Any Two Buyer Questions Share 2 Sources Of 10
Across the 220 pairs of buyer questions the Machine Relations Index publishes inside a single category, the two top-ten source lists share a median of 2 domains. Grouping the questions into shopping and research families predicts nothing: two shopping questions overlap no more than one of each. The sources that do carry are platforms, not publications.
The mechanic measured here is source carryover: when a shopper changes the question they ask about the same category, how much of the source list an AI engine draws on stays the same. Measured against the Machine Relations Index release mri_score_v2.0+2026-09-28+054b584266b8, generated 2026-09-28, the answer is a median of 2 domains out of 10. In 19 of 220 question pairs it is zero.
For a brand's leaders that number sets the scope of a placement programme. A brand does not need to be present in "AI answers about pet food." It needs to be present in six different source lists, one per question a shopper asks, and being in one of them predicts almost nothing about being in the next.
This is a cross-section, not an experiment. It reports which domains six engines cited over a fixed 141-day window. It makes no claim that any page, placement, schema change or campaign caused a citation, and it does not compare a before to an after. Para Labs — an applied research lab from the team behind the Machine Relations Index, and part of the same portfolio as AuthorityTech — publishes it as a measured baseline a later experiment can move against.
The measurement
| Field | Value |
|---|---|
| Release | mri_score_v2.0+2026-09-28+054b584266b8 |
| Window | 2026-05-10 to 2026-09-28, 141 days |
| Segments with a published top ten | 89, across 15 categories |
| Buyer question shapes | best_x best tools, top_list ranked roundup, x_vs_y head-to-head comparison, how_choose how buyers choose, is_x_worth is it worth it, problem_first problem-first research |
| Within-category question pairs compared | 220 |
| Engines | ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews, Perplexity |
| Evidence floor per segment | 10 observed runs across 7 distinct run dates |
A segment is one subject category paired with one buyer question shape, and it publishes a ranked leaderboard only after clearing that floor. Fourteen of the 15 categories publish all six shapes, giving 15 pairs each; HR & Talent publishes five, giving 10. That is the 220.
The comparison is set membership in the top ten, read off each segment's own public page — for example Consumer Products best-tools answers and Consumer Products problem-first research. The machine-readable release carries every row.
Two of ten
| Shared sources in both top tens | Question pairs | Share of pairs |
|---|---|---|
| 0 | 19 | 8.6% |
| 1 | 50 | 22.7% |
| 2 | 52 | 23.6% |
| 3 | 48 | 21.8% |
| 4 | 28 | 12.7% |
| 5 | 15 | 6.8% |
| 6 | 5 | 2.3% |
| 7 | 3 | 1.4% |
| All pairs | 220 | 100% |
The median is 2 and the mean 2.44. In 121 of 220 pairs — 55.0% — two questions about the same category share two sources or fewer out of ten. No pair anywhere in the release shares more than seven, and none shares all ten.
The grouping a brand would reach for does not work
The intuitive move is to sort the six questions into a shopping family — best tools, roundups, head-to-head comparisons — and a research family — how buyers choose, is it worth it, problem-first. Then buy for each family. The data does not support the split.
| Pair type | Question pairs | Median shared of 10 | Mean shared of 10 |
|---|---|---|---|
| Two shopping questions | 43 | 2.0 | 2.37 |
| Two research questions | 45 | 2.0 | 2.56 |
| One shopping and one research question | 132 | 2.0 | 2.42 |
Two shopping questions are no more alike than a shopping question and a research one. Same median, and the means sit inside a quarter of a slot of each other. Whatever makes two source lists diverge, it is not the family the question belongs to — so a two-list placement plan is not a smaller version of the right plan, it is the wrong unit.
Per category
Ordered by how little carries over.
| Category | Shapes published | Question pairs | Median shared of 10 | Fewest | Most |
|---|---|---|---|---|---|
| Fintech | 6 | 15 | 1.0 | 0 | 3 |
| Enterprise Software | 6 | 15 | 1.0 | 0 | 3 |
| HR & Talent | 5 | 10 | 1.0 | 0 | 2 |
| Healthcare Services | 6 | 15 | 1.0 | 0 | 2 |
| AI Security & Privacy | 6 | 15 | 2.0 | 1 | 4 |
| Education & Training | 6 | 15 | 2.0 | 1 | 5 |
| Consumer Products | 6 | 15 | 2.0 | 1 | 4 |
| Emergent Prosumer | 6 | 15 | 2.0 | 1 | 4 |
| Cybersecurity | 6 | 15 | 2.0 | 1 | 4 |
| Deep Tech & Hardware | 6 | 15 | 2.0 | 1 | 4 |
| Family Software | 6 | 15 | 2.0 | 1 | 7 |
| Consumer Health | 6 | 15 | 3.0 | 3 | 5 |
| AI Visibility & GEO | 6 | 15 | 3.0 | 2 | 7 |
| AI Infrastructure | 6 | 15 | 4.0 | 3 | 5 |
| Consumer Finance | 6 | 15 | 4.0 | 3 | 7 |
Fintech, Enterprise Software, HR & Talent and Healthcare Services each have at least one pair of questions whose top tens share nothing at all. At the other end, Consumer Finance and AI Infrastructure carry four of ten — still a minority of the list, and still six sources a brand would have to win separately.
