When an SMB buyer asks an AI assistant which global payroll software to buy, a shortlist comes back. We asked 44 different phrasings of that question across 2 engines, 88 answers in total, and counted every product named. This is the result. Nobody paid to be on it, and no vendor list was written in advance — the 30 vendors below are simply everyone the engines named.
Free to read, raw data included.
The board is re-measured every month. For $9.99 a month you get one email when your own row moves — your rank and answer count, who took the mentions you were absent from, and which of the 44 questions started or stopped naming you. Leave your address and I reply by hand with what your row reads today. What is in it, and the $29.99 one-off report →
sonar) searches the web while it answers and returned the sources used beside every one of those 44 answers — that column is what gets retrieved. Gemini (gemini-flash-latest) was called through its API with no tools and the near-empty system prompt quoted in the method below, so it answers out of the model’s own weights — not one of those 44 answers came back with a source. That column is what the model already remembers. Both are worth counting and they are not the same measurement: a product can be named often by what the models remember and rarely by what they retrieve, or the reverse, and the totals in the Answers column add the two together.“Answers” is how many of the 88 answers named that product.
Share is the same number as a percentage. The two right-hand
columns split it by engine.
Every row has its own link — the # beside a vendor name gives you
something like connexion.me/c/payrollglobal/#deel.
Click a vendor's name to see which of the 44 questions named them, and
click one of those to read the answer it was counted from, word for word.
| # | Vendor | Answers | Share | Gemini | Perplexity | |
|---|---|---|---|---|---|---|
| 1 | Deel# | 77 | 87.5% | 42 | 35 | |
| 2 | Remote# | 72 | 81.8% | 38 | 34 | |
| 3 | Rippling# | 63 | 71.6% | 41 | 22 | |
| 4 | Papaya Global# | 38 | 43.2% | 19 | 19 | |
| 5 | Oyster# | 30 | 34.1% | 18 | 12 | |
| 6 | Multiplier# | 19 | 21.6% | 7 | 12 | |
| 7 | ADP# | 11 | 12.5% | 4 | 7 | |
| 8 | RemoFirst# | 6 | 6.8% | 1 | 5 | |
| 9 | WorkMotion# | 6 | 6.8% | 0 | 6 | |
| 10 | Plane# Named together with Pilot in 1 of the 5 answers that name either. Gemini’s answer to question 18 writes the two as one product renamed: “**Plane (formerly Pilot.co)**”. Counted as separate rows — how names are counted. | 4 | 4.5% | 1 | 3 | |
| 11 | CloudPay# | 3 | 3.4% | 1 | 2 | |
| 12 | HiBob# | 3 | 3.4% | 0 | 3 | |
| 13 | Lano# | 3 | 3.4% | 0 | 3 | |
| 14 | Native Teams# | 2 | 2.3% | 0 | 2 | |
| 15 | One Global Payroll# | 2 | 2.3% | 0 | 2 | |
| 16 | Pilot# Named together with Plane in 1 of the 5 answers that name either. Gemini’s answer to question 18 writes the two as one product renamed: “**Plane (formerly Pilot.co)**”. Counted as separate rows — how names are counted. | 2 | 2.3% | 2 | 0 | |
| 17 | Playroll# | 2 | 2.3% | 0 | 2 | |
| 18 | Velocity Global# | 2 | 2.3% | 1 | 1 | |
| 19 | Alcor# | 1 | 1.1% | 0 | 1 | |
| 20 | Celergo# | 1 | 1.1% | 1 | 0 | |
| 21 | Dayforce# | 1 | 1.1% | 0 | 1 | |
| 22 | G-P# | 1 | 1.1% | 0 | 1 | |
| 23 | Gusto# | 1 | 1.1% | 0 | 1 | |
| 24 | Howdy# | 1 | 1.1% | 0 | 1 | |
| 25 | RemotePass# | 1 | 1.1% | 0 | 1 | |
| 26 | Safeguard Global# | 1 | 1.1% | 0 | 1 | |
| 27 | Skuad# | 1 | 1.1% | 1 | 0 | |
| 28 | Thera# | 1 | 1.1% | 0 | 1 | |
| 29 | Vistra# | 1 | 1.1% | 0 | 1 | |
| 30 | Workday# | 1 | 1.1% | 0 | 1 |
Everything above is free and stays free — the leaderboard, the method, all 88 raw answers and the machine-readable ranking, published in full at the bottom of this page. No vendor paid to be listed, no vendor can pay to move, and nothing here goes behind a login later.
