The Agent Gap: can an AI buying agent find, price and reach B2B software? 70 vendors, 3 tasks, one agent, October 2026

Abstract

We tested 70 selected B2B software vendors across 14 categories with ChatGPT Work with Cloud browser. The displayed product and mode are preserved in each receipt. Each vendor received three fixed tasks in a fresh conversation: describe the vendor, read its pricing page, and reach its empty demo or sales form. All 210 of 210 primary sessions were captured. Of 209 scored sessions, 159 passed and 50 failed. The pass rate was 76.1% (95 percent Wilson interval 69.9% to 81.4%). One of 210 primary sessions, Pipedrive S2, has unknown currency evidence and is unscored. Unknown is neither a pass nor a failure.

Found was yes for 210 of 210 sessions across 70 vendors. Pricing-task facts were correct in 67 of 69 evaluable S2 answers; this includes accurate no-public-price and access-limitation reports, not only numerical prices. Pricing destinations were reached in 52 of 70 S2 sessions. Demo or sales forms were reached in 44 of 70 S3 sessions. The required destination was not reached in 47 of 210 primary sessions.

The preselected repeat sample contains ten vendors and 30 task pairs. 30 of 30 repeats are captured, with 28 scored and 2 unknown. Complete Found, Correct and Reached triplets agree in 27 of 28 evaluable pairs among 30 captured pairs. 2 captured pairs include an unknown score and cannot establish complete-triplet agreement.

Selection and method

The frozen set contains the five most-named vendors in each category of Absence Index release 2026-09, DOI 10.5281/zenodo.22907695. Ties use vendor-domain cited count, then alphabetical order. The ten repeat vendors were selected before the first run by sorting category and vendor and taking positions 0, 7, 14, 21, 28, 35, 42, 49, 56 and 63. No vendor, task or scoring rule was substituted after observing results. The owner amended the study to one agent on 6 October; the original protocol, amendment and excluded evidence are retained separately.

The study is single-agent and single-run: one primary run per vendor and task. Repeat runs measure agreement separately; they do not replace, average or repair primary results. Every repeat began at least two hours after its corresponding primary observed end. Observed end times are upper bounds where collection occurred after completion; the product's displayed duration is retained in the receipts. The same exact filled prompt was reused for each repeat.

Perplexity Computer ran five sessions on two vendors before its seat showed no credits; they are not part of the study figures. The preserved receipts identify Procore (three sessions) and Buildertrend (two). They are excluded from every primary, repeat, category, pass-rate and agreement denominator.

ChatGPT sometimes used public content retrieval or web search instead of its Cloud browser. Receipts distinguish those operations from a rendered browser visit. A content fetch is not called a browser rendering. Results describe this product, mode, prompt set and day, not every future run.

Tasks and scoring

Each prompt began with: Do not type into or submit any form on any vendor site. Do not enter personal details, create an account, buy anything, accept terms, send a message, sign in, solve a challenge or request payment. When a demo or contact form loads, stop with all fields empty. Browse only public pages. Report the exact URLs and any block you actually encounter.

S1: Find the vendor <Vendor> (<domain>). From its own site, tell me what it is, who it is for and what it costs. Give the URLs you read.

S2: Open the pricing page on <domain> and list the plans and prices shown. Tell me the URL and whether anything stopped you from reading it.

S3: Get me to the page on <domain> where I can request a demo or contact sales. Open the form and stop. Do not fill in or submit anything. Tell me the URL and what the form asks for.

The original score definitions, verbatim: Found: the agent identified <Vendor> as the vendor in question. Correct: the agent stated no false fact about <Vendor> when checked against <Vendor>'s own site read the same day. Reached: the page the task needed loaded for the agent with no block, challenge page or empty render. Position: where <Vendor> stood in any shortlist or comparison the agent produced.

The study's fixed single-run adaptation passes a session only when Found, Correct and Reached are all yes. Source checks use the same UTC day. Position is recorded separately and is not a ranking score. A correct statement that no public price is displayed can pass. Unknown evidence remains unscored. The commercial Agent Test's two-run rule is not used to aggregate this study.

Results by task

Wilson intervals quantify binomial uncertainty for the observed scored sessions. They do not turn the purposively selected set into a random sample of the software market. Tasks share vendors, and outputs share one agent and one day; the pooled interval does not account for that dependence. Unknown evidence is excluded from the scored denominator and explicitly reported in the captured denominator.

Results across 14 categories

Vendor results

Each cell is the single ChatGPT Work primary result. Pass requires Found, Correct and Reached. Fail means at least one is no. Unknown is unscored. Every category contains five selected vendors.

construction project management

dealership management systems

dental practice management

field service management

hotel property management

insurance agency management

legal practice management

loan origination software

radiology medical imaging

residential property management

transportation management systems

veterinary practice management

crm

project management

Obstacles and factual checks

Terminal groups are mutually exclusive: a confirmed empty destination takes precedence over a later timeout. Other includes missing destinations, HTTP refusals and tool failures.

The required destination was not reached in 47 of 210 primary sessions. Reasons below are the recorded terminal obstacles; a 403 alone does not establish an edge challenge or a login wall. Recovered errors remain in receipts. Factual failures and pre-execution seat gates are distinct.

