Bayanat-Labs/arabic-web-rights-census

Arabic Web Rights Census

Which Arabic-language websites tell AI crawlers to stay out? This census records the machine-readable AI-use signals of 36,190 hosts listed: Arabic-language hosts from Common Crawl plus popular domains under Arab and Arabic-script country-code TLDs. For each host it records the robots.txt rule for each of 26 crawler tokens, Cloudflare Content-Signal values, TDMRep reservations (tdmrep.json, headers and meta tags), IETF AIPREF Content-Usage values and noai values in X-Robots-Tag or meta robots. Census 2026-10, collected from 2026-10-30 to 2026-10-31 from one vantage point in the United States. By Bayanat Labs.

As of 31 October 2026, 2.4% of 33,985 hosts checked block at least one AI-training crawler in robots.txt (824 hosts). For the 31,653 of them from the Common Crawl sample alone, the share is 1.8% (555 hosts). Weighted by each host's Arabic pages in the Common Crawl crawl CC-MAIN-2026-43 (a sample of the web), the share is 2.6%. 0.6% (213 hosts) fully block all four of GPTBot, CCBot, ClaudeBot and Google-Extended. 4.4% (1,511 hosts) reserve AI training by at least one signal: a robots.txt block, Content-Signal ai-train=no or a TDMRep reservation.

A robots.txt "allow", or a missing signal, is not permission to reuse a site's content. The census reports signals, not licences.

Look up a domain and see the charts in the browser: Bayanat-Labs/arabic-web-rights-census Space (English and Arabic).

In numbers

Rates are over the 33,985 hosts checked: Arabic-language hosts from Common Crawl plus popular domains under Arab and Arabic-script country-code TLDs, whose robots.txt request was answered with a file, with a 4xx reply or with more than five redirects (the last two mean there is no robots.txt); the tables call them reachable hosts. For the 31,653 of them from the Common Crawl sample alone, 555 (1.75%) block at least one AI-training crawler (aggregates, table = by_size_band, the bands other than 0). Page-weighted shares use each host's Arabic-primary pages in CC-MAIN-2026-43 (11,116,524 pages over the reachable hosts); this weight is a sample, not a count of the web.

Signal	Hosts	Share of hosts	Share of sampled Arabic pages
Full block for at least one AI-training crawler in robots.txt	824	2.42%	2.62%
Full block for GPTBot, CCBot, ClaudeBot and Google-Extended	213	0.63%	0.29%
Any AI-training reservation (robots.txt, Content-Signal or TDMRep)	1,511	4.45%	4.56%
Content-Signal ai-train=no	958	2.82%	2.32%
Any Content-Signal line	1,038	3.05%	2.73%
Cloudflare managed robots.txt block	189	0.56%	0.27%
Full block for an AI search or user-fetch crawler	111	0.33%	0.62%
User-agent: * disallows the whole site (not AI-specific)	127	0.37%	0.11%
Googlebot and Bingbot not fully blocked	33,874	99.67%	99.85%

• Of the 958 hosts with Content-Signal ai-train=no, 187 (19.5%) carry it inside a Cloudflare managed robots.txt block.
• TDMRep: 0 of the 31,520 reachable hosts where /.well-known/tdmrep.json answered or the homepage headers and <head> were read declare a TDM reservation.
• noai in X-Robots-Tag or meta robots: 1 of 31,025 hosts whose homepage headers were read.

Coverage

The host list holds 36,190 hosts. All of them were requested. Their robots.txt request ended as follows:

Outcome	Hosts	Share of requested hosts
File returned	31,709	87.6%
4xx reply or more than five redirects: no robots.txt, no restrictions (RFC 9309)	2,276	6.3%
Unreachable (DNS, connection, TLS or 5xx)	1,170	3.2%
Timed out	650	1.8%
Challenge or block page	385	1.1%

Of the 2,276 hosts counted as having no robots.txt, 78 answered 401 or 403; RFC 9309 treats that as no restrictions, and they were not contacted again. 14 were redirected more than five times.

