to know where it currently stands rather than where this cohort stands, the release's own domain lookup is public and reports the same fields used here.
FAQ #
What exactly was measured?
For each of the 24 domains, the count of distinct eligible answer runs in which that domain was cited, from the field mri_score_v2.overall.runs_cited in the Machine Relations Index release mri_score_v2.0+2026-09-26+2b779408cfda, window May 10 to September 26, 2026, 133 observed dates, 16,925 eligible runs. A domain with no record in the release has no citation event in the window.
Does "zero cited runs" mean AI engines do not know these companies? No. It means their domains were not cited as a source in the release's observed answer runs. An answer can name a company without citing its site, and a question set is a sample of demand rather than all of it. The claim is about source selection within a defined run set.
Why do six of the seven have no confidence grade?
The release publishes a graded rate only above its evidence floor of ten observed runs across seven observed dates, and reports collecting below it. Collecting means the evidence is not yet interpretable as a rate, not that the rate is zero.
Could the seventeen be cited on engines the Index does not observe? Yes. The release observes six production answer engines. A citation on a surface outside that roster is outside this measurement, and per-engine capture within the roster is not uniform, which is why the single breadth figure reported above is stated as a floor.
How were the 24 companies chosen? By a published inclusion rule applied before this measurement: a self-described AI company announcing a round between September 21 and 26, 2026, with amount, round and lead investor confirmed at the company's own announcement or the outlet that reported it. Eight companies announcing rounds that week were excluded for not describing themselves as AI companies.
What would falsify the reading? The October 26, 2026 re-read moving a substantial share of the seventeen off zero. That would show announcement coverage does feed source selection on a lag, and the conclusion about sequence would need restating.