Download
Four CSV files, downloadable without filling in any form. The first round was sampled September 12–13, 2026 and the day-30 rerun September 25–26; both use the same columns and classification rules.
AI Citation Study Datasets
AI Citation Study 2026-09: Run Log (runs.csv)
216 runs, one row each: question ID and type, engine, sample number, whether an external page was cited, number of cited pages, and test city.
AI Citation Study 2026-09: Cited Sources (citations.csv)
560 rows, one per run × cited page: source kind (web / map / chrome), domain, page URL and title. Every statistic in the article can be recalculated from it.
AI Citation Study 2026-09, Day-30 Rerun: Run Log (runs.csv)
216 runs, one row each, with the same columns as the first round. Sampled September 25–26, 2026.
AI Citation Study 2026-09, Day-30 Rerun: Cited Sources (citations.csv)
509 rows, one per run × cited page, with the same columns as the first round. Every day-30 statistic can be recalculated from it.
Methodology and Limitations
The analysis and conclusions for this data are in ChatGPT vs Gemini Citation Overlap: 17.8% in a 216-Sample Study. This page only documents how the samples were taken, the classification rules and the raw files.
Questions and Sampling
The question set has 36 questions written in Traditional Chinese, in three groups by intent: 18 recommendation questions looking for someone to hire, 9 knowledge questions about how to do something, and 9 comparison questions about differences and trade-offs. Each question was run three times on ChatGPT and three times on Gemini, 216 runs in total, all completed.
Each question was run three times rather than once because answers from generative assistants vary randomly, and a single result cannot represent how a question is usually answered. The first round ran from late on September 12, 2026 into the early hours of September 13, using the logged-out web interfaces, with identical conditions for all 216 runs.
The day-30 rerun took place September 25–26, 2026 with the same questions, engines, test machine and outbound IP as the first round, and all 216 runs completed. Eight of them were rerun at noon on the 26th, because ChatGPT's logged-out interface stopped producing answers after about 100 consecutive queries and recovered a few hours later. The series ended on day 30, and the originally planned day-60 and day-90 reruns were not carried out.
What Counts as a Cited Source
Only links pointing to a content page on an external website are counted. Two kinds of links are excluded:
Map listings. These link to real businesses, but they come from a map database rather than from a website's content being cited, and counting them together would distort the picture of who the AI actually cited.
Links belonging to the interface itself. Login, help and map attribution links are the same on every page and carry no information.
This distinction matters a great deal. ChatGPT often attaches a map panel to answers for recommendation questions, so counting map links would clearly overstate its citation rate. The same page also often appears several times as different passage fragments, so citations were deduplicated by page before counting to avoid counting one source more than once.
Scope of the Data
What is published is the classification for each run, plus the URL, title and domain of every cited page.
The full answer text is not included, because the terms of service of generative assistants restrict large-scale reproduction of their output, so only the metrics and sources are released. The test location is given only at city level.
Limitations
- Three samples per question per round from a single test city, with the two rounds 13 days apart; the results describe these two batches of observations only.
- Only the logged-out web interfaces of ChatGPT and Gemini are covered, and logged-in use or the API may behave differently.
- Recommendation questions are affected by where they are asked from. Although the questions specified Taiwan, both engines answered with businesses in the city where the test ran, so regional results cannot be generalized.
- Being listed as a source and actually supporting the answer are two different things, and this study counts only the former.
Column Reference
runs.csv (one row per run)
| Column | Description |
|---|---|
| run_id | Run ID |
| question_id | Question ID; A = recommendation, B = knowledge, C = comparison |
| question_type | recommendation / knowledge / comparison |
| question | Original question text (Traditional Chinese) |
| engine | chatgpt / gemini |
| sample_no | Sample number for this question on this engine (1–3) |
| requested_at_utc | Time the run was queued (UTC); for runs automatically requeued after a failure, the requeue time |
| completed_at_utc | Completion time (UTC) |
| grounded | 1 = cited at least one external content page |
| web_page_count | Number of external content pages cited (deduplicated) |
| cited_page_count_all_kinds | External content pages plus map listings |
| runner_city | Test city |
citations.csv (one row per run × cited page)
| Column | Description |
|---|---|
| run_id, question_id, question_type, engine | Same as runs.csv |
| kind | web = external content page; map = map listing. Interface links are not included |
| domain | Domain with www. removed |
| page_url | URL with passage fragments (#:~:text=) removed; source parameters added by the engine are kept |
| page_title | Captured page title, up to 160 characters; Gemini's title field often includes snippet text |
File Checksums (SHA-256)
| File | SHA-256 |
|---|---|
| seobar-ai-citation-2026-09-runs.csv | 0cf24f594f37a5d2ff71af82dad7169926db8d85e50e4927ae060b5cdc9dc4d6 |
| seobar-ai-citation-2026-09-citations.csv | f7cbb978235389d89828d08aaf387ad33a779dbee073a2156176b28f56d2a47f |
| seobar-ai-citation-2026-09-day30-runs.csv | 5ae58f5ff78385cb4456654cc723d56f39bfaf3b4a140b4f4ab0596cc4223794 |
| seobar-ai-citation-2026-09-day30-citations.csv | 63d38e02d45a15d8f3182905045ff4f1f1ab01516962ac52a1bd3b81142c244d |
License and Citation
This data is released under CC BY 4.0. You are free to use and adapt it; please credit the source:
SEOBAR (2026). AI Citation Sources Dataset 2026-09. https://seobar.dev/research