Unreachable, timed-out and challenged hosts are reported here and are never counted as allowing anything; they are left out of every rate. The live run stopped once under its error-rate rule (more than 20% of the last 1,000 hosts unreachable or timed out) and was resumed with the hosts whose own name does not resolve, while the machine's resolver answers for stable reference names, left out of that count.

By crawler

Crawler token	Operator	Purpose	Hosts with a full block	Share of hosts	Share of sampled Arabic pages	Hosts with a partial block
Bytespider	ByteDance	training	489	1.44%	1.33%	165
GPTBot	OpenAI	training	381	1.12%	1.37%	491
Amazonbot	Amazon	training	367	1.08%	0.72%	302
ClaudeBot	Anthropic	training	360	1.06%	1.27%	476
Meta-ExternalAgent	Meta	training	292	0.86%	0.61%	204
CCBot	Common Crawl Foundation	training	280	0.82%	0.37%	354
Diffbot	Diffbot	training	252	0.74%	0.49%	47
Google-Extended	Google	training	241	0.71%	0.43%	445
anthropic-ai	Anthropic	training	237	0.70%	0.64%	313
Applebot-Extended	Apple	training	232	0.68%	0.63%	364
cohere-ai	Cohere	training	220	0.65%	0.60%	164
omgili	Webz.io	training	211	0.62%	0.30%	24
ImagesiftBot	ImageSift	training	104	0.31%	0.59%	14
ChatGPT-User	OpenAI	user-fetch	91	0.27%	0.55%	435
PerplexityBot	Perplexity	search	67	0.20%	0.46%	474
OAI-SearchBot	OpenAI	search	55	0.16%	0.14%	455
Timpibot	Timpi	training	24	0.07%	0.02%	21
Perplexity-User	Perplexity	user-fetch	22	0.06%	0.08%	319
Claude-User	Anthropic	user-fetch	19	0.06%	0.08%	300
Claude-SearchBot	Anthropic	search	13	0.04%	0.01%	308
AI2Bot	Allen Institute for AI	training	11	0.03%	0.28%	34
DuckAssistBot	DuckDuckGo	search	8	0.02%	0.00%	122
cohere-training-data-crawler	Cohere	training	6	0.02%	0.00%	20
Bingbot	Microsoft	control	24	0.07%	0.09%	412
Googlebot	Google	control	7	0.02%	0.00%	472
BayanatLabs-OpenData	Bayanat Labs	control	0	0.00%	0.00%	0

full_block: a group naming the token disallows the whole site. partial_block: it disallows some paths. Tokens with no group of their own follow the User-agent: * group, which is counted separately (wildcard_full_block), so a site that blocks every crawler is not counted as an AI-specific opt-out.

By sector and country

Sector (rule-based)	Requested	Reachable	Block ≥1 training crawler	Share	Page-weighted share	Any reservation	Share
other	33,287	31,520	737	2.3%	2.5%	1,356	4.3%
government	909	660	15	2.3%	0.0%	15	2.3%
education	893	753	26	3.5%	0.9%	27	3.6%
e-commerce	831	798	33	4.1%	4.0%	100	12.5%
religious	153	148	1	0.7%	0.4%	1	0.7%
news	117	106	12	11.3%	11.0%	12	11.3%

Country is the ccTLD (Arabic-script IDN ccTLDs mapped to their country); hosts under generic TLDs such as .com or .net are labelled gTLD. Labels with at least 200 reachable hosts:

Country label	Requested	Reachable	Block ≥1 training crawler	Share	Page-weighted share	Any reservation	Share
gTLD	27,092	26,231	466	1.8%	2.6%	1,042	4.0%
SA	1,732	1,497	37	2.5%	1.0%	131	8.8%
AE	1,278	1,102	75	6.8%	4.0%	79	7.2%
MA	994	824	51	6.2%	2.6%	52	6.3%
EG	542	413	18	4.4%	3.5%	18	4.4%
TN	455	353	26	7.4%	0.3%	26	7.4%
DZ	382	266	6	2.3%	0.0%	6	2.3%
LY	339	310	22	7.1%	3.3%	22	7.1%
PS	335	303	10	3.3%	3.4%	10	3.3%
JO	304	246	12	4.9%	0.3%	12	4.9%
IQ	292	267	8	3.0%	3.1%	8	3.0%

All labels are in the aggregates config (table = by_country).

