1 of 3 Predicted AI Shopping Carriers Arrived

Seven releases ago we named hubspot.com, zapier.com and adyen.com as the next vendors likely to hold a top-10 AI citation slot in two shopping question shapes. On the 2026-09-26 Machine Relations Index release hubspot.com does. The other two are unchanged to the single observed run, because their segments were never observed again — which means our own falsifier was untestable as written.

Para Labs Research··
ai-visibilitybrand-visibilitycommercemeasurementmethodology
Para Labs research image for 1 of 3 Predicted AI Shopping Carriers Arrived

The mechanic measured here is prediction settlement: whether a falsifiable read we published against a future release can be settled on that release, and what the answer is. On 2026-09-19 Para Labs published a shopping-shape carrier cross-section that ended with three named vendor–category pairs and a condition. If decision-shape standing and shopping-shape carry-over travel together, hubspot.com in AI visibility, zapier.com in enterprise software and adyen.com in fintech are the next carriers; if they are not, the property is coincidence at n=32.

Seven releases later, on mri_score_v2.0+2026-09-26+2b779408cfda, hubspot.com carries. zapier.com and adyen.com do not, and both are unchanged to the individual observed run — same rank, same denominator, same cited runs, same observed runs as the figures we published on 2026-09-19.

That third fact is the one that matters more than the verdict. Two of the three predictions could not have come true in either direction, because the segments they live in were not observed again between the two releases. The falsifier we wrote named a release. It should have named a run date.

This is a cross-section read on two releases, not an experiment. It reports what six engines cited over a fixed 133-day window. It does not claim that decision-shape citation causes shopping-shape carry-over, and it says plainly which part of its own prior test design was wrong.

The measurement

The source is the public view of the Machine Relations Index, contract machine_relations_index_public_view_v2.0, methodology mri_score_v2.0, release mri_score_v2.0+2026-09-26+2b779408cfda. The machine-readable release carries every number below and hashes to SHA-256 2b779408cfda38bab1170b66162973372c7229e64eddb08cbbf019d8c34eba17.

Field 2026-09-19 release 2026-09-26 release
Window 2026-05-10 to 2026-09-19 2026-05-10 to 2026-09-26
Days observed 126 133
Answer runs 15,883 16,925
Citation events 125,115 132,514
Cited source domains 22,213 23,978
Eligible monitored queries 912 1,010
Categories publishing all six buyer shapes 13 14

A segment is one subject category paired with one buyer question shape. Six shapes are published: three shopping and selection shapes (best_x, top_list, x_vs_y) and three decision shapes (how_choose, is_x_worth, problem_first). A segment is published only after it clears the evidence floor of at least 10 observed runs across at least 7 distinct run dates. Every rank below is the release's own published rank inside a segment, read with its denominator, never re-derived by sorting.

The verdict on the three named vendors

Vendor, category Decision-shape slot that made it a candidate Shopping shapes held, 2026-09-19 Shopping shapes held, 2026-09-26 Carrier now
hubspot.com, AI visibility is_x_worth #5 of 467 x_vs_y #1 of 417 x_vs_y #1 of 417; top_list #6 of 558 Yes
zapier.com, enterprise software problem_first #2 of 258 best_x #9 of 167 best_x #9 of 167 No
adyen.com, fintech is_x_worth #3 of 206 top_list #9 of 200 top_list #9 of 200 No

One of three. The property named the right vendor in the one case where the instrument could answer, and that is a real confirmatory data point rather than a rhetorical one: hubspot.com was the only one of the three whose shopping segment was observed again.

Read the unchanged rows literally. zapier.com's enterprise-software best_x line is rank 9 of 167 on 11 cited runs out of 101 observed on both releases. adyen.com's fintech top_list line is rank 9 of 200 on 12 cited runs out of 102 observed on both. Not approximately the same. The same numbers, seven days and 1,042 new answer runs apart.

Why two of the three could not move

Those two segments have 7 distinct run dates each across the 133-day window, exactly as they did at 126 days. The segment hubspot.com entered has 8.

The Index publishes a segment once it clears the evidence floor and then observes it again only when a new run date lands in it. Machine Relations has already measured and published this: across two releases three days apart, all 84 comparable published segments were numerically identical, because a published buyer-decision segment is a fixed small-date sample while news segments are observed continuously. That research went live on 2026-09-21, two days after the prediction it invalidates as a seven-day test.

