# Gadex B2B AI Citation Benchmark 2026 — dataset README

Version 1.0 · Fieldwork: 11 August 2026 · Market: United States, English

## Package contents

| File | Rows or records | Purpose |
|---|---:|---|
| `gadex-b2b-ai-citations-2026.csv` | 1,576 data rows | One row per citation occurrence after within-answer URL deduplication |
| `gadex-b2b-ai-prompts-2026.csv` | 50 data rows | Complete frozen prompts, seed queries, categories and demand fields |
| `gadex-b2b-ai-answers-2026.jsonl` | 150 records | One metadata record per answer, including integrity hash and cited-source array |
| `gadex-b2b-ai-methodology-2026.json` | 1 object | Machine-readable study design, definitions and release status |
| `gadex-b2b-ai-response-text-rights-review-2026.md` | — | Publication-rights review for full generated responses |

The package represents 50 prompts × three engines: ChatGPT, Gemini and Perplexity. It contains 1,576 citation occurrences, 1,211 distinct normalized URLs and 632 distinct root domains.

## Important definition

A **citation occurrence** is one distinct normalized URL within one answer. Repeated links to the same normalized URL inside one answer count once. The same URL cited in a different answer counts as another occurrence.

This definition is why the citation CSV contains exactly 1,576 rows even though a source can appear several times across the study.

## Citation CSV fields

- `citation_id`: stable row identifier, `C-0001` to `C-1576`.
- `answer_id`: joins the citation to an answer JSONL record.
- `prompt_id`: joins the citation to the full prompt CSV.
- `engine` and `engine_key`: display and normalized engine names.
- `model_name`: exact model label returned by the API.
- `query_index`: public 1-based prompt order.
- `seed_query`: shorter phrase used for demand and Google-SERP comparison.
- `full_prompt`: exact conversational prompt sent to the model.
- `citation_url` and `normalized_url`: the normalized cited URL retained by the analysis. They are identical in version 1.0 because the raw pre-normalization URL was not retained as a separate field.
- `root_domain`: registrable root domain used for domain-level analysis.
- `citation_title`: title supplied in the response annotation where available.
- `source_type`: rule-based ownership/type class.
- `page_format`: rule-based format class based on domain, URL and title signals.
- `google_top10_domain_match`: whether the root domain appeared among the first ten organic results for the associated shorter seed query.
- `google_organic_result_rank_absolute`: absolute SERP rank reported by DataForSEO. It can exceed 10 when non-organic SERP features precede an organic result.

## Prompt CSV fields

The prompt file contains the exact 50 submitted prompts, not merely the seed queries. It also retains category, market, language, fieldwork date, engines, generation settings and the contemporaneous DataForSEO keyword-demand fields used for opportunity validation.

Search volumes are rounded, directional estimates—not traffic forecasts. DataForSEO returned a US volume record for 49 of 50 seeds; `best chatbot software for business` had no returned demand record in this collection.

## Answer JSONL fields

Each line is a valid JSON object. Important fields include:

- `answer_id` and `prompt_id` for joins;
- collection date/time, engine and requested/returned model;
- exact prompt and experiment settings;
- input/output token counts;
- `answer_character_count` and `answer_text_sha256` for integrity checks;
- `answer_text_included: false` and a machine-readable omission reason;
- citation count, distinct cited-domain count and an array of that answer's cited sources.

The file intentionally excludes full response text and all reasoning sections. See the rights review before changing that release decision.

## Classification limits

Source ownership/type and page format were assigned by deterministic domain, URL and title rules. They were not created through manual inspection of all 1,211 pages.

`publisher_agency_or_unclassified` is a residual group. In particular, `software_vendor_owned` means **identified vendor-owned under the version 1.0 mapping**, so it should be read as a conservative classified share rather than an exhaustive census of all possible vendor domains.

Four Gemini citations retained a `vertexaisearch.cloud.google.com/grounding-api-redirect/...` URL because the redirect could not be resolved during collection. Their destination root domains were inferred from the API-provided citation title. The rows remain in the data and are disclosed in the methodology rather than silently removed.

## Reuse and licensing

Copyright © 2026 ALPHAOSCAR EURL, publisher of Gadex AI. All rights reserved. No permissive public data license is granted in version 1.0. See [`LICENSE.txt`](./LICENSE.txt) for the scope of the notice and the contact for reuse or redistribution permission.

The notice covers the Gadex-authored prompt inventory, methodology, classification annotations, dataset compilation, editorial text and research visualizations. It does not claim ownership of third-party URLs, page titles, trademarks or other third-party material referenced by the study. Full generated response text is not included.

When citing the dataset, use:

> Gadex Research (2026), *The B2B AI Citation Benchmark 2026*, version 1.0. Fieldwork conducted 11 August 2026.

## Validation checks

- 50 prompt rows.
- 150 answer JSONL records.
- 1,576 citation data rows.
- Every answer has at least one citation.
- Every citation joins to one answer and one prompt.
- No credentials or hidden reasoning text are included.