Why I Build for the Citation Market

I would build the source list from citation rate on one denominator, and I would retire the press list that ranks outlets by how they look in a deck. There is no universal best publication. The same domain changes standing when the question changes shape. A static list cannot see that, so it keeps sending pitches to newsrooms the answer does not retrieve and ignoring domains a founder would never put on a slide.

The dataset is Which Publications AI Search Engines Cite Most in 2026. That page is the deeper resource. This essay is how I would use it as a founder: how I would read a rate, which segment I would open before a pitch, and what new release would put a prestige list back in charge.

One denominator, read before the rank

Machine Relations Index release mri_score_v2.0+2026-09-20+47973f373a20 covers 2026-05-10 to 2026-09-20, 127 days, 919 monitored prompts, 16,039 answer runs, 22,320 cited domains, and six engines. A domain counts once per answer run in which it is cited. Nothing gets a rate or a rank until it clears 10 observed runs across 7 distinct dates. I would not compare a vendor's "top publications" slide to this table unless the vendor can match those fields. A count with no denominator is inventory.

Publication	Citation rate	Cited runs	Days cited	Engines	Confidence	Standing
medium.com	5.04%	808	109 of 127	6	A	rank 1 of 1,230
forbes.com	4.15%	665	106 of 127	6	A	rank 2 of 1,230
techradar.com	2.87%	460	—	—	A	rank 3 of 1,230
yahoo.com	1.93%	310	—	—	A	rank 4 of 1,230
businessinsider.com	0.62%	100	41 of 127	5	B	rank 20 of 1,230
prnewswire.com	1.13%	182	75 of 127	6	B	rank 8 of 1,230
techcrunch.com	1.02%	164	44 of 127	5	B	rank 9 of 1,230
fortune.com	0.37%	59	28 of 127	4	C	rank 36 of 1,230
venturebeat.com	0.29%	46	29 of 127	6	C	rank 61 of 1,230
thenextweb.com	0.12%	20	14 of 127	5	Collecting	below the floor for a rank
reuters.com	0.08%	13	13 of 127	3	Collecting	below the floor for a rank
techbullion.com	0.07%	12	11 of 127	3	Collecting	below the floor for a rank

The editorial-class rank comes off the Index source_role field. In this release that field mixes source-type evidence with a placement catalog. 96 of 1,231 class members enter through the catalog, and the classification is moving to source-type evidence alone. The class currently includes six platform domains next to edited newsrooms: medium.com at 1, amazon.com at 7, prnewswire.com at 8, apple.com at 16, wordpress.com at 239, and blogspot.com at 330. Substack, beehiiv, and dev.to are filed as community platforms even though they are the same kind of hosted publishing product as Medium. Those six platform domains hold 1,358 of the class's 13,525 cited runs, 10.04%, and they include the top slot. The class denominator itself moves, with published figures from 1,025 to 1,230, because each release counts the domains it classified. Rate, cited runs, days, engine breadth, and confidence are observations on a fixed run set. I would decide from the rate. I would use the rank as a label that can shift when the class rules shift.

TechRadar and Yahoo are third and fourth on this denominator, both confidence A. Neither sits on the press list I was handed for years. Lower down, entrepreneur.com is cited in 18 runs and techbullion.com in 12, both below the floor. I would not pay a premium for a name that has not cleared 10 runs on 7 dates, and I would not ignore Yahoo because it feels like a portal. The engines are selecting it.

Counts rank syndication. Rates rank selection

A 30-day window can fall. A cumulative URL count only climbs, and what it climbs with is how many URLs a source pushed into the corpus. One press release becomes hundreds of syndicated copies. One newsroom story stays one story. Counting URLs ranks the copy machine. Counting answer runs ranks the selection.

The wire and press-release distribution class, nine domains, takes 232 cited runs out of 112,518 across the nine source classes, 0.21% of classified citation, and held 0 of 880 top-ten segment slots on the 2026-09-19 release. prnewswire.com, the largest wire domain in the Index, is classified as an editorial publication in this release rather than inside that wire class, so the nine-domain class excludes it. Counted with the wires, the class is ten domains and 414 of those 112,518 runs, 0.37%. prnewswire.com's best standing in any published segment is 15 of 448, so the class still holds none of the 880 top-ten slots. Rank 8 and the class share sit on two different denominators. I would say both out loud before I let a wire deck use the rank 8 as proof that distribution is the strategy. The methodology log is the place the class rules are corrected in public (methodology).

