skaldanalytics/ai-bots-robots-txt-top1000

Who blocks which AI bots? robots.txt of the Tranco top 1,000

What the robots.txt of the 1,000 most popular domains (Tranco list 56Z8N) says to 16 crawlers: the AI crawlers of OpenAI, Anthropic, Perplexity, Google, Common Crawl, ByteDance, Apple, Meta and Amazon, plus Googlebot as a baseline. Fetched on 23 October 2026. 527 of the 1,000 domains gave us a file we could read; the rest are marked unknown (or have no file at all), see Coverage.

• Data license: CC BY 4.0. The scripts in code/ are Skald's and are covered by the Elastic License 2.0 (source-available, free to self-host). The raw files in raw_robots_txt.zip are copies of public files written by the sites themselves, reproduced as served for research; we claim no rights in their content.
• Published by: Skald · code: https://github.com/spcaeo/skald
• Version: 1.0 (23 October 2026 snapshot)
• Companion tool: an AI bots checker that runs the same kind of check on a single site; its source is in the Skald repository.

Files

File	What it is
robots_ai_bots_top1000.csv	Main table. One row per domain (1,000 rows, 41 columns): fetch result, then status and matching rule source for each of the 16 agents, plus the exact robots.txt lines with line numbers.
robots_ai_bots_long.csv / .parquet	The same results in long form: one row per domain × agent (16,000 rows). Easier to group and plot.
raw_robots_txt.zip	Every response body we saved, as served (762 files, raw/<domain>.txt). Includes error and HTML pages for domains that did not return a robots.txt.
raw_manifest.csv	For each raw file: domain, rank, HTTP status, content-type, final URL, size, SHA-256, and our fetch class.
fetch_log.json	Per-domain fetch log (status, final URL, URL tried, errors, retry status).
tranco_meta.json	Metadata of Tranco list 56Z8N.
charts/	The two charts below.
code/	The scripts that produced the data (fetch.py, retry.py, parse.py, build.py, analyze.py, charts.py) and verify_numbers.py, which recomputes every number quoted on this page from the CSV. The fetch script expects the Tranco list file tranco_top1000.csv (rank,domain), which you can download from the Tranco link below.

How the data was collected

Domain list. Tranco list ID 56Z8N, top 1,000. Created 2026-10-22; daily list; combination method Dowdall; providers crux, farsight, majestic, radar and umbrella; window 2026-09-23 to 2026-10-22; pay-level-domain filter on. Tranco: Le Pochat et al., Tranco: A Research-Oriented Top Sites Ranking Hardened Against Manipulation, NDSS 2019.

Fetching. Fetched on 23 October 2026 (evening, America/Toronto) from a single cloud server, with the user agent SkaldResearch/1.0 (+https://skaldanalytics.com). For each domain we requested https://<domain>/robots.txt, then https://www.<domain>/robots.txt, then http://<domain>/robots.txt, and kept the first response. Redirects were followed (up to 10), with a 10 second timeout per request. Domains with no response, or with status 202, 400, 401, 403, 405, 409, 418, 421, 429, 498, 500, 503 or 520 in that pass, were retried once with curl (15 second timeout, up to 10 redirects, https://<domain> then https://www.<domain>). A retry replaced the first result only if the first pass had no response or the retry returned 200. TLS certificate verification was switched off in both passes, so a response from a host with a bad certificate was accepted.

Parsing. Our own parser (code/parse.py), following RFC 9309:

• A group is one or more consecutive User-agent lines followed by its Allow / Disallow rules. Comments are stripped; useragent and user agent spellings are accepted; empty Disallow: lines are ignored.
• Product tokens compare case-insensitively, after cutting any /version suffix. A group that names the agent applies to it, and a group that names the agent beats the * group (all groups naming the agent are combined). Only if no group names it does * apply. If neither exists, nothing applies.
• Inside the applicable rules the longest matching path wins; on a tie, Allow wins. * and $ wildcards are supported. We test the path /.
• A 404/410 means "no file, everything allowed". A 401/403, a 5xx, a timeout or an HTML page served in place of the file means we have no usable file, so the agent is recorded as unknown (we do not apply RFC 9309's "unreachable means disallow everything" rule to those; they are simply out of scope).
• Differences from RFC 9309 and from some real crawlers: no 500 KiB size cap, no percent-encoding normalisation, Crawl-delay and Sitemap lines are ignored.