Pipedrive S2 reported EUR 14, 39, 59 and 79 per seat per month with annual billing, totals 168, 468, 708 and 948, VAT extra. The vendor's English page and German locale selected through its visible language control served USD at the same figures on 6 October. Those observations neither verify nor refute the EUR claim. Correct is unknown, the session is unscored, and its repeat pair has no complete-triplet agreement classification.

The Pipedrive S1 and S2 repeats also reported EUR figures. The same unknown-evidence rule is applied to those repeat answers; identical unverified numbers are not proof of factual correctness or verified agreement. The two affected task pairs are listed as unknown. The primary S1 answer was in USD and was independently verified, so that original score is retained.

Dentrix Ascend S2 claimed its Packages and Suites page named no tiers; the same-day comparison table named Ascend Essentials, Ascend Pro and Ascend Accelerate. AppFolio S1 and S2 broadly claimed no dollar prices were published; the same-day rendered pricing page displayed Core, Plus and Max starting at $1.49, $3.20 and $5.00 per unit monthly. These three of 210 primary sessions are factual failures. A regional or rendering difference alone was not treated as a false fact.

EZLynx S3 had a pre-execution unusual-activity seat event. The owner-requested fresh retry completed and is counted once, with its actual form-render timeout. A Sectra pre-navigation approval-service 429 and other pre-execution fetch failures are separately retained and excluded. No study task was paid for, no challenge was solved, and no vendor form was typed into or submitted. In three form observations, masked phone-field displays could not prove literal emptiness; receipts retain that uncertainty rather than claiming it was visually established.

Repeat agreement

Across all 30 captured pairs: 27 agree, 1 disagree, and 2 are unknown. Each count uses 30 captured pairs as its base; the planned base is 30.

Complete-triplet agreement is 27 of 28 evaluable pairs (96.4%; 95 percent Wilson interval 82.3% to 99.4%). All 30 of 30 captured pairs are shown, including 2 pairs with an unknown score. Agreement over all 30 pairs cannot be assigned a point value while those comparisons are unknown; no missing comparison is counted as agreement or disagreement.

What the study does not prove

This records one agent product on stated dates with fixed prompts. It is not an agent comparison, a vendor ranking, an endorsement, or a judgment of software quality. It does not measure buyer adoption or promise a shortlist, position, inclusion, lead or revenue. Selection of highly named Index vendors is not random sampling of all software companies. A reached page does not prove an entire site is accessible. Answers, website content, localization, model behavior and tool reliability can change.

The domain-only S2 and S3 prompts can reach another product family. Oracle S2 returned OCI examples, SAP primary S2 returned application-package categories and its repeat reached the S4HANA Cloud ERP pricing section, and IDEXX S2 returned Petly Plans. DaySmart S2 returned four product families, including Vet. Scoring follows the literal submitted prompt. These outcomes do not establish that Oracle Transportation Management, SAP Transportation Management, Cornerstone or Neo prices were found. Prompts and results were not changed after this limitation became visible.

Dispute route

Any tested vendor may request one re-run of its three tasks at no cost by writing to index@broadcastwell.com. Both original and re-run results will be published, with dates and reasons. Original receipts are retained. No vendor is ranked.

Data, dates and reuse

The CSV contains all 210 primary session rows and all 30 repeat rows. Blank Correct values for Pipedrive primary S2 and repeat S1/S2 mean unknown evidence, not zero or no. Receipt and screenshot columns are local evidence paths, not a claim that private conversation files are hosted publicly. The frozen selection and repeat files and the dated protocol amendment accompany the dataset. License: Creative Commons Attribution 4.0 International, https://creativecommons.org/licenses/by/4.0/. Attribute Broadcastwell and retain the dates and limitations.

Vendor-set SHA256: a11035069a3cd90fbcff98de56c28cdef82561ec1ec08b633723e044ad041299.

Repeat-set SHA256: 4e4d6ab02ad56a1ff8e2522eb6dad32719f4719a8d1b0a683844cdd9fb0c6e40.

Study DOI: https://doi.org/10.5281/zenodo.23197560

Archive correction: release v1.0.1 preserves the original CSV bytes and both frozen hashes. Git normalized line endings in the first v1.0 archive, while all parsed rows, selections and reported figures were identical. Both dated releases remain available; this report cites the corrected archive.

Method notes

The study's factual rule is unchanged: Correct requires no false vendor fact against the vendor's own page read the same UTC day. A price omission is not automatically a false fact; a broad false claim that no price or tier is published is. Missing source verification remains unknown. Found, Correct and Reached are separately retained, and a reported access failure does not become a factual failure merely because the page worked for the reviewer.

The safety instruction and exact filled task were read back before Send. Fresh conversations were used, visible activity was retained without reconstructing hidden activity, and repeat prompts were checked against primary receipts. Reviewer source retrieval is independent of the measured agent's tool and is labeled as such. No rerun replaced an adverse primary result. Pre-execution operator or seat events are retained outside the vendor-session denominator.

The publisher sells measurement and remediation services in this market. These dated findings are not peer reviewed, do not rank vendors, and are separate from the Index release's named or cited counts. A vendor may use the dispute route regardless of any commercial relationship.