Buy Every Publication AI Cites in Your Category and You Still Miss Half the Answers

Your media plan has a ceiling, it is lower than anyone quotes you, and you can compute it before you spend.

I read the Machine Relations Index release of September 23, 2026 and asked the budget question rather than the pitch-list question: if you won a placement in every publication AI engines cite in your category, what share of the answers would you be in? Across the 86 published buyer-question segments, the best single publication is cited in a median 12.6% of that segment's answer runs. The top five summed reach a median 33.8%. Every cited publication in the segment, summed, reaches a median 51.0% — and in 39 of those 86 segments, buying the entire editorial class still cannot reach half the answers.

Those are ceilings, not estimates. I explain below why the real number is lower.

That changes one thing on Monday: earned media stops being the plan and becomes a line item with a known maximum, and the brief has to say what the other half of the answer is made of.

What a ceiling means here, and why it is generous

The Index runs real buying questions against six answer engines every day and records which source domains get cited. The September 23, 2026 Machine Relations Index release covers May 10 to September 23, 2026 — 130 days, 16,475 monitored answer runs, 23,280 cited domains, 129,264 source events.

A category paired with a question shape is a segment. Six shapes are what a buyer asks while deciding: best tools, how buyers choose, is it worth it, problem-first research, top lists, and head-to-head comparisons. A segment publishes once it clears the evidence floor of ten observations across seven distinct run dates. Ninety-three segments are published; 86 sit in those six buyer shapes and cover 9,757 answer runs between them.

For each segment I took every domain the release classifies as an editorial publication, read the runs it was cited in within that segment, sorted them, and added them up. That sum is an upper bound on coverage, because adding two publications' run counts assumes they never appear in the same answer.

They do appear in the same answer. The median buyer segment cites 6.29 distinct domains per run. With six sources in a typical answer and sixteen publications competing for those slots, overlap is not a risk, it is arithmetic. Every number in this piece is therefore the best case, and the true coverage sits at or below it. I published the ceiling rather than an estimate because a ceiling is falsifiable and an estimate of overlap I cannot measure would not be.

The ceiling, by how many placements you buy

Median across the 86 published buyer-question segments, with the count of segments where that many placements could in principle reach half the answers.

Publications bought	Median coverage ceiling	Segments that could reach 50%
The single best one	12.6%	0 of 86
Top 3	25.8%	10 of 86
Top 5	33.8%	19 of 86
Top 10	43.6%	35 of 86
Every publication cited at all	51.0%	47 of 86

Read the first row and the last row together. There is no segment in this release where one placement, however good, can put you in half the answers. And there are 39 segments where the whole class cannot, at any budget, because the publications simply are not in those answers.

The median segment has 16 editorial publications cited in it at all. That is the entire buyable universe for that buyer question, and in most segments it is not enough.

The shape you are pitching against matters more than the outlet

The same six shapes behave very differently. Medians across the segments in each shape:

Buyer question shape	One placement	Top 5	Every publication
Best tools	19.2%	46.6%	66.5%
Top lists	17.3%	38.1%	65.8%
Head-to-head comparisons	15.3%	33.7%	43.9%
Is it worth it	12.1%	33.8%	59.2%
How buyers choose	11.7%	31.9%	42.4%
Problem-first research	6.9%	21.4%	33.0%

Earned media buys the most answer share on "best tools" and "top lists." It buys the least on problem-first research — the question a buyer asks before they know your category exists, where the median ceiling for the entire editorial class is 33.0%.

If your campaign brief says "thought leadership in the consideration stage," it is aimed at the two shapes where placements do the least.

A worked example you can check

Enterprise software, head-to-head comparisons. 114 monitored answer runs. Every editorial publication the Index has observed being cited in that segment, with the running ceiling:

#	Publication	Runs cited	Share of the 114 runs	Running ceiling
1	TechTarget	14	12.3%	12.3%
2	Forbes	7	6.1%	18.4%
3	TechRepublic	4	3.5%	21.9%
4	Spiceworks	4	3.5%	25.4%
5	TechRadar	3	2.6%	28.1%
6	OpenReview	3	2.6%	30.7%
7	ITPro	1	0.9%	31.6%
8	HR Executive	1	0.9%	32.5%
9	beqom	1	0.9%	33.3%

Nine publications. That is the whole list. Win all nine and the ceiling is 33.3% of the answers, with the realistic figure lower because those nine overlap.