Statuses (per domain and agent, in the 527 parsed files):

Status	Meaning
blocked	The rules that apply to the agent disallow /, and there is no real Allow left. Allows for /robots.txt, /ads.txt, /app-ads.txt, /sitemap*, /favicon.ico, /llms.txt and /.well-known/* are ignored for this purpose.
partial	At least one rule applies to the agent but it is not a full block. This covers the common "keep out of /search, /cart, /account" files, and files that disallow / but then allow real paths.
allowed	No rule applies, or there are only empty Disallow: lines.
unknown	No usable file (see below).

matched_by / <Agent>_src says where the verdict came from: named (a group names the agent), * (only the catch-all group), or none (no group applies).

Cross-check. We re-ran all 527 parsed files through the separate parser used by the AI bots checker (robots_rules.py). For blocked the two parsers agree on every one of the 527 × 15 agent decisions (the checker has no anthropic-ai verdict). The only disagreements are partial here versus allowed in the checker: the home page is open and either the file only closes generic paths (for example /admin/ under *) or the agent's own group contains only Allow rules. The checker counts that as allowed; this dataset calls it partial. All headline numbers on this page use blocked only.

Coverage

Result (fetch)	Domains	Agent columns
ok: robots.txt parsed	527	blocked / partial / allowed
error_timeout: no response in time or connection failed (mostly CDN, DNS and API hosts with no website)	238	unknown
no_robots_404: the site answered 404/410, so no file	75	allowed (no file, nothing asked), matched_by = none
html_not_robots: an HTML page came back instead of a file (e.g. linkedin.com served a reCAPTCHA page)	65	unknown
forbidden: 401/403 to our user agent (e.g. oracle.com, godaddy.com, lowes.com, state.gov)	56	unknown
error_http: any other HTTP error or odd status	34	unknown
empty: a 200 with an empty body	5	allowed, matched_by = none
Total	1,000

473 of 1,000 domains therefore gave no parseable file (393 unknown plus 80 with no rules). Every percentage and count below is of the 527 sites with a parsed robots.txt unless stated. A different user agent, network or day could give a different split; sites that blocked us at the CDN are among the unknown, so the sample leans toward sites that answer anonymous requests.

Column dictionary: robots_ai_bots_top1000.csv

Column	Type	Meaning
rank	int	Rank in Tranco list 56Z8N (1 to 1000).
domain	string	The Tranco domain.
category	string	news_media for a hand-picked list of 41 news and media domains (see code/analyze.py), otherwise other.
http_status	string	HTTP status of the response we kept; empty if nothing answered.
fetch	string	Fetch class: ok, error_timeout, no_robots_404, html_not_robots, forbidden, error_http, empty.
final_url	string	URL the response came from after redirects.
n_groups	int	Number of user-agent groups parsed (0 unless fetch is ok).
<Agent>	string	One column for each of the 16 agents: GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, Claude-User, Claude-SearchBot, anthropic-ai, PerplexityBot, Perplexity-User, Google-Extended, CCBot, Bytespider, Applebot-Extended, Meta-ExternalAgent, Amazonbot, Googlebot. Value: blocked, partial, allowed or unknown.
<Agent>_src	string	Where that verdict came from: named, * or none; empty when unknown.
all_user_agents	string	Every distinct user-agent name in the file, separated by |.
relevant_lines	string	The lines that decided the result: each group that names an AI-related token, with its rules and line numbers ([L350] UA: GPTBot => Disallow: /), plus the * rules. Compact; the full file is in the raw zip.

robots_ai_bots_long.* columns: rank, domain, category, fetch, agent, owner, agent_role (training, search, user fetch and so on), status, matched_by.

Agents used for the headline comparisons: training crawlers = GPTBot, ClaudeBot, CCBot, Google-Extended; referral agents (the ones that can send a reader back) = ChatGPT-User, OAI-SearchBot, Perplexity-User, Claude-User.

Key findings

All numbers recomputed from the CSV with code/verify_numbers.py (it exits non-zero if any number here stops matching).

Sites that fully block each agent (of 527)

Agent	What it does	Blocked	Of those, named in the file
Bytespider	ByteDance crawler	111	75
CCBot	Common Crawl (training dataset)	105	70
ClaudeBot	Anthropic training	94	66
GPTBot	OpenAI training	91	65
Meta-ExternalAgent	Meta training	83	50
Google-Extended	Google AI training control	82	56
Applebot-Extended	Apple AI training control	77	46
PerplexityBot	Perplexity search index	73	47
Amazonbot	Amazon	72	40
anthropic-ai (retired)	old Anthropic token	68	30
Perplexity-User	user-requested fetch	68	31
Claude-User	user-requested fetch	63	32
ChatGPT-User	user-requested fetch	60	31
Claude-SearchBot	Anthropic search	58	26
OAI-SearchBot	ChatGPT search	49	20
Googlebot	Google Search	14	1

• 122 sites (23%) fully block at least one training crawler; 65 block all four.
• 81 sites (15%) fully block at least one referral agent. 37 of the 81 never name one of them: they are caught by a blanket User-agent: * / Disallow: /.
• 44 sites block a training crawler but leave all four referral agents unblocked.
• 179 sites (34%) name at least one of the 15 AI agents we checked.

