The #1 'Editorial Publication' in Our Own Index Has No Editors

I have spent the last week correcting my own posts about which publications AI engines cite. Every correction leaned on the same source-role column in the Machine Relations Index: editorial publications over here, wire distribution over there, and the argument falls out of the gap between them.

That column has a defect, and it sits at the top of it.

medium.com is filed as an Editorial publication and ranks #1 of 1,230 of them. substack.com is filed as a Community and social platform. Those are the same product. A company runs the hosting, an author writes and presses publish, and nobody at the company reads it first.

One of them is the headline row of our editorial leaderboard. The other is not in the class at all.

The measurement

Machine Relations Index release mri_score_v2.0+2026-09-20+47973f373a20, window 2026-05-10 to 2026-09-20, 127 observed days, 16,039 monitored answer runs across six engines, 22,320 cited domains. A domain counts once per answer run in which it is cited. Nothing publishes a rate or a rank until it clears an evidence floor of 10 observations across 7 distinct run dates.

Domain	Source role in the release	Citation rate	Cited runs	Days cited	Engines	Confidence	Standing
medium.com	Editorial publication	5.04%	808	109 of 127	6	A	#1 of 1,230 editorial publications
forbes.com	Editorial publication	4.15%	665	106 of 127	6	A	#2 of 1,230
substack.com	Community and social platform	1.55%	249	76 of 127	5	B	#3 of 27 community platforms
dev.to	Community and social platform	1.30%	209	56 of 127	6	B	#4 of 27
wordpress.com	Editorial publication	—	11	11 of 127	4	below floor	no published rate
blogspot.com	Editorial publication	—	6	6 of 127	3	below floor	no published rate

The counts are not in dispute. Medium really is cited in 808 of 16,039 monitored answer runs, on 109 of 127 days, by all six engines, at confidence A. That is a real, stable, high-breadth observation and nothing below takes it away.

What is wrong is the word next to it.

WordPress.com and Blogspot are also filed as editorial publications. Beehiiv is filed as a community platform. So the release does not have a rule that says "hosted open publishing is editorial" or "hosted open publishing is community." It has both rules at once, applied domain by domain, and the split runs straight through the middle of one product category.

This is not one bad row

Once you look for it, the same failure shows up in three separate families inside the same taxonomy. In each one, a domain sits in the editorial class while the class that actually describes it exists in the same release, already populated with its structural peers.

Domain	Filed as	Peers already filed elsewhere	The class that fits
medium.com (#1)	Editorial publication	substack.com, beehiiv.com, dev.to	Community and social platform (27 domains)
amazon.com (#7)	Editorial publication	walmart.com, target.com, ebay.com	Other observed source / vendor-owned
prnewswire.com (#8)	Editorial publication	businesswire.com, globenewswire.com, prweb.com	Wire and press-release distribution (9 domains)
apple.com (#16)	Editorial publication	microsoft.com, ibm.com, salesforce.com	Vendor-owned source (823 domains)

Para Labs published the Amazon half of this in detail on 2026-09-20: Amazon ranks #7 of 1,230 editorial publications while Walmart, Target, Best Buy and eBay are filed as other observed sources. I corrected the PR Newswire half across my own archive the same day. I had both pieces in hand and still did not look at the top of the list.

What it costs, in the index's own arithmetic

The editorial class holds 13,525 cited runs — 12.02% of the 112,518 classified cited-domain observations in the release. Six domains inside it have a better-fitting sibling class in the same taxonomy:

Domain	Cited runs
medium.com	808
amazon.com	236
prnewswire.com	182
apple.com	115
wordpress.com	11
blogspot.com	6
Total	1,358

That is 10.04% of the editorial class's cited runs, and it includes the rank-1, rank-7 and rank-8 rows. Three of the class's top eight.

Every editorial rank computed off this release — including the ones I published — is a rank inside a set assembled that way. The per-domain rates survive it completely. The class does not.

Why I am writing this instead of quietly fixing it

This is our index. Nobody outside had found this. The ordinary move is to patch the classifier, re-cut the release, and let the corrected leaderboard appear without comment.

Two reasons not to.

The first is that people are acting on it right now. "Medium is the #1 editorial publication AI cites" is the single most quotable row in the release, and read as written it tells a founder that publishing on Medium is earned media. It is not. Earned media means somebody who does not work for you decided your thing was worth publishing. A Medium post is owned media on rented land — you wrote it, you pressed publish, no one gated it. The citation rate is real and the distribution is real; the word "editorial" is what turns it into a recommendation, and the word is wrong.

