I Would Not Budget Off One HR Shortlist

I would monitor the first HR buyer-choice shortlist, name its evidence floor, and hold any budget shift until the same sources repeat. In the September 21, 2026 Machine Relations Index release, the HR and talent "how buyers choose" segment cleared the publication floor for the first time: 101 observed answer runs across seven run dates, and 235 cited source domains. Rank one was cited in 14 of those 101 runs. Rank ten was cited in eight. Ranks seven through thirteen were separated by nothing on cited-run count. A six-run gap on a 101-run sample is a finding I can cite. It is a thin base for a category rule.

The reading of why that category is still unfinished is One HR AI-Citation List Is Live, and It Isn't Finished. That page is the deeper resource. This essay is how I would use it as a founder: what I would watch, what I would refuse to fund, and the repeat test that would turn the list into a budget signal.

What I would decide, and what would reverse it

Field	What I would do
Decision	Keep the first published buyer shortlist in a monitoring file. Hold category-wide spend until the pattern repeats.
Evidence	HR and talent, how-buyers-choose, cleared publication at 101 observed runs over seven dates (September 21 release, unchanged on September 25). Rank one has 14 cited runs. Rank ten has eight. At least six of the buyer top ten are cited by ChatGPT, Claude, Perplexity, and Gemini alike. That is a floor: Gemini has intermittent blind days (22 through the September 18 window, a count from the collection ledger that the release file does not carry), a recorded citation can only be real, up to four of the ten are not in that four-engine set, and three of those have fewer than fifteen citations each in the entire release.
Scope	First-published buyer shortlists, especially where only one buyer-question shape has cleared the evidence floor.
Action	Read the ranked sources. Do not move spend until a source repeats in at least two further buyer shapes, or the same rank band survives a later release.
Reversal	A source that holds a top-three position in three buyer shapes, across three releases, and is cited by at least three of the engines whose collection actually covers the window, can move from the monitoring file into the budget.

I would rather miss an early spike than let one thin leaderboard become the plan. The cost I am protecting is leadership attention. A page is cheap. Pointing the company at the wrong proof standard is not.

The list, read as a list

The segment is HR and talent paired with the how-choose question shape. The Index route reports 101 observed runs across seven dates in release mri_score_v2.0+2026-09-21+d73e239e7c5b, and the same 101 runs over the same seven dates in the September 25 release. The order below is stable across both.

Rank	Source	Cited runs	Segment rate	How I would label the source
1	Pin	14 of 101	13.86%	Vendor-owned
2	LinkedIn	13 of 101	12.87%	Community and social
3	Reddit	9 of 101	8.91%	Community and social
4	Selenios	9 of 101	8.91%	Other observed source
5	Umantis	9 of 101	8.91%	Other observed source
6	US Tech Automations	9 of 101	8.91%	Other observed source
7	ClearCompany	8 of 101	7.92%	Other observed source
8	CVViZ	8 of 101	7.92%	Other observed source
9	Gem	8 of 101	7.92%	Vendor-owned
10	Lever	8 of 101	7.92%	Other observed source

Pin at rank one appears in 14 of 101 runs. Lever at rank ten appears in eight. People Managing People sits just outside the top ten at rank eleven with the same eight cited runs as rank ten. If one citation moved on the next release, the story a slide would tell could look cleaner than the evidence. I would brief the near-tie as a band, ranks seven through thirteen, and I would forbid a sentence that calls rank ten a different tier from rank eleven.

This is one question shape. It says how engines cite sources when someone asks how buyers choose HR software. On September 21 it did not yet say how they cite best-tool questions, comparison questions, problem-first questions, top-list questions, or whether-the-tool-is-worth-it questions. On the September 21 release, one buyer shape had cleared the floor and five were still collecting. By the September 25 release three more had cleared (best-tool, worth-it, problem-first) and two (comparison, top-list) were still collecting. I would not have generalized from the shape that happened to publish first.