Facts you can check in the CSV (filter domain, read the agent column and relevant_lines)

1. Training out, search in. GPTBot is fully blocked by 91 sites, OAI-SearchBot by 49. ClaudeBot by 94, Claude-SearchBot by 58. 42 sites block GPTBot but not OAI-SearchBot, among them forbes.com, espn.com, techcrunch.com, apnews.com, cbsnews.com, latimes.com, airbnb.com, quora.com, tripadvisor.com, yelp.com, figma.com, ted.com and nature.com. (Most still keep OAI-SearchBot out of some paths under their general rules.)

2. forbes.com (rank 217) spells the split out. L350 User-agent: GPTBot / L351 Disallow: /, then L360 User-Agent: OAI-SearchBot / L361 Allow: /. Same for Anthropic: L334 ClaudeBot / L335 Disallow: /, then L376 Claude-User / L377 Claude-SearchBot / L378 Allow: /.

3. An allowlist with a gap. netflix.com (rank 43) opens with L1 User-agent: * / L2 Disallow: /, then lets in a list of 28 bots (L4 to L31), including GPTBot, ChatGPT-User, OAI-SearchBot, ClaudeBot, Claude-User, Claude-SearchBot and PerplexityBot, with L32 Allow: /. Perplexity-User is not on the list, so the blanket block applies to it. The same kind of gap appears on 7 more sites: netflix.net, disneyplus.com and trustpilot.com (Perplexity-User), flickr.com (GPTBot, ClaudeBot, Perplexity-User), weibo.com (GPTBot, ClaudeBot, Claude-User, Claude-SearchBot, PerplexityBot), academia.edu (ClaudeBot, Perplexity-User) and wsj.com (ClaudeBot, Claude-User, Claude-SearchBot, PerplexityBot, Perplexity-User, see fact 10). A robots.txt does not say whether a gap is a choice or an oversight.

4. A hyphen that isn't a hyphen. chatgpt.com (rank 112) has L42 User-agent: Perplexity‑User / L43 Disallow: /, but the dash is U+2011, a non-breaking hyphen. A parser that compares names as text will not match it to Perplexity-User. The same character appears on launchpad.net (L17) and academia.edu (L1923). Perplexity's crawler documentation writes the name once with that same character in its table, so it was probably copied from there. (Perplexity also says Perplexity-User "generally ignores robots.txt rules", so in practice the line may not matter either way.)

5. AI companies block each other. claude.ai (rank 412): L6 User-Agent: GPTBot / L7 OAI-SearchBot / L8 ChatGPT-User / L9 Google-Extended / L10 Disallow: /. chatgpt.com blocks CCBot (L3), Google-Extended (L9), Bytespider (L33), PerplexityBot (L39) and the retired anthropic-ai (L18) and Claude-Web (L21), but has no ClaudeBot group. openai.com and anthropic.com do not block any of these bots from the home page (openai.com keeps one path, /microsoft-for-startups/, closed to everyone).

6. Old names live on. 89 sites still name the retired anthropic-ai or Claude-Web tokens. Five name them with no ClaudeBot group at all: chatgpt.com, espn.com, repubblica.it, pornhub.com, and cloudflare.com (as an Allow). Anthropic told 404 Media in July 2024 that ClaudeBot respects rules written for those old names, so the intent likely still holds, but only because the company chose to honour it.

7. Blanket blocks hit the referral bots too. x.com ends with L97 # Every bot that might possibly read and respect this file / L99 User-agent: * / L100 Disallow: /. twitter.com, pinterest.com, imdb.com, baidu.com, t.co and xiaohongshu.com also end in a blanket block and let Googlebot in. On x.com, twitter.com, pinterest.com, baidu.com, t.co and xiaohongshu.com every one of the 15 AI agents is blocked (on x.com and twitter.com, Google-Extended and Meta-ExternalAgent have groups of their own, the other 13 fall under *). On imdb.com 14 of the 15 fall under the blanket block; Amazonbot has its own group.

8. reddit.com blocks Googlebot too, as we were served. L7 User-agent: * / L8 Disallow: /. Every agent, Googlebot included, is blocked in the file we received. Only 14 of 527 sites block Googlebot; the rest are mostly ad-tech and infrastructure hosts.