The same wall is visible in this cross-section. Of the 84 segments belonging to the 14 categories that publish all six shapes, 78 sit at exactly 7 distinct run dates and 6 sit at 8. Every figure that changed between the two releases in this read belongs to one of those six segments. Every figure that did not change belongs to one of the 78.

Segment Distinct run dates Observed runs, 2026-09-19 Observed runs, 2026-09-26
ai-visibility-geo top_list 8 71 101
cybersecurity best_x 8 75 105
enterprise-software best_x 7 101 101
fintech top_list 7 102 102
consumer-finance best_x 7 132 132
fintech x_vs_y 7 102 102
cybersecurity x_vs_y 7 107 107

A new run date adds about 30 observed runs to a segment. Seven days of release cadence added zero to five of the seven segments above.

What one new run date does to a top ten

The prior read gave us two published rows inside the one segment that moved, so the before and after are both on the record. ai-visibility-geo top_list went from 412 cited domains over 71 observed runs to 558 over 101, and its vendor-owned rows re-sorted.

Domain 2026-09-19 2026-09-26
semrush.com #7 of 412, 9 of 71 runs #3 of 558, 17 of 101 runs
therankmasters.com #5 of 412, 11 of 71 runs #4 of 558, 17 of 101 runs
hubspot.com outside the top ten #6 of 558, 15 of 101 runs

One new observation date moved a domain from outside the top ten to rank 6, moved another four places up, and grew the competing pool by 146 domains. This is the movement a brand team would attribute to a campaign if they read the two release dates and not the run dates.

The other segment that gained a date, cybersecurity best_x, moved less at the top: sentinelone.com held #2, cynet.com held #3, and paloaltonetworks.com went from #7 of 213 to #5 of 247. huntress.com now leads that segment at 56 cited runs of 105. Two segments is not a sample; what they show is that a new run date is the unit in which a leaderboard can change at all, and that the size of the change is not fixed.

The cross-section, re-measured

Filtering the top ten of each of the 42 shopping segments of the 14 qualifying categories to the vendor_owned class gives 33 vendor–category pairs over 29 domains: 24 hold exactly one shopping shape, 9 hold two, none holds all three.

Read 2026-09-19 release, 13 categories 2026-09-26 release, 14 categories 2026-09-26, 13 categories
Vendor-owned top-10 slots, best_x 9 of 130 10 of 140 9 of 130
Vendor-owned top-10 slots, top_list 12 of 130 14 of 140 13 of 130
Vendor-owned top-10 slots, x_vs_y 18 of 130 18 of 140 18 of 130
Vendor–category pairs 32 over 29 domains 33 over 29 domains 32 over 29 domains
Pairs holding exactly one shape 25 24 24
Pairs holding two shapes 7 9 8

The third column is a like-for-like read. Fourteen categories now publish all six shapes against 13 on 2026-09-19, and the Index does not archive prior release files, so which category joined cannot be read directly from the published data. It can be inferred: excluding healthcare-services from today's 14 reproduces the 2026-09-19 counts exactly on all four of best_x slots, x_vs_y slots, pair count and domain count. Of the thirteen other single exclusions, twelve match on two of the four or fewer. The one that matches three is consumer-finance, and it is ruled out by the prior read itself, which names experian.com in consumer finance as one of its seven carriers. On that basis the like-for-like change across seven releases is one pair: hubspot.com moved from single-shape to carrier. All seven original carriers held.

The property, and the second vendor it does not explain

The prior read found that a carrier's distinguishing feature is a top-10 slot in one of the same category's three decision shapes. Re-measured on this release it still separates the groups. On the like-for-like 13 categories it separates them slightly more sharply than before: 7 of 8 carriers hold a decision-shape top ten against 3 of 24 single-shape pairs, where the prior read had 6 of 7 against 3 of 25. The table below is the full 14-category read.

