Skip to content

Data

Broadcastwell publishes its measurement data under the Creative Commons Attribution 4.0 license (CC BY 4.0). You may copy, share and build on it. Attribute Broadcastwell and link the DOI.

A release is a fixed set of files with a version DOI on Zenodo. Each release states its own method version, engines, questions, repeat count and capture dates. A release is never edited in place. A correction gets a new version and its own DOI, and the earlier version stays online with a notice. Standings in one release are not compared with standings in another unless a study says so and shows the test.

To cite a dataset, use the "How to cite" line under its entry. Each line carries the version DOI. The DOI resolves to the record page, where the files, checksums and README live.

Datasets

The Absence Index, release 2026-09

Standings for 14 vertical B2B software categories and 286 checked vendors. Five engines answered ten buyer questions per category: ChatGPT, Claude, Perplexity, Google AI Overviews and Google AI Mode. Each engine and question pair ran three times as scheduled repeats on 22 and 23 September 2026 (UTC). The release reports 2,250 observations: 2,100 targets and 150 controls. Five excluded target rows leave 2,095 valid target answers. Every answer is included verbatim with its cited URLs. The package holds release-2026-09.json, answers-2026-09.jsonl, 14 category CSVs, control-questions.json, a README with the data dictionary and SHA256SUMS.txt. Method v1.1. Standings only. No change between releases is asserted.

The Absence Index, August 2026 release

The first Index release: CRM and project management, ten checked vendors each. Four engines answered ten questions per category, three repeats each, captured 28 August 2026: ChatGPT, Claude, Perplexity and Google AI Overviews. CRM has 117 valid observations of 120 (three engine errors); project management has 120. Version 2026-08.1, published 12 September 2026, corrects the citation counts and records the erratum in the release file. Answer text was not retained for this release. The zip holds the release JSON, crm.csv, project_management.csv and a README with the erratum. The first version of this release, DOI 10.5281/zenodo.22150322, is superseded and stays online with a notice.

The 2026 State of GEO, Volume I

Ten standardized buyer questions for each of 85 B2B software companies across 61 categories, put to one AI engine with live web search between 18 and 23 July 2026. The 860 answers were scored for whether the company was named and whether its domain was cited. 5,160 citations were logged across 1,753 unique domains. Company identities are withheld (C01 to C85). Files: company_visibility_anonymized.csv (85 rows), most_named_brands_top50.csv (50 rows), cited_source_domains_top100.csv (100 rows) and the write-up.

The 2026 State of GEO, Volume II: The Absence Ladder

The 616 answers from the Volume I frame in which the engine answered the buyer question without naming the company, classified by question shape and by the company's visibility tier. Files: absence_questions_classified.csv (616 rows), absence_shape_by_tier.csv (four tiers), challenger_visibility_v2.csv (85 companies). Same single-engine frame as Volume I. The Absence Manual below is built on this volume.

The 2026 State of GEO, Volume III: cross-engine divergence

280 buyer questions across 40 B2B software categories, put to four engines on 4 August 2026: ChatGPT, Claude, Perplexity and Google AI Overviews. Vendor sets were extracted from each answer and agreement between engines was measured. A 25-question subsample ran three times per engine for within-engine consistency, with a length control. Engine coverage is uneven (Google AI Overviews 280 answers, Claude 273, Perplexity 175, ChatGPT 18), and every statistic carries its engine set. Files: raw.csv (821 answers), scored.csv, vendor_mentions.csv, citations.csv, questions.csv, universe.csv, results.json, the paper (State-of-GEO-Volume-III.pdf and PAPER.md) and six figures. A corrections file is kept in the GitHub repository. The paper is also posted on SSRN, abstract 7234418.

The 2026 State of GEO on figshare

A deposit of the Volume I and Volume II files on figshare, made 31 July 2026: six CSVs, PAPER.md and five figures. The data is the same as in the Zenodo records above. The deposit is filed under the author's name, Sairam Sivakumar.

Hugging Face dataset: Broadcastwell/state-of-geo-2026

The six Volume I and Volume II CSVs on Hugging Face (98.6 kB), for loading with the datasets library. The data is the same as in the Zenodo records above.

The Absence Manual

A free technical manual on measuring vendor absence in AI answers, built on Volumes I to III. Thirteen chapters and five appendices, including the ten-question buyer bank, the classification rules, the scoring specification and a chapter on measurement noise. Version 1.0, 41 pages. Prose and figures are CC BY 4.0. The site code is MIT.

Study 1. The Shortlist Gap: who AI engines name when buyers ask, and who they never name

Vendor, category, engine and source-domain figures computed from the Absence Index release 2026-09: 14 categories, 286 checked vendors, 140 buyer questions, five engines, three scheduled runs per engine and question, 2,095 valid target answers. Headline: 46 of 286 checked vendors (16.1%) were never named in any of the 2,095 answers. Files: vendors_2026-09.csv, category_summary_2026-09.csv, source_domains_2026-09.csv, top10_sources_2026-09.csv, domain_classes_2026-09.csv, study1_numbers.json, vendor_rules_2026-09.json and the scripts that recompute every figure from the release package.

Study 2. What AI search visibility costs in 2026

Prices, guarantees and checkout paths of sellers of AI search visibility products and services, read from each seller's own live pages on 25 September 2026. 137 seller profiles were read. 123 sellers had verified in-category offer rows, including rows without a published paid price. Headline: 56.9% of the 123 sellers publish a paid price. Broadcastwell sells in this market. Its own price rows are listed in a separate file and are not scored. Files: prices_2026-09-25.csv, guarantees.csv, checkout.csv, bw_prices.csv and the script that computes every figure.

Study 3. Run-to-run noise: what one AI answer tells you and what three do

Within-day repeat variation in the Absence Index release 2026-09, where each engine and question pair ran three times. A cell is one question, one engine and one checked vendor; 14,192 cells have three valid runs. Headline: 22.2% of single-run vendor namings were not repeated by both other runs. Files: run_cells_2026-09.csv, vendors_by_run_2026-09.csv, study3_numbers.json and the recompute scripts. The study measures repeat variation on one day. It asserts no change over time.

The September Index figure (16.1% of 286 vendors, five engines, three runs) uses a different sample from Volume I's 35% (85 companies, one engine). These are different samples, not a trend.

The monthly letter of the Absence Index

The Absence Report

Who AI engines name, what the evidence shows and what it means for B2B software teams.

Subscribe on Substack

Free. Monthly. Unsubscribe any time. Record Watch is a separate subscription.

Broadcastwell

AI search visibility for vertical B2B software companies.

Broadcastwell LLC
517 E Kirkwood Ave
Bloomington, IN 47408
Product
Research
Company
Contact
Broadcastwell LLC. Indiana, USA. Copyright 2026.