Founder Decision Ledger: Is Your Vertical the Right Unit?

The decision is this: stop treating your vertical as the unit of an AI citation strategy. The unit is the domain's observation count, and the vertical is a filter you apply after it.

Every media list template in circulation is organised by category. Tier your outlets, the guidance says, and put the trade press in its own tier because trade publications carry more authority inside a category than a general-interest outlet with ten times the circulation (Everything-PR, Cision). Segment by outlet type, beat and industry vertical (PRLab). Build an outlet universe of national, trade, regional and niche (ContentGrip, AMW). A placement in a high-quality industry vertical creates more qualified pipeline than a brief mention in a general business daily (Presskid). Tier 2 is where the industry publications go (MentionAgent, PR.co, Prezly).

That is good advice for reaching humans. I am no longer sure it survives contact with how AI engines actually select sources, and this post is me writing down why, and what I am going to do about it.

The ledger decision

Field	Decision
Decision	Do not build an AI citation target list from a category-scoped outlet universe. Build it from domains with a measured observation count, then filter for category fit.
Evidence	In the September 23, 2026 Machine Relations Index release, 86.6% of 23,280 cited domains appear in exactly one category — but 46.9% of those domains were observed exactly once in 130 days, and a domain cited once can only ever appear in one category. Conditioned on evidence the relationship inverts monotonically: at 50 or more observations, only 14.0% are single-category, and the median domain spans 5 of 20 categories.
Scope	Applies to earned-media and placement targeting for AI answer visibility. It does not apply to reaching human readers, where vertical relevance is still the right filter.
Action	For every outlet on a proposed list, ask the vendor or agency for its observation count and its category span. Cut any domain that cannot produce one.
Reversal condition	If a conditioned cut of a later release shows category span falling rather than rising as observation count rises, the unit is the vertical after all and this decision is wrong.

Why the vertical thesis looks so strong

Because the published numbers support it, loudly. A 2026 analysis of behavioral-health LLM citations segmented sources by domain rating and found DR 50-70 vertical publications carrying 64% of citations while DR 70+ generalist press carried under 3%, with a companion argument that mid-DR vertical publications are the most productive placement targets available (Webserv). The same firm's audit playbook tells you to cluster cited URLs by DR tier and turn the cluster into a pitch list, mid-DR vertical publications first (Webserv).

We have published a version of this argument ourselves, on the research side, reading those same third-party studies (Machine Relations Research).

Here is the problem. Put the widely cited concentration figures next to each other and they cannot all describe the same world:

Published figure	Source	What it implies
Top 15 domains capture 68% of all AI citation share	5W AI Platform Citation Source Index, via PR Newswire	A tiny, fixed, nameable market
97.4% of AI citations come from non-Tier-1 sources	Profound 2026, reported by Webserv	A very wide, long-tailed market
No single domain exceeds 5% of total citations across platforms	Profound, reported by Webserv	A flat market with no dominant domain
ChatGPT cites Wikipedia 47.9% of the time	Profound, reported by Webserv	One domain dominating one engine
Top 1% of domains, roughly 12 sites, capture 47% of citations	Everything-PR AI Overviews Citation Source Index, June 2026	A tiny market again, at a different magnitude

Two of those come from the same dataset and still disagree about whether any domain dominates. This is not a measurement dispute. It is a denominator problem: each number is computed over a different population — one engine or five, answers or links, domains or URLs, a curated head list or everything observed — and almost none of them say which. A founder cannot build a pitch list out of numbers that are not comparable to each other.

What our own data says once you condition it

We publish the Machine Relations Index, and the September 23, 2026 release is the first one where I could test the vertical thesis on our own population rather than on someone else's headline.

Release mri_score_v2.0, contract machine_relations_index_public_view_v2.0, generated September 23, 2026, window May 10 to September 23, 2026, 130 observed days, 16,475 monitored answer runs, 129,264 source events, 23,280 cited domains, six engines. Evidence floor: 10 observations across 7 distinct run dates. Confidence grades in this release: 14 A, 59 B, 441 C, and 22,766 still collecting.