What is for sale is one vendor's row, watched: $9.99 a month to be told when your rank moves, who took your mentions, and which phrasings started or stopped naming you — the board is re-run monthly and you get what changed. A one-off $29.99 deep report on a single row is the other option. What is in each →
Every claim on this page is reproducible from the files linked at the bottom.
Gemini (gemini-flash-latest), Perplexity (sonar). Each was called through its public API with a near-empty system prompt —
“answer the user's question directly and concisely” — and nothing else. The engines were
not asked to produce a list, rank anything, or name vendors. Steering them toward a list
would have manufactured the thing being measured.
44 distinct phrasings of one purchase decision, written before any measurement was run. No vendor name appears in any of them.
Each of the 88 answers was read back and every named product extracted. The vendor list is discovered: a product appears here if and only if an engine named it. Counted are global payroll platforms, Employer of Record (EOR) services, multi-country payroll systems, and broader HR suites whose global payroll or EOR module was named. Excluded, even when named: domestic-only payroll software, US-only PEOs, corporate expense management tools, cross-border payment processors, immigration-only agencies, contractor-only invoicing tools, accounting software, and companies named only as an example employer.
Spelling variants of one product are folded together, and a trailing
“by company” is dropped. This run collapsed 24 spellings onto 10 of the rows below: “Deel”, “Deel Payroll” and “Deel Global Payroll” are one row — the one the table prints as Deel. An edition name is folded into its product when
it is that product’s name followed by a further word, and the plain name is at least three
times as common. This run folded 6 such names, among them “ADP GlobalView” counts as ADP and “RemoFirst/Remote People” counts as RemoFirst. remote is exempt: it is an umbrella name in this category rather than a product, so a longer name that starts with it keeps its own row however common the short one is. A name that merely begins with the same
letters is a different product and keeps its own row. Three further pairs were merged by hand, and they are the only place a human judgement touches the list — it merges duplicates, it never adds or removes a vendor:
pilotco counted as pilotpebl counted as velocitypeblvelocity counted as velocityBeyond those folds a product is counted as the engines spelled it, and two rows in the table are never merged into one. When a product has been renamed or acquired and the engines still answer with both names, both names keep their own row: this page carries no hand-maintained list of who is now who, because that list would be a human claim about a mechanical count and nothing here could check it. What it has instead is the answers. Where an answer itself writes two of the rows below as one product renamed — the construction “X (formerly Y)” — the line beside those rows quotes that sentence and gives the overlap and the total, so the union is on the page. Nothing else produces that line. Two rows being named in the same answers does not, however often: the engines answer in shortlists, so the leaders of a category are named together whatever they are called. One pair of rows below carries that line in this run: Plane and Pilot. What each one quotes is what an engine wrote, not a judgement of ours about who owns what.
A product is counted once per answer, however many times that answer mentions it.
A leaderboard whose top ten reshuffles between runs would not be worth publishing, so this check was run before the page was, and the page ships only because it passed.
This measures what two answer engines said on 11 August 2026, in response to these questions, through their APIs. It is not a quality ranking, not a review, and not advice on which global payroll software to buy. A product can be excellent and rank low here — that gap is the point.
All 88 answers, verbatim, one JSON object per line: run A · run B. The per-answer extracted product names: run A · run B. Machine-readable ranking: ranking.json.
What Perplexity was reading. It is the only one of these engines that returns a source
list, and the markers in its answers above — [1], [2] — point into
it. Every URL it cited, with the title, date and the excerpt it actually retrieved, one JSON
object per answer: run A ·
run B. Join on index to the
answers file, and read [n] as citations[n-1]. Sources came back on
44 of its 44 run A answers; ChatGPT and Gemini return none, so this
file is Perplexity only.
The answer a vendor panel shows you is read straight out of the run A file linked above — the page fetches the one line it needs and displays it unaltered, so what you read there and what you get by downloading the file are the same bytes.
What one line of each file holds, and the terms you may re-use them under — the raw answers may be re-used in any medium, commercially and inside open-source software, with attribution.