Engine reach is uneven, and I would brief it with the collection gaps attached. A whole-window reading of the same top ten, on the September 25 release's 132-day window, found four of the ten missing from at least one of the three engines the Index collected continuously through the window, ChatGPT, Claude, and Perplexity. ChatGPT misses four of the ten, Claude misses one, and Perplexity is present on every one. Those counts are exact for those three engines. Gemini was not observed on three of the ten, which is an upper bound rather than a count, because Gemini has intermittent blind days inside the window (22 through September 18, per the collection ledger) and a blind day can hide a citation it cannot invent. The three narrowest domains, Selenios, Umantis, and CVViZ, were observed cited only by Claude and Perplexity, and each has fewer than fifteen citations in the whole release. Their narrowness is thin evidence as much as it is an engine preference.

The two Google surfaces stay out of that engine-reach fraction for a reason that belongs to the collection. In the September 25 release's 132-day window, the Index's collection record shows nothing from Google AI Overviews on 41 days (August 15 to September 24) and nothing from Google AI Mode on 11 (September 14 to 24). The release file carries no per-engine daily counts, so those two figures come from the collection ledger. An engine that was not collected cannot be shown to have cited nothing. I would say that sentence in any meeting where someone wants to claim a source is "missing from Google."

What the list is allowed to justify

A founder needs to know which verb the list supports. This list supports "watch" and "inspect the source types." It does not support "reallocate the category budget toward whoever is rank one."

If a vendor page is rank one on a newly published shape, the question I would ask is whether that source type shows up when the question changes. A page can fit one prompt pattern, one engine mix, one moment of category maturity, and one thin window. Copying the page is how a local artifact becomes a roadmap.

I would keep three verbs separate. A one-shape rank says watch. A repeated rank says study. A repeated rank across engines and releases says allocate. Mixing the verbs is the failure mode.

The stop rule I would write into the planning doc has four tests, and a source has to pass all four before anyone calls it a category pattern.

1. Shape. It appears in the top three on at least three buyer question shapes.
2. Release. It holds the same rank band across at least three releases, so the finding survives new runs.
3. Engines. It is cited by at least three of the engines whose collection covers the full window.
4. Coverage. Before counting engines, state which engines actually ran through the window. An engine with missing days can show presence and cannot show absence, so it belongs in neither the numerator nor the denominator of an engine-reach test.

The HR buyer list could not pass that rule when it published, because only one buyer shape had cleared the floor. On the September 25 release four had, and no source is in the top three of three of them (Pin and LinkedIn each reach two), and no source has held a band across three releases. Publishing the boundary is the point of having a floor. I would rather explain a boundary than invent a maturity the release does not claim.

What I would do if we sell into this category

I would read source types before ranks. The top ten contains two community platforms, LinkedIn and Reddit, two vendor-owned sources, Pin and Gem, and six other observed sources. The answer is assembling a selection layer from decision-support surfaces, not from HR trade press alone. A pitch list made only of trade magazines would be aiming beside the observed layer.

I would separate buyer answers from news answers. The HR news segment is thicker: 608 observed runs across 53 run dates, the same on both releases. At least nine of its top ten are cited by ChatGPT, Claude, Perplexity, and Gemini alike, against at least six of ten in the buyer list. Both figures are floors, for the same Gemini blind-day reason, and the comparison holds because the same blind days sit under both. The news layer is more mature than the buyer-choice layer. Maturity is not the same as value. I would not move buyer budget onto news outlets because the news sample is larger. I would use the larger sample as a warning that the buyer sample is earlier.

I would treat ranks seven through thirteen as one evidence band. The edge of a top ten is a cutoff in a table. It is not a cliff in the market.

I would read the next buyer shapes as they publish. Best-tool, comparison, problem-first, top-list, and worth-it questions are the test of whether Pin, LinkedIn, Reddit, and the vendor pages repeat. On September 25, best-tool, worth-it, and problem-first had published, and the repeat was partial: Pin is rank one in best-tool (32 of 107 runs) and LinkedIn is rank two in problem-first (16 of 101), while Reddit is not in the top three of any other published shape, and the worth-it top three (PerformYard, Lattice, People Managing People) shares no source with the how-choose top ten. The first list was a starting line, and the later lists only partly repeat it.

