Two Publications With the Same AI Citation Rate Are Not the Same Buy

Your pitch list ranks outlets by how often they are cited. It does not tell you how many different days they showed up, and those are two different buys.

I read the Machine Relations Index release of September 23, 2026 and put a second column next to the one every media plan uses. Of the 1,266 domains the release classifies as editorial publications, 99 carry enough evidence to be scored at all. The median one of those 99 was in the cited set on 24 of 130 observed days. Seventy-seven of the 99 were present on 31 days or fewer.

And the two columns disagree. Rank those 99 publications by citation rate, then rank them by days present, and the two top-five lists share three members. The top ten share eight. NerdWallet and TechTarget sit next to each other on the rate column — 1.54% and 1.49% — and 31 days apart from each other on the second one. NerdWallet was present on 31 days. TechTarget on 83.

That changes one thing on Monday: before you fund an outlet, ask what its days-present number is, because rate alone cannot tell you whether you are buying a standing presence or a good week.

What was measured

The Index runs real buying questions against six answer engines every day and records which source domains get cited. Release mri_score_v2.0+2026-09-23+fdec6388001a covers May 10 to September 23, 2026 — 130 days observed, 16,475 monitored answer runs, 23,280 cited domains, 129,264 source events.

Two fields on every domain matter here.

Citation rate is runs cited over runs observed. It answers: across everything the panel asked, what share of answers included this domain.

Days present is the release's days_cited over days_observed: the count of distinct days in the window on which the domain was cited at least once anywhere in the panel. It answers a different question — on how many separate days did this source exist in the answer surface at all.

A domain can score well on the first and poorly on the second by being cited heavily on a few days. NerdWallet's 253 citations arrived across 31 days, an average of 8.2 per active day. TechTarget's 246 arrived across 83 days, 3.0 per active day. Same rate, different shape.

Only graded domains are in this analysis. A domain publishes a rate after clearing the evidence floor of ten observations across seven distinct run dates, and 1,167 of the 1,266 classified publications are still marked collecting. That matters: the single most common way to fool yourself with an index like this one is to compute a share over a population where most members were observed once. None of the 99 here were.

The publications a media plan actually reaches for

The fifteen highest-rate editorial publications in the release, with the second column beside the first.

Publication	Citation rate	Runs cited	Days present of 130	Categories published in
Medium	4.90%	808	109	17
Forbes	4.10%	675	109	18
TechRadar	2.79%	460	101	16
Yahoo	1.91%	315	71	18
NerdWallet	1.54%	253	31	5
TechTarget	1.49%	246	83	11
Amazon	1.44%	237	67	11
PR Newswire	1.12%	185	77	16
TechCrunch	1.00%	164	44	9
People Managing People	0.92%	152	51	4
The New York Times	0.86%	142	32	7
Axios	0.80%	131	60	10
CNBC	0.78%	129	31	10
Fierce Healthcare	0.75%	123	37	2
PCMag	0.74%	122	31	6

Read the fifth and sixth rows together, then the eleventh and twelfth. NerdWallet outranks TechTarget on rate and is present on 52 fewer days. The New York Times outranks Axios on rate and is present on 28 fewer days.

Most of the gap is footprint, and you can control for it

The honest reading first, because it is the part that would otherwise make this number look better than it is.

Days present rises with the number of subject categories a publication is scored in, and it rises steeply. The Index does not run every category on every date, so a publication that appears in two categories simply has fewer dates on which it could have been cited.

Categories published in	Publications	Median days present
1 to 2	35	19
3 to 4	12	24
5 to 7	30	24
8 to 11	16	30
12 or more	6	89

So the NerdWallet and TechTarget pair, taken alone, is partly a story about breadth: five categories against eleven. Comparing days present across publications with very different footprints measures the footprint.

The fix is to hold footprint constant and look inside a band. Thirty of the 99 publish in five to seven categories. Within that band, citation rate and days present correlate at r = 0.55 — related, but leaving most of the variation in presence unexplained by rate.

What survives the control

Inside the five-to-seven-category band, matched pairs:

Publication	Categories	Citation rate	Runs cited	Days present
NerdWallet	5	1.54%	253	31
Tech Insider	5	0.58%	96	40
The New York Times	7	0.86%	142	32
Built In	7	0.43%	71	33
PCMag	6	0.74%	122	31
TechRepublic	6	0.38%	62	36
WIRED	6	0.29%	47	13

Three pairs, footprint held constant, and in every one the lower-rate publication is present on as many days or more. Tech Insider is cited at roughly a third of NerdWallet's rate across the same number of categories and shows up on nine more days. Built In is cited at half The New York Times' rate and shows up on one more day. TechRepublic is cited at half PCMag's rate and shows up on five more days.

WIRED is the other end of the same axis: 47 citations compressed into 13 days, 3.6 per active day. On 117 of the 130 observed days, WIRED was not in the cited set anywhere in the panel.