What actually carries is a platform
Across the 220 pairs there are 536 shared slots, held by 102 distinct domains. Two domains hold 234 of them.
| Domain | Pairs it carries across | Source role |
|---|---|---|
| 139 | Community and social platform | |
| YouTube | 95 | Search or media platform |
| Arxiv.org | 20 | Academic and government source |
| Medium | 18 | Editorial publication |
| 18 | Community and social platform | |
| Healthline.com | 18 | Other observed source |
| TechRadar | 16 | Editorial publication |
| Nerdwallet.com | 15 | Editorial publication |
| Forbes | 11 | Editorial publication |
| Hubspot.com | 10 | Vendor-owned source |
Reddit and YouTube are 43.7% of every shared slot in the release. In 42 of the 220 pairs the shared set is only platforms — the two questions have no publication, retailer or vendor site in common at all.
Set the composition of the carried slots against the composition of the top tens they came from, and the size of the effect is visible in one column.
| Source role | Share of all 890 top-ten slots | Share of the 536 carried slots |
|---|---|---|
| Other observed source | 55.7% | 20.9% |
| Editorial publication | 12.8% | 15.3% |
| Community and social platform | 10.3% | 29.7% |
| Vendor-owned source | 9.9% | 9.7% |
| Search or media platform | 5.3% | 17.7% |
| Academic and government source | 3.4% | 5.4% |
| Market and company database | 1.9% | 1.3% |
| Analyst and consulting research | 0.7% | 0.0% |
Community and search platforms are 15.6% of the top-ten slots in this release and 47.4% of the slots that survive a change of question — three times their weight. Vendor-owned sources carry at exactly their own weight, 9.9% against 9.7%: a brand's own domain is no more durable across questions than the average source, it is simply present or absent question by question. Editorial publications get a slight lift, 12.8% to 15.3%, and analyst and consulting research carries nothing anywhere in the release.
One category in full
Consumer Products, all six questions, top three sources each.
| Buyer question | Segment denominator | Top three cited sources |
|---|---|---|
| Best tools | 161 runs / 8 dates | Forbes 30.4%, Reddit 23.6%, YouTube 22.4% |
| Top lists | 131 runs / 7 dates | Chewy.com 23.7%, Dogfoodadvisor.com 23.7%, Forbes 19.9% |
| Comparisons | 131 runs / 7 dates | YouTube 29.8%, Petmd.com 25.2%, Reddit 22.9% |
| How buyers choose | 137 runs / 7 dates | YouTube 24.8%, Reddit 23.4%, Livetinted.com 22.6% |
| Is it worth it | 136 runs / 7 dates | Reddit 27.9%, Daily-harvest.com 23.5%, YouTube 22.1% |
| Problem-first research | 132 runs / 7 dates | Reddit 22.0%, YouTube 15.2%, Dialavet.com 14.4% |
Pool the six top tens and 45 distinct domains appear. Five appear on both the shopping and the research side: Reddit, YouTube, Allure, Healthline.com and Petmd.com. Three of the fifteen pairs inside this category share exactly one source, and in each case that source is YouTube.
What this does not say
It does not measure citation rate. Two questions can share a domain that is cited in 30% of one and 11% of the other, and that counts as carried here. It does not distinguish rank 11 from rank 400: a source just outside a top ten reads the same as one absent from the category.
It does not isolate the question from its subject. Inside a category the release samples different specific products per shape — Consumer Products roundups run heavily on pet food while how-buyers-choose runs on skincare — so part of the divergence is subject drift, not the question wording alone. The family comparison does not remove that; what it shows is narrower and still useful, that sorting the questions into shopping and research families predicts nothing about which pairs diverge.
"Other observed source" is 55.7% of the top-ten slots here and is the release's unclassified layer, not a named class with a shared character. Our own AuthorityTech brief on the unclassified majority in source-role coverage sets out why a reader should not treat it as a finding about any class. The role figures above are shares of slots, not of citations.
Whether the same source list holds still over time is a separate question with a separate answer; the Index publishes citation stability by source role, and whether the kind of source that leads changes with the question is measured in source-class capture by category. This page measures neither. It measures how much of a specific list of ten survives a change of question, which is the number a placement list is built from.
What a brand's leaders do with it
Count the questions, not the categories. If a shopper reaches a brand through two questions, that is two source lists to win and about two sources of overlap to reuse. A programme that measures itself on one question and reports "we are cited in AI answers" is reporting on a sixth of its own surface.
The platform result is the part that pays for itself. A brand's presence on Reddit and YouTube is the only thing in this release that carries across questions at more than its own weight, and it carries at three times. Everything else — the publication placement, the retailer listing, the brand's own product page — is won and lost question by question.