AI Cited 155 of the 1,621 Domains in Our Placement Catalog

Every AI visibility proposal you will read this quarter is built on a list of publications. Ours included. Before you buy one, there is a single number to ask for, and it is rarely in the proposal: what share of the sources AI engines actually cite in your category does that list reach?

We ran the question against our own catalog. The answer is uncomfortable enough to publish.

Our placement catalog holds 1,621 distinct domains. The Machine Relations Index observed 24,639 domains being cited across 17,868 monitored answer runs in the 135 days from 10 May to 28 September 2026. The overlap is 155 domains, which is 9.6% of our list and 0.6% of what the engines cited.

That is not an argument against earned media. It is an argument against buying a list before you have measured a category.

What the two numbers mean, separately

They are different failures and they need different responses.

Read from the list's side: 1,466 of the 1,621 domains in the catalog went unobserved in this window. Not cited rarely. Not cited at a low rate. Absent from the measurement entirely, across six engines and 1,010 buying and research questions. A placement in any of those is a real placement. It can carry a brand, a founder, a claim, a link. It produced no observed citation inside this basket.

Read from the engines' side: 155 of 24,639 is 0.6% coverage. The list is a rounding error against the cited web. The gap widens as you move down the ranking.

The Index publishes a rank only for domains that clear its evidence floor of ten observations across seven distinct dates. 567 domains clear it. That ranked set is the honest field of play, and it is where the bands below are measured.

Rank band	Catalog domains in it	Share of the band
Top 10	3	30.0%
Top 50	7	14.0%
Top 100	11	11.0%
Top 200	13	6.5%
All 567 ranked	22	3.9%
All 24,639 cited	155	0.6%

Read the last two rows together, because they are the finding. Of 1,621 domains we can place in, 22 are cited often enough and regularly enough to carry a published rank. The other 133 of the 155 were cited, but below the floor where the Index will state a rate.

The catalog is strongest exactly where every catalog is strongest, among the famous names at the top, and it thins immediately below them.

Six of the top ten are not placements at all

This is the part a media plan cannot fix by adding publications, because the sources doing the most citation work are not publications.

The ten most-cited domains in the release are reddit.com, youtube.com, linkedin.com, medium.com, forbes.com, nih.gov, arxiv.org, techradar.com, g2.com and gartner.com. Three are on our catalog. Of the seven that are not, six sit outside editorial placement in any sense a programme can sell: two community platforms, two academic and government archives, a software review marketplace and an analyst firm.

Each of those has its own mechanism, and each mechanism sits outside the earned-media line of a budget. Citation by nih.gov or arxiv.org follows from publishing research. Citation by g2.com follows from a customer review programme. Citation by gartner.com follows from an analyst relationship. Citation by reddit.com and linkedin.com follows from a community presence people choose to quote. A placement retainer buys none of these, and pretending otherwise is how a programme ends up accountable for results it was never built to produce.

The misses inside editorial are not obscure

It would be easier if the gap were all platforms and archives. It is not. Restrict the field to the source class a placement programme is built for, classified editorial publications, and the Index measures 1,340 of them. Our catalog covers 154, which is 11.5%.

The editorial properties inside the top 50 cited domains that our catalog does not carry: techradar.com at rank 8, nerdwallet.com at 16, techtarget.com at 19, cnbc.com at 36, nytimes.com at 45 and axios.com at 48.

Those are titles any communications lead would recognise, cited more often in this measurement than most of the titles on the list.

Why this is not the same as the ceiling question

Two adjacent findings are already public and this page is neither of them.

Christian Lehman published the class ceiling: win a placement in every publication AI engines cite in your category and, across 86 buyer-question segments, the entire editorial class reaches a median 51.0% of answers, with 39 segments where it stays under half. The full working is here. That question assumes you can buy them all and asks what the class is worth.

The Machine Relations Index separately measures how much citation weight editorial publications carry at all, which is the ceiling above any list, and we have argued that the right pitch list is category-specific rather than a fixed set of 15 sites.

This page asks the question underneath all three: what does an actual catalog reach? The ceiling is theoretical and generous. The catalog is what a vendor can really deliver, and the distance between the two is the part of a proposal that usually goes unpriced.

What "not observed cited" does and does not mean

The honest bound, because the number is only useful if you know its edges.

The Index runs a fixed basket: 1,010 eligible prompts across 24 measured subject categories plus a legacy bucket, against six answer surfaces, which are ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity, every day. A domain is recorded as cited when an engine cites it on one of those questions.

So "not observed cited" means not cited on these questions, by these engines, in these 135 days. A trade publication serving an industry the basket does not reach can be genuinely influential and still sit among the 1,466. That is a real limitation of the reading.

It is also the entire point. The test that matters is whether a publication is cited on the questions your buyers ask, not whether it is cited somewhere. A list assembled without reference to those questions will pass that test only by luck, and 1,466 of ours did not.

Two further limits worth stating. Rank in the Index is citation volume across the whole window rather than authority, and a domain cited on questions irrelevant to you is worth nothing to you. And most cited domains carry no source-role classification yet, which bounds every class-level share on this page; we wrote about what that unclassified majority does to source-class claims last week.

What to do with this on Monday

The order of operations is the whole recommendation, and most programmes run it backwards.

1. Name the questions. Six shapes cover how a buyer actually decides: best tools, how to choose, is it worth it, problem-first research, top lists, head-to-head comparisons. Write yours down in your buyer's words, not your category's marketing words.
2. Rank your category's cited domains before you look at any vendor list. The Index publishes a leaderboard per category and question shape. That ranking is the field you are competing in.
3. Intersect the vendor's list with that ranking and ask for the number. Not the list's size. Not its tier labels. The share of your category's cited domains it reaches, and the rank of the highest one it can deliver. Treat an unanswered question there as the answer.
4. Budget the unreachable classes separately. Community platforms, review marketplaces, analyst research, academic and government sources, and your own product pages arrive through their own mechanisms and their own owners. If they carry a meaningful share of your category's citations, they need their own budget line.
5. Re-measure after the placements land. A citation is an observable event. If a placement produced none in ninety days on the questions you named, that is data about the publication rather than about the campaign.

We run steps one through three for clients and publish the method rather than the client list. If you want the intersection computed against your own category, that is what AuthorityTech does.

The reason we can publish our catalog's 9.6% is that the measurement came first and the list is downstream of it. That is Machine Relations as a working discipline rather than a label: you measure which sources the machines actually read in your category, and you earn your way into those, in that order. A list built the other way around is a guess with an invoice attached. The category definition and the public method live at machinerelations.ai.

The measurement

Release mri_score_v2.0+2026-09-28+054b584266b8, published at machinerelations.ai/index. Observation window 10 May to 28 September 2026, 135 days. 17,868 monitored answer runs, 139,633 source events, 1,010 eligible prompts, 24,639 cited domains, six answer surfaces, 25 taxonomy nodes. A category paired with a question shape is a stratum; a stratum publishes a rate only after clearing ten observed runs across seven distinct dates, and 96 of 157 clear it in this release. A domain publishes a rank on the same evidence floor, and 567 of the 24,639 cited domains clear it.

The catalog figure of 1,621 distinct domains is read from the AuthorityTech opportunity catalog on 28 September 2026, counted as distinct domains so that it is comparable to the Index, which also counts domains. The 155-domain overlap is the Index's own at_catalog_member flag, which the Index records as a label and never uses to decide what gets scored: every domain an engine cites is measured on equal terms. Both counts were computed directly and reconcile in both directions.