None of this says the low-rate outlet is the better buy. It says the two numbers are answering different questions, and a list ordered on one of them is silent about the other.

What to do with this on Monday

Add the column. For every outlet on the target list, pull its domain profile from the public index and write down days present next to citation rate, plus the number of categories it is scored in. Three numbers, one row per outlet. Outlets that are still collecting have no rate and no presence figure, and that is information too — it means the Index has not seen them often enough to say anything, which is a different risk from a low score.

Compare inside a footprint band, never across. A specialist trade title in two categories and a general business title in sixteen are not comparable on days present. Group the list by category count first, then rank inside the group. Ranking the whole list on presence rewards breadth you were not buying.

Decide what you are actually buying. A high-rate, low-presence outlet is a burst: strong when its categories run, absent otherwise. A steady-presence outlet is coverage insurance. Most plans want some of each, and almost no plan says which it is buying, because the brief only ever carried one number.

Do not treat one placement as a presence change. Neither figure here is an outlet's ability to carry your brand into an answer. They describe the outlet's own standing in the cited set. What a placement inside that outlet earns you is a separate measurement, and I would not let a rep present the outlet's number as the campaign's number.

The sales handoff

The line "we are cited in AI answers" does not survive contact with this data, and it is worth retiring before a prospect retires it for you.

Give the sales team a sentence with a denominator in it instead. If your plan rests on an outlet present on 31 of 130 observed days, then on most days a buyer asking a question in your category met an answer that outlet was not in. That is not an argument against the placement. It is the argument for what else has to be in the plan, and it is the sentence that stops a board asking in November why one placement did not move pipeline.

The worksheet fits on one row per outlet: outlet, rate, days present, categories, and the one question the outlet holds. I covered that last column separately in Most Publications AI Engines Cite Answer Exactly One Buyer Question, and the ceiling the whole list adds up to in Buy Every Publication AI Cites in Your Category and You Still Miss Half the Answers. Presence is the third column, and together the three are the shortest honest description of what an earned-media line item buys.

Limits

Days present is a panel-wide field, not a per-category one. It records that a domain was cited somewhere in the panel that day, so a domain scored in many categories has more chances to register, which is exactly the effect the footprint table shows and the reason the matched comparisons are restricted to one band.

The release reports how many runs cited a domain, not which runs, so I cannot say whether two publications were present on the same days or different ones. Every comparison here is between marginal distributions.

A 130-day window is a window. A publication that changed its coverage in August carries its July behaviour in these totals, and nothing in the public artifact separates the two halves — the release's drift series records daily run counts, not per-domain daily presence.

And the direction of the whole finding is one-sided in a useful way: presence measured this way can only overstate how often an outlet is available to a buyer, because being cited once anywhere in a six-engine panel on a given day is a very low bar to clear. The real figure a buyer experiences is at or below these numbers.

Prior work on how much the cited set churns day to day, across all source classes rather than publications alone, is at AI Citation Volatility.

FAQ

Is a low days-present number a reason to drop an outlet? No. It is a reason to stop describing that outlet as standing coverage. A publication present on 31 of 130 days with a high citation rate is concentrated, not weak, and concentration is fine if you know you bought it.

Why are only 99 of 1,266 publications in this? The other 1,167 have not been observed enough to publish a rate. The evidence floor is ten observations across seven distinct run dates, and a share computed over domains observed once or twice describes the instrument rather than the market.

Does this replace ranking outlets by citation rate? It adds to it. Rate is still the right first sort. Presence is the second sort, and the footprint count is the control that makes the second sort mean anything.

Can I check these numbers myself? Yes. Every figure is from the public release file at machinerelations.ai, and each domain has its own profile page carrying its rate, its confidence grade and its category footprint.

Methodology and sources

Single source for every figure: Machine Relations Index release mri_score_v2.0+2026-09-23+fdec6388001a, artifact SHA-256 fdec6388001adba06a783b83bbd8a0c0f38a182b91fb81f3079524d6773710eb, window May 10 to September 23, 2026, 130 days observed, 16,475 monitored answer runs, 23,280 cited domains, 129,264 source events, six engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity.

Method: filter the release's cited domains to the 1,266 it classifies as an editorial publication; keep the 99 carrying a graded overall record, being 4 graded A, 16 graded B and 79 graded C; read each one's overall citation rate, runs cited, days cited against 130 days observed, and the count of distinct subject categories in which it holds a published segment record. Rankings are read from the release's own fields. The footprint bands group publications by that category count, and the correlation quoted for the five-to-seven band is Pearson's r over the 30 publications in it.

Limits are stated in full in the section above, and the two that bound every comparison are that days present is panel-wide rather than per category, and that the release publishes marginal counts rather than the runs themselves.

Machine Relations as an operating framework — treating the answer surface as a system to be sourced and measured rather than a channel to be pitched — is documented at AuthorityTech.