9. Newsrooms block training far more, but not OpenAI's search bot. Among the 41 news and media sites (all of which gave a parsed file): CCBot 59% vs 17% for the 486 other sites, ClaudeBot 46% vs 15%, PerplexityBot 37% vs 12%, GPTBot 32% vs 16%. OAI-SearchBot is 7% for news vs 9% for the rest. 27 of the 41 block a training crawler; 9 of those 27 leave all four referral agents unblocked. The news list is hand-picked (see code/analyze.py) and small, so treat the percentages as indicative.

10. Three newsrooms, three policies.

    • nytimes.com blocks all four referral agents by name (e.g. L176 ChatGPT-User, L265 OAI-SearchBot, L281 Perplexity-User) and GPTBot (L232). But L145 User-agent: Amazonbot / L146 Disallow: /wirecutter/ keeps Amazonbot out of Wirecutter only.
    • theguardian.com fully blocks PerplexityBot, anthropic-ai, ClaudeBot, Claude-SearchBot, Claude-User, Applebot-Extended, meta-externalagent and Amazonbot in one group (L65 to L102). OpenAI's bots are not named anywhere in the file.
    • wsj.com starts with L13 User-agent: * / L14 Disallow: /. Then it allows L71 ChatGPT-User / L72 GPTBot / L73 OAI-SearchBot in the same group as social-preview bots (L90 Allow: /). Every other AI agent stays blocked.

    A robots.txt doesn't say why. Publicly known context: News Corp (WSJ's owner) signed a content deal with OpenAI in May 2024, and the NYT signed one with Amazon in 2025. This context does not come from the dataset.

11. Amazon closes the door on rival assistants. amazon.com and 11 other Amazon storefronts block ChatGPT-User, OAI-SearchBot, Claude-User and Perplexity-User by name (amazon.com: L273 ChatGPT-User, L276 OAI-SearchBot, L246 Claude-User, L255 Perplexity-User). tiktok.com, msn.com, nytimes.com and cnn.com do the same.

Limitations

• robots.txt is a request, not enforcement. It shows what a site asks for, not what a bot does or whether a bot obeys. OpenAI says ChatGPT-User is user-initiated and that robots.txt rules may not apply to it; Perplexity says Perplexity-User generally ignores robots.txt rules; Anthropic says its three bots honour robots.txt. We did not test anyone's behaviour.
• One snapshot. One day, one self-identified user agent, one server. Some sites serve different files to different bots, networks or regions, and files change often.
• Unknown sites. 473 of 1,000 domains gave no parseable file. The 238 timeouts are mostly infrastructure (CDNs, DNS, API hosts). The 56 sites that answered 401/403 and the 65 that served HTML are unknown, not allowed. The 75 that returned 404 and 5 empty files are recorded as allowed because no rules exist; they are left out of every count above.
• Certificate verification was off during fetching, so a file served over a connection with an invalid certificate was accepted as is.
• Our parser, our definitions. Statuses come from code/parse.py, not from any crawler's own implementation. partial is a wide bucket; the headline numbers use blocked only. Real crawlers can differ in edge cases (size caps, encoding, merging of groups).
• Only the path / is tested for blocked; the dataset does not say which sections of a site are open.
• The news and media list is hand-picked (41 domains) and the category column has only two values.
• Statements about vendor behaviour or business deals (facts 4, 6 and 10 and this section) come from public vendor documentation and press reports, not from this dataset.

Reproduce

pip install pandas
python3 code/verify_numbers.py robots_ai_bots_top1000.csv   # recomputes every number on this page

To re-fetch, download Tranco list 56Z8N (https://tranco-list.eu/download/56Z8N/1000000), keep the first 1,000 rows as tranco_top1000.csv (rank,domain), and run the scripts in code/ in order: fetch.py, retry.py, build.py, analyze.py, charts.py. Results will differ from this snapshot as robots.txt files change.

Citation

@dataset{skald2026robotsaibots,
  author    = {{Skald}},
  title     = {Who blocks which {AI} bots? {robots.txt} of the {Tranco} top 1,000},
  year      = {2026},
  version   = {1.0},
  note      = {Snapshot fetched 23 October 2026, Tranco list 56Z8N},
  publisher = {Skald},
  url       = {https://skaldanalytics.com/}
}

Cite Tranco as well if you reuse the domain list: Le Pochat, V., Van Goethem, T., Tajalizadehkhoob, S., Korczyński, M., Joosen, W. (2019). Tranco: A Research-Oriented Top Sites Ranking Hardened Against Manipulation. NDSS 2019.

About Skald

Skald is self-hosted web analytics built from your server access log. It is source-available (Elastic License 2.0) and free to self-host. Source: https://github.com/spcaeo/skald.