The first reading was clean and would have made a great post: 86.6% of cited domains appear in exactly one category. Citation authority is category-local. Pitch your vertical.

It is false, and the tell was in the same output. 46.9% of all cited domains were observed exactly once in 130 days. A domain cited once can only ever appear in one category. The statistic was mostly a census of things we have barely seen.

Condition it on evidence and it inverts, monotonically:

Minimum observations in 130 days	Share of those domains appearing in exactly one category
1 or more (all 23,280)	86.6%
2 or more	74.8%
5 or more	60.5%
10 or more	48.6%
25 or more	27.3%
50 or more	14.0%

The median domain in the release spans 5 of the 20 categories carrying citation data. The 228 domains at 50 or more observations hold 25.0% of all 115,066 citation observations in the window. All 14 of the release's grade-A domains appear in more than one category — that is 14 rows, not a rate, and I am not going to dress fourteen observations up as a percentage.

So the sources that actually carry weight in AI answers are the ones that travel. Category span is not a nice-to-have on top of citation volume; on this data it is what citation volume looks like from the side.

The part where the vertical thesis is still right

Both shapes are real, and a founder decision that only carries one of them is a worse decision.

Thirty-two domains in this release cleared 50 observations inside a single category: omnimd.com at 105 in Healthcare Services, huntress.com at 94 in Cybersecurity, allaboutcookies.org at 91 in Family Software. Forty-one cleared 50 observations across ten or more categories. Deep category-local sources exist and they are reachable, which is exactly what makes the trade-press instinct feel right.

The difference is that there are 32 of them, they are nameable, and you find them by reading an observation count — not by pulling every publication that serves your industry and calling it a tier.

Correcting my own post

On April 14, 2026 I published a post on share of citation that said the top 3 cited domains in a single category capture a median of 47% of weighted citations, the top 10 capture 78%, and that below the top 10 you are functionally invisible.

Our own measurement says the ten most-cited domains in a category hold between 7.0% and 24.2% of that category's citations depending on the category, and that reaching half of a category's citations takes between 49 and 304 distinct domains (AuthorityTech). The 78% figure is wrong by a factor of three to ten against our data, and the strategic conclusion drawn from it — that anything outside a top ten is invisible — is wrong in the direction that costs founders the most money, because it tells you to stop looking at exactly the middle of the market where the reachable sources are.

The number came from a third-party study read as a portfolio constant rather than as one segment's result. That is the same defect as the table above: a share with an unstated denominator. I am correcting that post rather than leaving it standing, and this is the record of it.

The question to ask instead

One question replaces the whole tiering exercise. For any domain someone proposes you pitch:

How many times has this domain been observed being cited, over what window, and in how many categories?

A source list that cannot answer that is a list of publications someone thinks are important. A source list that can is a market map. If the answer for a given domain is "once", you are not looking at a citation target — you are looking at a row in a long tail, and 46.9% of the domains in our index are that row.

Methodology and sources

Our figures are computed from the public Machine Relations Index release mri_score_v2.0, contract machine_relations_index_public_view_v2.0, generated September 23, 2026, covering monitored answer runs from May 10 to September 23, 2026 across six engines: 130 observed days, 16,475 answer runs, 129,264 source events, 23,280 cited domains, 20 of 25 taxonomy nodes carrying citation data. The evidence floor for a scoreable domain is 10 observations across 7 distinct run dates. Category-span figures are computed per domain over the 20 categories carrying data; the conditioned series reports, for each minimum observation count, the share of qualifying domains appearing in exactly one category. Per-category run denominators sum to 16,475, which is the release's own published run count.

Third-party figures are reported as their publishers state them and are not reconciled to each other, which is the point of the comparison table. The 5W figure is carried via its PR Newswire release; the Profound and behavioral-health DR-tier figures via Webserv; media-list construction practice via Everything-PR, Cision, Presskid, ContentGrip, PRLab, AMW, MentionAgent, PR.co and Prezly.