On September 25, 2026 I would also keep the correction that belongs in any briefing of this release. An earlier reading said four of the top ten were visible across all six engines, that three were visible across only two, and that Google AI Overviews was missing five of the ten. That reading is withdrawn. Collection had failed on both Google surfaces inside the window, no AI Overviews citations on 41 of 132 days and none from AI Mode on 11, and a whole-window presence flag only accumulates. Missing days can remove a domain from an engine's set and cannot add one. The engine that looked narrowest was the engine measured least. Corrected, on the engines whose collection covers the window, six of the top ten are cited by all four of ChatGPT, Claude, Perplexity, and Gemini, not four of ten across six engines. The decision is unchanged. The fourth test, collection coverage before engine counts, is what that mistake earned.

The habit I want from this category, a public count with a stop rule, is the one I describe on Machine Relations. The term for the rate I would show beside a rank, so a 13.86% is never left naked, is citation rate: cited runs divided by observed runs, which is what 14 of 101 is.

The decision

I would cite the shortlist. I would not obey it. The monitoring file gets the ten domains, the 101-run denominator, the seven dates, the release id, and the engine-coverage notes. The budget file gets nothing from this segment until a source passes the four tests.

Scope: first-published buyer shortlists in a category where other buyer shapes are still collecting. The HR and talent how-choose segment is the example as it stood on the September 21 release; by September 25 three more buyer shapes had published, so it is now a worked example rather than a live one. The rule is the same for the next category that publishes one shape and calls it the market. It does not apply to a source that has already passed the four tests, and it does not apply to the HR news segment's 608 runs, which I would read as a more mature news layer rather than as a buyer map.

What would change my mind: a source in the top three of three buyer shapes, across three releases, cited by at least three engines that were actually collected across the window. That is the reversal written into the table above. I would also reopen the Google surfaces if a later release filled the 41 and 11 missing days and the presence counts changed, because the withdrawn reading failed exactly by treating missing collection as a property of the engine. I would not move spend because a competitor's slide declared a winner from rank one.

FAQ

Is rank one on this list a category pattern?

Pin leads with 14 cited runs out of 101, 13.86%, in release mri_score_v2.0+2026-09-21+d73e239e7c5b, stable through the September 25 release. Lever, at rank ten, has 8 runs, 7.92%, and rank eleven ties that count. On the September 21 release one question shape had cleared the floor. On September 25 four had, and Pin is rank one in two of them, how-choose and best-tool. That is a repeat, short of the three-shape, three-release test. I would call Pin the leader of those shapes. I would call a category pattern something that repeats on more shapes and later releases.

Why leave Google out of the engine-reach count?

Because the collection record shows nothing from Google AI Overviews on 41 of the September 25 release's 132 days and nothing from Google AI Mode on 11. A gap in collection can hide a citation. It cannot prove an engine never cited the domain. Gemini's intermittent blind days (22 through September 18, per the collection ledger) are why I describe the four-engine overlap as a floor: at least six of the buyer top ten, and the Gemini misses are an upper bound.

What should an HR vendor do with the list this month?

Inspect source types. Two of the top ten are community platforms, two are vendor-owned, six are other observed sources. Compare that mixture with the news segment, which has 608 runs over 53 dates and at least nine of ten sources cited by all four of ChatGPT, Claude, Perplexity, and Gemini. Watch the seven-through-thirteen band as one band. Read the buyer shapes that have published since (best-tool, worth-it, problem-first) before changing a pitch list or a content budget.

What mistake is already on the record, so we do not repeat it?

The September 25 correction. A whole-window flag made Google look like the narrow engine because Google was the engine we had failed to collect, 41 days and 11 days. The corrected statement is that six of the top ten are cited by ChatGPT, Claude, Perplexity, and Gemini. The stop rule gained a coverage test because of that error. I would rather amend a public count than leave a briefing that treats our blind days as the market's structure.