The second is that this is the whole argument. We tell people not to trust a citation leaderboard they cannot reproduce. A leaderboard whose top row is misfiled, published by us, discovered by us, and then silently edited, is the thing we warn about. The instrument is public — the full release JSON and every per-domain profile are there — which means the only version of this that survives contact with a reader is the one where we say it first.

What I am changing, and what I am not

Not changing: any per-domain citation rate, count, engine breadth or confidence grade. Those are observations over a fixed run set and the classification error does not touch them. Medium is genuinely the fourth most-cited domain in the index out of 22,320. That is worth knowing.

Changing: every sentence I have published that reads a class rank or a class share off the source-role column. Sixteen posts on this site carry "rank 1 of 1,230 editorial publications" as a load-bearing phrase. They get the limit named in the page, the same way I handled the wire-class correction.

Asked for: the classifier needs one rule per object, not one per domain. Medium, WordPress.com and Blogspot belong wherever Substack and Beehiiv are. Amazon and Apple belong with the vendors. PR Newswire belongs with the wires.

The rule I should have been using

A source-role label is an interpretive judgment wearing the clothes of a measurement. The citation rate is counted. The class is decided, by someone, using a rule you usually cannot see. When a leaderboard sorts by one and is titled by the other, the title is doing work the data never did.

It is a reasonable habit to check whether a number is real. This week taught me to also check whether the noun in front of it is.

So: before you quote any "top cited publications" list — ours included — open the top row and ask what the publication actually is. If the answer is "a place where anyone can post," you are not looking at the editorial market. You are looking at a hosting company that got filed in the wrong drawer, and everything ranked below it inherited the error.

---

FAQ

Why is Medium ranked the top cited publisher by AI engines?

Because it is genuinely cited a lot, and because the Machine Relations Index files it in the editorial-publication class. On release mri_score_v2.0+2026-09-20+47973f373a20, medium.com is cited in 808 of 16,039 monitored answer runs (5.04%), on 109 of 127 observed days, across all six engines, at confidence A. The citation observation is sound. The class label is not: Substack, Beehiiv and Dev.to — the same kind of hosted open-publishing product — are filed as community and social platforms in the same release. The measured unit is domain-level citation presence in the observed run set. Boundary to preserve: this does not establish a universal source-selection mechanism, citation causation, a guaranteed future citation, recommendation lift, pipeline, revenue, or any other business outcome.

Does publishing on Medium count as earned media for AI visibility?

No. Earned media means an independent party chose to publish about you. A Medium post is self-published on a hosted platform with no editorial gate, which makes it owned content on someone else's domain. It may still get cited — Medium's citation rate is real — but a citation leaderboard that calls it an editorial publication is describing the hosting, not an editorial decision.

Is the Machine Relations Index citation data wrong?

The rates and counts are not. Citation rate, cited runs, days cited, engine breadth and confidence are per-domain observations over a fixed set of monitored answer runs, and a classification error does not touch them. What is affected is anything computed across a source-role class: class share, class rank, and any "editorial vs. wire" or "editorial vs. community" comparison. Six domains holding 1,358 of the editorial class's 13,525 cited runs — 10.04% — have a better-fitting class in the same taxonomy.

How much of the editorial class is misclassified?

By cited runs, 10.04%: medium.com (808), amazon.com (236), prnewswire.com (182), apple.com (115), wordpress.com (11) and blogspot.com (6), against a class total of 13,525. By position it is worse than the percentage suggests, because three of those six hold ranks 1, 7 and 8.

What should I use instead of a class rank?

The per-domain figures, which are measured rather than decided: citation rate against a stated denominator, cited runs, days cited out of days observed, engine breadth, and the confidence grade. If you need a comparison across source types, define the classes yourself and say what rule you used — regrouping alone can move a leaderboard without changing a single count.

Where can I check this myself?

Every figure here is on public pages. The release JSON carries by_source_role with the class totals, and each domain has a profile page — medium.com, substack.com, amazon.com, apple.com, prnewswire.com — showing source role, standing, citation rate and confidence.

For the peer platforms' own descriptions of what they are, read them side by side: Substack is a publishing platform for independent writers, Dev.to is a community for software developers, and WordPress.com hosts sites for anyone who signs up. Medium is the same kind of product and describes itself the same way. Then open a newsroom's masthead and note that there is one. That masthead is the whole difference, and it is the thing a source-role column is supposed to encode.

Independent work on the source composition of AI citations is published by BuzzStream, Fractl, Fullintel, Presenc AI and Ahrefs; each uses its own sample and denominator, so read the methodology before comparing any two. The mechanism behind why some pages are easier to extract than others is quantified in the GEO paper (Aggarwal et al., KDD 2024). AuthorityTech's own take on why placements and citations are different objects is here.