Cybersecurity, how buyers choose, runs the same way: 131 runs, 20 publications cited, and all twenty summed reach 36.6%.

So what is in the other two thirds

This is the part the media plan never shows you. Across all 86 buyer segments there were 64,165 cited-domain observations. By source class:

Source class	Observations	Share
Other observed source	46,594	72.6%
Editorial publication	5,752	9.0%
Vendor-owned source	3,503	5.5%
Community and social platform	2,889	4.5%
Academic and government source	2,087	3.3%
Market and company database	1,427	2.2%
Search or media platform	1,403	2.2%
Analyst and consulting research	465	0.7%
Wire and press-release distribution	45	0.1%

Editorial publications are 9.0% of what gets cited in a buying answer. Press-release distribution is 0.1%.

In that enterprise software comparison segment, the most-cited domain of any class is not a publication — it is erpfocus.com at 20.2% of runs, then LinkedIn at 15.8%, then TechTarget at 12.3%, then two vendor documentation sites, ServiceNow at 11.4% and Atlassian at 10.5%. In cybersecurity's how-buyers-choose segment the top two are Microsoft at 21.4% and Reddit at 18.3%, both ahead of every publication in the segment.

Independent niche sites, community threads and vendor documentation are where the answer is actually assembled. None of them take a pitch.

What to do with this on Monday

Compute your ceiling before the campaign, not after. Pull your category and the question shape you are targeting from the public index, list the editorial publications cited in it, add their run counts, divide by the segment's runs. That number is the most your earned-media budget can buy. If it is 33%, stop presenting the campaign as an AI visibility strategy and start presenting it as a third of one.

Pick the shape, then the outlet. A placement aimed at "best tools" is worth roughly three times one aimed at problem-first research, measured as answer share. Put the shape in the brief.

Give sales the ceiling, not the placement. When a rep says "we were in Forbes," the useful number is the share of buyer answers that changed. In most segments one placement moves you from zero to about an eighth of the answers. That is a real result and it is not the whole shortlist, and saying so in the handoff is what stops the board asking why pipeline did not move.

Budget the other 91%. The classes that hold it — your own documentation, comparison and database listings, community presence — are not PR work and will not arrive from a PR budget. That reallocation is the actual decision this data forces.

FAQ

Is this saying earned media does not work for AI visibility? No. It says earned media has a computable maximum in each segment, the maximum is usually well under half, and treating it as the whole strategy guarantees a gap. A 12.6% median for one placement is a meaningful move from zero.

Why is the real coverage lower than these numbers? Because the sum assumes two publications are never cited in the same answer. The median segment cites 6.29 domains per run, so they frequently are. Every figure here is an upper bound.

Why 86 segments and not all 93? Seven published segments are news-topic rather than buyer-decision shapes. They behave differently — the editorial class is much larger there — and mixing them into a buying analysis would flatter the result.

Does this change the per-question pitch list? It bounds it. The pitch list still has to be built per question shape, because a publication's value is capped by the shape it holds. This adds the second question: whether the whole list is worth the line item in that segment.

Methodology and sources

Single source for every figure: Machine Relations Index release, 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.

Method: filter the release's cited domains to the 1,266 it classifies as an editorial publication; for each, read its own per-segment record, which carries runs cited and runs observed for every published segment it stands in; restrict to the 86 published segments in the six buyer-question shapes; within each segment sort publications by runs cited and take the running sum over the segment's runs observed, capped at 100%. Segment standings are read from that field, never re-derived by sorting a leaderboard. Cited domains per run is the segment's total cited-domain observations over its runs observed.

Evidence floor: a segment publishes at ten observations across seven distinct run dates. Segments still collecting are excluded, which is why 93 of 157 are published.

Limits: the running sum is an upper bound on coverage and not a measurement of it, because the release reports how many runs cited a domain and not which runs, so the overlap between two publications cannot be computed from the public artifact. A ceiling that is already low is informative; a ceiling that is high says little. Source-class labels are the Index's classification and a domain sits in exactly one class.

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