Decision-shape standing in the same category 9 carrier pairs 24 single-shape pairs
Top-10 slot in at least one decision shape 7 of 9 3 of 24
Top-10 slot in two or more decision shapes 3 of 9 0 of 24
Cited in all three decision shapes 7 of 9 8 of 24
Cited in none of the three 0 of 9 5 of 24
Cited runs across the three decision shapes, median per pair 28 5
Cited runs across all six buyer shapes, median per pair 80 29
Cited by all six engines in the window 9 of 9 19 of 24
Confidence grade 0 A, 6 B, 3 C 2 A, 8 B, 11 C, 3 collecting
Cited runs across the whole index, median 152 78.5

Grade and volume still run the wrong way: the only A-grade rows in the set are single-shape, and the largest domain in it is cited in 306 runs and holds one shape in each of two categories.

Two carriers do not fit the property, one more than last time. therankmasters.com remains the original exception: best_x #3 of 351 and top_list #4 of 558, with no decision-shape top ten and an x_vs_y rank of 378 of 417. The new one is guideflow.com in healthcare services, holding best_x #10 of 199 and top_list #6 of 223 while sitting at #82 of 214 in how_choose and #148 of 216 in is_x_worth, with problem_first not published in that category. It is also the pair that arrives with the category, so it is the weakest row in the table and is reported rather than dropped.

What a brand-side operator can take from this

  • Read run dates, not release dates. Every published segment in the Index exposes its distinct run-date count next to its rank and denominator. A rank that is identical across two releases has usually not been re-tested; a rank that moved has almost always had a new date land in it. Any dashboard that refreshes daily is making the same distinction, whether or not it shows you.
  • A week is not an interval over which your position can be judged. Two of three predictions here were unanswerable at seven days. Freeze a rank on a run date, not on a calendar date, and read the next one.
  • Carry-over is still rare and still does not come with size. Nine of 33 vendor pairs hold two shopping shapes, none holds three, and the biggest domain in the set holds one.
  • Decision-shape standing remains the best single predictor available, and it is now one for one where it could be tested. The one candidate whose segment was re-observed became a carrier. That is a sample of one, and it should be treated as such.
  • Write your own internal predictions against the measurement event. Preregistering an analysis is worth little if the trigger it names cannot occur in the stated interval. Ours named a release and the release is not the measurement.

Limits

The unit is 33 pairs and the carrier group is nine. A property that fits seven of nine at that size is a hypothesis with a good second read, not a result. Ranks are read from published segments only; a rate in a collecting segment is not a rank and was not used. The 2026-09-19 column throughout is the figures published in the prior cross-section, not a re-run on an archived release file, because prior releases are not retained at a public URL. The identity of the fourteenth category is an inference from count arithmetic and is labelled as one. Two segments gaining a run date is not a measurement of what a run date does in general. The Index observes six engines on a fixed English-language commercial basket and cannot see feed-driven product surfaces separately from text answers, so "shopping shape" here means the question text, not a shopping product. Nothing here measures whether any page, schema change or campaign moved a rank. Same-date runs may not be independent replications, which is a separate reason to read a small-date segment cautiously: the design-effect adjustment used for clustered observations in trials, and the sampling assumptions behind any binomial interval, both apply to a 7-date sample more sharply than to a 133-day one.

Sources and method

FAQ

What was predicted, exactly? That three named vendor–category pairs holding one shopping-shape top-10 slot and at least one decision-shape top-10 slot would be the next to hold two shopping shapes. The prior article stated the falsifier in the same sentence.

Is one of three a pass or a fail? Neither, and the honest answer is that the test as written could not run. Two of the three segments were not observed again, so those two rows carry no information about the property. The one row that could move, moved in the predicted direction.

Why does a daily release not produce daily numbers? A published buyer-decision segment is sampled on a small number of distinct run dates, not continuously. Between these two releases, 78 of the 84 segments in this read gained no new run date, so their ranks, denominators and cited-run counts are identical.

Does this weaken the Index? It is the Index's own published field that makes the check possible. run_dates sits beside every rank, and Machine Relations published the segment-cadence finding itself on 2026-09-21. A measure that tells you when it last looked is more usable than one that does not, not less.

How would I test this properly on my own domain? Read your domain's profile at machinerelations.ai/index/domains/<domain>, record the rank, denominator and run-date count for each of your category's six segments, and compare only against a release whose run-date count for that segment has increased.