AI assistants now hand buyers a shortlist before any vendor gets a word in. This program measures those shortlists the way they deserve to be measured: repeatedly, across every frontier model, with published methods and raw data. The reports tell you what moved. What it means for your company is a conversation.
Current readings from Wave 2 (July 2026, 8 models, 5 AI labs, 1,328 runs). Updated at every wave.
72%
Convergence Index
Average agreement between frontier models on top-five shortlists, across 25 categories. Higher means the AIs are reading from one map.
12 of 25
Contested categories
Categories where the models disagree on the number one vendor. The open territory in B2B software.
18%
Big-four concentration
Share of all 875 top-five shortlist slots held by Salesforce, Microsoft, HubSpot, and Adobe.
13%
Turnover per release
Share of shortlist slots that changed hands across one model release, with no market event in between.
SandsDX original data: identical buyer questions through Claude Fable 5 and its predecessor across 21 B2B SaaS categories, 258 runs. Three category leaders flipped, 13% of top-five places changed hands, and the models know about rebrands they refuse to use.
Read the reportSandsDX original data: identical buyer questions through 8 frontier models from 5 AI labs, 26 B2B SaaS categories, 1,328 runs. The frontier is converging on one map of B2B software, half the categories still have no agreed leader, and four companies hold 18% of every shortlist slot the models hand out.
Read the reportSandsDX original data: 8 frontier models, 26 B2B SaaS categories, five kinds of buyer questions, 8,188 runs. HubSpot replaces Microsoft at the top when the buyer is small, describing a problem instead of naming a category produces a different leaderboard entirely, and at renewal time the AI tells the buyer to shop around in at least 30 of 35 answers, for every vendor tested.
Read the reportWave 3 collected the baseline under a protocol published before collection began, including a hypothesis the data half-refuted, which we report. The next frontier model release triggers Wave 4: the identical instrument, read against the baseline, with the release's effect isolated for the first time. The protocol and its dated amendment are public.
Read the protocolFrozen questions, published verbatim
Every wave asks the same buyer questions, word for word, through the API with no system prompt and no web access except a search-grounded control. The exact prompts are printed in each report. Anyone can rerun the instrument.
Consensus, never a single answer
Every question runs five times per model and results are scored across runs. Our calibration study shows five runs recovers the same top five as twenty, while a single query misreads the noisiest categories a third of the time.
Corrections in public
Findings that fail a stricter retest get corrected in print, not quietly. The corrections log below is part of the method.
July 7, 2026
Wave 1 reported three category leaders losing the number one slot. Wave 2's retest showed those single-slot changes are sensitive to how many vendors the question asks for; the membership churn, naming gap, and maturity gradient all replicated. Both reports now state this, and crown changes are held to a multi-phrasing standard from Wave 3 forward.
Every wave's complete run data, the collection harness, the metric definitions, and the versioned company-mapping tables are maintained as a private archive. Researchers, journalists, and operators who want to verify a number can request access.
The index tells you the market moved. Where you stand in it, under which name, at which stage of the buyer's journey, and what to do between now and the next model release is the work.
Schedule a Conversation