What 2,095 AI answers say about who gets named for B2B software
A buyer asking an AI system for software receives more than a list of links. The answer can name a shortlist, describe the products and set out the requirements the buyer should use to choose between them. We publish research on each of those layers so that a software team can trace a statement back to the answer that made it and the evidence used to check it.
The September 2026 Absence Index records 2,095 valid answers across 14 B2B software categories. Its 286 vendor category records show which vendors the answers named, with the count, the valid answer base and a 95 percent interval. The release is fixed to its capture dates and has a permanent DOI, so a reader can distinguish a published observation from a later change.
Start with the question and the denominator
The release covers ChatGPT, Claude, Perplexity, Google AI Overviews and Google AI Mode. Each category has ten buyer questions and three scheduled runs per engine. That makes 150 scheduled observations per category. Completed answers and failures are recorded separately, so the valid denominator can be smaller than the scheduled panel. The release's 2,095 valid category answers come from 2,100 scheduled observations.
This distinction matters when comparing vendors. A named count describes how often a vendor appeared in those captured answers. It is useful only beside the matching base, question set, engines and dates. It does not establish market share or the quality of a product. A vendor can appear in more than one category, which is why the record count should not be silently converted into a count of unique companies.
The Index makes those details available at the point of comparison. A team can inspect a category, open a vendor record and follow the evidence behind a count. A reader can also return to the frozen release instead of treating the latest page as though it were the original observation. That makes a disagreement specific: a particular answer, a particular vendor match or a particular source can be examined.
Being named and being described accurately are separate
An appearance in an answer is one question. The facts stated about that vendor are another. Wrong About You examines factual statements about 70 vendors selected from the same September release and classes them against the vendors' own public pages.
The published version contains 2,922 engine statements: 1,924 classed true, 113 wrong, five stale and 880 unverifiable. Among the 2,042 verifiable engine statements, 118 were wrong or stale: 5.8 percent, with a descriptive 95 percent Wilson interval of 4.8 to 6.9 percent. Unverifiable statements are reported separately; they are not counted as correct or incorrect in that rate.
The useful part is the connection between the classification and the evidence. A flagged statement carries its engine, run, question, capture date and the vendor page used to check it. That lets a company inspect the exact wording and the dated source, rather than respond to an unattributed claim about what AI supposedly says. The published team review remains in progress. Any changes accepted through that review belong in a dated, traceable revision.
The checklist can shape the shortlist
Before naming a product, an answer may tell the buyer what to require. The AI Written RFP studies that earlier part of the answer: the buying criteria stated before the first vendor name. It uses the frozen September answers rather than collecting a different set of responses for the comparison.
The study reports 2,003 distinct category criteria across 2,093 eligible source answers, after its documented exclusions. Those criteria are supported by 10,547 answer spans. The published data also records 2,417 explicit vendor association passages and 10,015 dated page comparisons across 70 mapped vendor pages. These are different units: a criterion, a supporting passage and a page comparison are not interchangeable counts.
A criterion frequency counts distinct eligible answers that state that requirement. Several passages in one answer still count as one answer for that frequency. The category's actual eligible base stays beside the count and interval. The study also distinguishes an explicit vendor association from a general list or an unsupported implication. That gives a software team a way to examine the requirements around its category without treating every nearby product name as evidence of a capability.
The page comparisons are equally specific. They describe what was stated on the exact dated page that was read. A feature not stated there is not proof that the product lacks it. Partial scope remains unknown. Each vendor comparison sits beside the corresponding median across the five mapped vendors in its category. Independent analyst review of this study remains in progress in the published version.
Make evidence useful to the next person
The open AI Source Standard v1 provides a common reference for recording source evidence. Together, the standard and the studies support a practical sequence: establish the observed shortlist, inspect the facts stated about a company, then examine the requirements that appeared before that shortlist.
For a marketing or product team, that sequence can turn a vague visibility concern into a reviewable piece of work. Start with the dated record. Identify the statement or requirement that matters. Read its evidence. Decide which company facts or pages need clarification. Keep the original observation available when a correction is made, and compare later observations on their own stated dates and bases.
The studies preserve the difference between an observation and a commercial result. They do not establish that an individual answer caused a sale, or that a change to a page will produce a particular future answer. Their value is a shared record that a buyer, vendor or researcher can inspect. Counts become more useful when the underlying questions, exclusions, sources and review state remain visible with them.
Every figure comes from the published method, v1.1. The cited releases are published under CC BY 4.0 with attribution and source references. Any named company may request one no-cost re-run at index@broadcastwell.com; both results are published.
Read the accompanying study pages for the Absence Index, Wrong About You and The AI Written RFP, and the AI Source Standard. Each connects the published result to its method and source material.