How AI Citation Behavior Changes Over Time: Recurrence, Not Recalibration, in the Machine Relations Index

Does AI citation trust update in discrete windows, or does it accumulate? The column that answers it is days cited: the number of distinct dates on which a publication was cited at all. On that column, the publications engines trust are not the ones that spike in a given week. They are the ones engines come back to on most days, on every engine.

What the measurement shows

Machine Relations Index release mri_score_v2.0+2026-09-20+47973f373a20, window 2026-05-10 to 2026-09-20, 127 days, 919 monitored prompts, 16,039 answer runs, 22,320 cited domains, six engines. A domain counts once per answer run in which it is cited, and nothing publishes a rate or a rank until it clears an evidence floor of 10 observed runs across 7 distinct dates.¹

Publication	Citation rate	Cited runs	Days cited	Engines	Confidence
forbes.com	4.15%	665	106 of 127	6	A
techcrunch.com	1.02%	164	44 of 127	5	B
businessinsider.com	0.62%	100	41 of 127	5	B
fortune.com	0.37%	59	28 of 127	4	C
csoonline.com	0.22%	35	22 of 127	6	C
reuters.com	0.08%	13	13 of 127	3	Collecting
prnewswire.com	1.13%	182	75 of 127	6	B
medium.com	5.04%	808	109 of 127	6	A

Forbes is cited on 106 of 127 separate days across all six engines. Medium on 109. That is not a publication that won one re-weighting window and coasted. That is a source engines return to, on different questions, on most days, on every engine. Recurrence is the signal, and only a measure with a concept of a day can see it.

Medium is the single most-cited editorial publication in the release, first of 1,230, and its pattern is the same as Forbes's: steady, broad, daily.

How to read the editorial-class rank. The editorial-class rank above is read off the Machine Relations Index's source_role field, which in the current release combines source-type evidence with AuthorityTech's placement catalog; 96 of the 1,231 class members are admitted through the catalog, and the Index is moving editorial classification to source-type evidence alone. The class currently includes six platform domains alongside edited newsrooms: medium.com (class rank 1), amazon.com (7), prnewswire.com (8), apple.com (16), wordpress.com (239) and blogspot.com (330). At the same time substack.com, beehiiv.com and dev.to — the same hosted open-publishing product as Medium — are filed as community platforms. Those six hold 1,358 of the editorial class's 13,525 cited runs, 10.04%, and they include the class's top slot. The class denominator also moves between releases (published figures carry 1,025 to 1,230) because it counts the domains classified in each release. What is unaffected: the citation rate, cited runs, days cited, engine breadth and confidence on this page are direct per-domain observations over a fixed run set, and are the primary measure; read the rate first.

Why weekly spikes mislead

A cumulative counter's seven-day delta is not a spike. It is just recent coverage. Whether that delta looks like 85% of the total or 5% of it depends on how big a base the source had already accumulated. A source with a large indexed base barely moves; a source with a small base looks like it exploded. The shape that produces is exactly what you get when a new index ingests a high-volume source first and reaches the long tail later: steady wire, spiky editorial. It describes the ingestion curve, not machine cognition.

Whether AI engines update source weighting continuously or in discrete steps is a real and open question, and the Index does not claim to answer it. What it does measure is recurrence, and recurrence is what a founder can act on.

Conductor's 2026 AEO/GEO benchmarks report documents that AI answer presence shifts in ways that do not track organic search movement. That is their finding at their own source.

Three checks for any citation trend

Ask whether the number can go down. If a metric has never once decreased, it is not measuring a window; it is measuring an accumulation, and every derived claim about velocity, timing, spikes or momentum is arithmetic about the counter's own age.

Ask how many engines observed it, because a source seen by one engine is a crawl artifact rather than a market position. And ask for the evidence floor below which the thing refuses to publish a rate at all. The engines are specific about what they crawl and how to control it: Google's common crawlers, OpenAI's bots, Perplexity's bots, Anthropic's crawler policy. A citation claim should be able to match that specificity.

What this means for founders

Consistent editorial presence beats quarterly bursts. Not because you must be present when a window opens, but because the engines come back on most days, and the publications they return to are the ones that keep showing up. Forbes on 106 days of 127. Medium on 109. There is no window to catch. There is a habit to be part of.

If you want to see where your brand sits on a measured denominator rather than a counter, run an AI visibility audit.

Updated 2026-09-23: figures reflect the current Machine Relations Index methodology; see the MRI methodology and update log.

Notes

Footnotes

1. Machine Relations Index, release mri_score_v2.0+2026-09-20+47973f373a20, window 2026-05-10 to 2026-09-20, 919 monitored prompts, 16,039 answer runs, 22,320 cited domains, six engines. Read live 2026-09-20 at https://machinerelations.ai/index and the linked domain profiles. ↩