Original research · Open data · Version 1.0

The 2026 B2B AI Citation Benchmark: full dataset

Every observation behind the report. Thirty B2B companies, eight fixed prompts, three AI answer engines, 720 coded observations collected on 18 and 19 August 2026. Nothing was excluded, re-run or re-scored after collection.

What is in it

Ten sheets. The data dictionary defines every column; read it first.

Read me

Provenance, licence, citation, and seven live integrity checks.

Data dictionary

Every column in every sheet, defined, including the coding rule.

Prompt tests

The core table. 720 rows, one per company × prompt × engine, each with the response excerpt it was coded from.

Discovery runs

The 54 independent discovery runs collapsed to one row each, with the 18 permanent Perplexity URLs.

Companies

The 30-company sample: category, use case, city, website, and exactly how each was selected.

AI Confidence Scores

The 17 August audit: overall score and five layer scores per company, with the gap notes.

Headline figures

Every published number, recomputed live from the Prompt tests sheet. Change a row and these move.

Concentration

The Finding 6 working: 191 firm-slots naming 133 distinct firms across the 18 Perplexity discovery runs.

Verification log

How the zero was tested, including the false positives and the false negative it produced.

Changelog

Corrections made to the report after publication, and what changed.

The headline figures

Each of these is a formula in the workbook, not a typed number. The denominators are the point: one discovery answer covers all five companies in its category, so 270 rows are 54 independent runs.

The largest effect in the study is not what a company has done. It is whether its name is already in the question.

What the engines offered instead

Across the 18 Perplexity discovery runs — the only subset with permanent, re-checkable URLs — Perplexity named 133 distinct firms a reader could hire or buy. Not one was in the sample.

Study-wide distinct is lower than the sum of the categories because Accenture, Deloitte, IBM and Capgemini are each offered in two of them. In project management the locked set is Asana, ClickUp, Jira, monday.com, Smartsheet and Wrike; in consulting it is Accenture, Bain, BCG, Deloitte, L.E.K., McKinsey, Oliver Wyman and PwC. In the other four categories no firm at all survived all three rephrasings.

How the zero was tested

A finding of zero is the easiest kind to get wrong, because a matcher that misses a name produces exactly the result the author was hoping for.

Every discovery answer was scanned in full against all thirty company names plus twenty-six additional tokens — trading names, hyphen and spacing variants, and the distinctive word in each name — using whitespace-bounded normalised matching rather than a plain substring test. 36 of the 54 runs were verified against complete source text; Perplexity's eighteen were re-opened at their permanent URLs and re-scanned twice, on 20 and 26 August. Zero matches, both times.

The matcher produced two false positives and one false negative, and all three are documented in the workbook. “Arketi” fires inside the word marketing. “Out of the box” fires on the phrase not a traditional end-to-end PMO platform out of the box. And H2R-Solutions is written with a hyphen, so a matcher that tokenises on punctuation misses it. We publish these because they are the failure modes most likely to have produced a false zero, and a reader should be able to see that we looked for them.

What the data does not show

The study set out to test whether a company's AI Confidence Score predicts whether engines cite it. On discovery prompts that question cannot be answered at all, because every company scored zero and there is no variance to explain.

The 95% interval on the Pearson r runs from −0.03 to +0.62 and includes zero. Power at the observed effect is 44%; roughly n = 69 would resolve it. The result is unresolved, not null, and we publish it that way. If you run this correlation yourself, please report the interval rather than the point estimate. The four companies with a Spotlight on Startups founder profile are too few to compare against the rest, and no such number appears in the report.

Limits, stated up front

• One snapshot. Each prompt was run once per engine. This measures what three engines returned on two days in August 2026; it cannot separate a stable pattern from run-to-run variation. The 2027 edition is designed to.
• n = 30. Wide enough to show a zero, too narrow to resolve a modest correlation.
• B2B SaaS means project-management software only. The category was fixed to one use case and is not a read on B2B SaaS generally.
• Uneven reproducibility. Perplexity results are permanently addressable; ChatGPT and Gemini sessions are not. A third of the discovery evidence rests on excerpts recorded at collection, and the workbook marks which rows those are.
• The composite score is not a mean. The overall AI Confidence Score weights earned media and entity clarity most heavily and cannot be reproduced from the five layer columns. The layer scores are the auditable part.
• The publisher is a marketing company. Four of the thirty companies have a Spotlight on Startups founder profile. One of those four is also a paying client of Kell Web Solutions, Inc. — and is one of the companies the engines failed to recognise. The other three have no commercial relationship. Every company is flagged in the workbook.

Citing this dataset