The segment is the list

The Index publishes segments by question shape, each with its own denominator and its own grade. A domain's standing on a best-tools question, a how-buyers-choose question, a top-list, or a news question can diverge from the overall table. Fortune ranks 36 overall and 31 of 236 on consumer-finance best-tools questions. That gap is the whole argument. If my buyer asks a best-tools question in a specific category, the overall rank is a context line and the segment is the target list.

Prestige does not rescue the overall table either. Reuters is cited in 13 of 16,039 runs, on 13 days, across 3 engines, below the floor for a rank. Medium is cited on 109 of 127 days across six engines. Forbes is on 106 across six. PR Newswire manages 75 days. TechCrunch manages 44. Being retrievable in a lot of places is a different event from being chosen on most days. I would read days-cited and engine breadth before I fell in love with a rate that came from a short burst.

The engines publish what they crawl and how a site can refuse them: Google, OpenAI, Perplexity, Anthropic. I would hold a citation claim to that standard. Window, denominator, engines, floor. The frame I would run the pitching calendar through is Machine Relations. When a list comes back ranked by "tier," I send the word to the glossary and replace it with the rate on the segment that matches the buyer's question.

What I would do

I would open the segment for the question shape we actually sell into, write down the domains above the evidence floor, and aim the next placements there. I would keep a wire release in the budget that announces a fact to the audiences the wire was built for, and I would not move that budget into the citation line because prnewswire.com is rank 8 on a class that also contains open publishing domains. I would ask any firm quoting a leaderboard for the release id. This one is mri_score_v2.0+2026-09-20+47973f373a20, read live on 2026-09-20. Segment top-ten counts in the companion analysis come from the 2026-09-19 release. A number that cannot name its release is a number I would not steer by.

Scope: commercial questions of the kind the Index monitors, in English, across the six engines in the release. A trade newsletter that a specific human buyer reads every morning can still be the right buy for that human. I would pay for it from the relationship budget. I would not let it displace a domain that is cited on 100 of 127 days if the goal I stated was the answer.

What would change my mind: a later release in which the overall ranking and the segment rankings converged, so a single press list described every question shape, and in which Reuters-class prestige started clearing the evidence floor on the commercial segments at rates near Medium's 5.04% or Forbes's 4.15%. If the best-tools segment and the news segment wanted the same domains, and prestige predicted recurrence, I would go back to a shorter list and spend less time in the segment tables. Fortune at 36 overall and 31 of 236 on one segment is a small version of the split. The wire class at 0 of 880 top-ten slots is the large one. While the split is that wide, the press list is a habit, and the citation market is the plan.

FAQ

Which publication would I call the best?

None, as a standing fact. On this release Medium leads the editorial class at 5.04% and 808 runs, Forbes is second at 4.15% and 665, TechRadar third at 2.87%, Yahoo fourth at 1.93%. The class rank is a label with a moving denominator. The rate is the observation. The segment that matches the buyer's question can reorder the table, the way Fortune's overall rank and its consumer-finance best-tools rank already differ.

Why would I not aim at Reuters first?

Because on this denominator Reuters is 13 runs, 13 days, 3 engines, below the floor. Trust in a newsroom and selection by an answer engine are different measurements. I would still brief Reuters when the human reader is the goal. I would not brief it as the way to enter a commercial answer that is currently being built from other domains.

How should I read a wire's rank?

prnewswire.com is rank 8 of 1,230 in the editorial class on this release, 1.13%, 182 runs, 75 of 127 days, six engines, confidence B. The wire class it is excluded from, nine domains, is 0.21% of classified citation and 0 of 880 top-ten segment slots. Include it, and ten domains are 0.37%, and its best segment standing is 15 of 448, so the top-ten count stays zero. I would quote the rank and the class share together or not at all.

What would I demand from a firm that shows a different top ten?

The window, the run count, the engines, and the evidence floor. This release uses 16,039 runs, six engines, and a floor of 10 runs on 7 dates. A top ten that cannot survive those questions is a sales list. I would also ask whether TechRadar and Yahoo appear. If a "founder press list" has no room for the third and fourth domains on this denominator, it is ranking taste.

When would a 30-day spike change the plan?

When it survived into days-cited and engine breadth on the next release. A cumulative URL count cannot fall, so it cannot tell me the spike ended. An answer-run rate can. I would rather miss a 30-day burst than retarget the company around a source that was selected on 13 days out of 127.