CiteCurrent, October 2026: which sources AI search engines cite https://citecurrent.com/reports/2026-10/ Question panel: core-v1 (version 1). Domain list: 1.5. Collected 2026-10-20 to 2026-10-20 (UTC), from US. Not collected in this report: Perplexity. We could not read its answers signed out, so it was left out of the collection and the files have no runs for it. Claude, Anthropic's AI assistant, classed the 1,156 entries new to the domain list. A second coder, a separate Claude instance with a fresh context, classed a random 116 of them. It drew them from the 1,153 entries with an evidence record and did not see our classes. It agreed on 106 (91.4%, Cohen's kappa 0.86). We decided the 10 disagreements, with advice from 3 AI reviewers who did not know which class came first. Our class stood for 6, the second coder's for 4. Changed after the collection: - Websites are grouped by the Public Suffix List's ICANN section (October 22, 2026 copy), a rule set after the collection. - A list entry for a suffix, such as nhs.uk, now covers each website under it. Together they changed no class or share. - Most-cited lists keep every website tied for last place (October 22), which added usda.gov to ChatGPT's and walmart.com to Copilot's. - List versions 1.1 and 1.2 gave 3 sections of websites already on the list their own class: techcommunity.microsoft.com/discussions, learn.microsoft.com/en-us/questions and capterra.com/resources. - That moved 9 citations. Copilot's earned media gained 3 of them: it holds 582 of 1,238 citations, against 579 without them (47% either way). - Google's AI Mode and AI Overviews each lost 3 brand-owned citations to user-generated. - Without these sections, 8 rounded shares would move by 1 or 2 points, such as AI Overviews' brand-owned share (34% rather than 33%). - List version 1.3 (October 26) classes calculator, cost-estimate and statistics sites as reference rather than earned media. - That moved 41 citations, 37 of them Copilot's, and took Copilot's earned-media share from 50% to 47%. No most-cited list changed. - It also classes file hosts such as googleapis.com as other (2 of Copilot's citations, from user-generated) and marketplaces as retailers. - List version 1.4 (October 27) classes endlesstravelplans.com as brand-owned rather than earned media: it sells a trip-planning app. - That moved 5 citations, 3 of Copilot's and 2 of Gemini's. No most-cited list changed. - 5 rounded shares moved by 1 point, such as Gemini's earned-media share (31% rather than 32%), and some 95% intervals moved too. - List version 1.5 (October 27) changed no class in this report. Method: https://citecurrent.com/methodology/ License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Reuse the data freely, with credit to CiteCurrent. Suggested citation: CiteCurrent (2026). Which sources AI search engines cite, October 2026. https://citecurrent.com/reports/2026-10/ The publisher's interest: CiteCurrent's publisher works in digital PR, which has an interest in how often AI answers cite earned media. The questions are fixed before each collection and the counting rules before each report, and a report names any rule changed after its collection. Details: https://citecurrent.com/disclosure/ Files: every column, in the order the CSVs give them - runs.csv: one row per run: a question put to one service once. run_id: the run's ID. sweep_id: the collection day. prompt_id: the question's ID, as on https://citecurrent.com/questions/. surface: the AI service: chatgpt (ChatGPT), perplexity (Perplexity), copilot (Microsoft Copilot), gemini (Gemini), google-ai-overview (Google AI Overviews), google-ai-mode (Google AI Mode). repeat: which of the question's runs on that service it was, 1 to 3. attempt: the try that collected it; a try that failed is sent again. started_at: when the run started (UTC). status: ok (an answer), refused (an answer that declines to help), no_answer (for Google AI Overviews: no Overview was shown) or failed (nothing was collected). citations: the sources the service marks in or beside the answer. more: ChatGPT's extra sources, which it did not cite: mostly pages it used, and some shop and brand buttons. They are not counted as citations. panel_version and domain_list_version: the versions of the question panel and the domain list used. - citations.csv: one row per source per answer. run_id, prompt_id and surface are as in runs.csv. position: the source's place in the service's list, from 1. shown: inline (linked in the answer text), listed (in its source list) or more (one of ChatGPT's extra sources; no figure counts these). url, title and domain: the page's address, its title and its website. A blank title in a url row means that answer's record has no title for the page. Another row for the same page may have one. source_class: the kind of website: earned-media (Earned media), ugc (User-generated), brand-owned (Brand-owned), reference (Reference), social-video (Social and video), gov-academic (Government and academic), other (Other), unclassified (Not yet classified). A report is published only when every counted citation has a class, so unclassified rows are among the more rows. subclass: a finer kind within the class, such as vendor, news or forum, blank when there is none. page_type: the kind of page, such as article, company page or forum thread. published_as: how much of the address is published: url (the full address), domain (the website and the kind of page, for videos, social posts, search pages and pages that may be by or about a private person) or type (the kind of page only). Fields left blank are withheld under that rule. panel_version and domain_list_version: as in runs.csv. We publish counts and the addresses of cited pages. We do not publish the text of the answers. Anyone can ask us to remove a page about them: https://citecurrent.com/contact/