Hosts outside the Common Crawl sample

The 2,332 reachable hosts with no Arabic pages in CC-MAIN-2026-43 (added from the Tranco list of popular domains or the fixed baseline) block an AI-training crawler at 11.5%, against 1.8% for the hosts in the crawl sample. Common Crawl does not collect pages from hosts that block CCBot, so a host list built from its pages under-represents blocking hosts; the Tranco and baseline hosts reduce this bias but do not remove it.

Named hosts

Baseline hosts (organisations whose robots.txt was also read rule by rule, separately from the parser, for this census; 10 of 10 hosts and 260 of 260 host and token rules agree with the parser output, qa/baseline_check.md):

Host	Training crawlers with a full block	AI search or user-fetch crawlers with a full block	Content-Signal ai-train
aljazeera.net	GPTBot, ClaudeBot, anthropic-ai, Bytespider, cohere-ai	ChatGPT-User, PerplexityBot	blank
arabic.cnn.com	GPTBot, CCBot, ClaudeBot, anthropic-ai, Applebot-Extended, Bytespider, Amazonbot, cohere-ai, cohere-training-data-crawler, Diffbot, AI2Bot, Timpibot, ImagesiftBot, omgili	OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot, Perplexity-User, DuckAssistBot	blank
bbc.com (whole host)	GPTBot, CCBot, ClaudeBot, anthropic-ai, Google-Extended, Applebot-Extended, Bytespider, Meta-ExternalAgent, Amazonbot, cohere-ai, Diffbot, omgili	OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User	blank
dw.com (whole host)	GPTBot, CCBot, ClaudeBot, anthropic-ai, Google-Extended, Applebot-Extended, Bytespider, Meta-ExternalAgent, Amazonbot, cohere-ai, Diffbot, AI2Bot, Timpibot, ImagesiftBot, omgili	OAI-SearchBot, ChatGPT-User, PerplexityBot, Perplexity-User	blank
arabic.euronews.com	GPTBot, CCBot	none	blank
elaph.com	GPTBot, CCBot, ClaudeBot, anthropic-ai, Google-Extended, Applebot-Extended, Bytespider, Meta-ExternalAgent, Amazonbot, cohere-ai, omgili	OAI-SearchBot, ChatGPT-User	blank
sabq.org	GPTBot, CCBot, ClaudeBot, anthropic-ai, Applebot-Extended, Bytespider	none	no
spa.gov.sa	ClaudeBot, anthropic-ai, Google-Extended, Applebot-Extended, Bytespider, Meta-ExternalAgent, Amazonbot, cohere-ai, Diffbot, omgili	none	no
wam.ae	none	none	blank
youm7.com	none	none	blank

"(whole host)": the census lists the Arabic service of bbc.com and dw.com, but their robots.txt covers the whole host. A host with no AI-crawler block in robots.txt has not given permission to reuse its content; its terms of use still apply.

How to use

from datasets import load_dataset

hosts = load_dataset("Bayanat-Labs/arabic-web-rights-census", "hosts", split="train")
rules = load_dataset("Bayanat-Labs/arabic-web-rights-census", "bot_rules", split="train")
agg = load_dataset("Bayanat-Labs/arabic-web-rights-census", "aggregates", split="train")

row = hosts.filter(lambda r: r["host"] == "aljazeera.net")[0]
print(row["robots_outcome"], row["rule_gptbot"], row["content_signal_ai_train"])

-- DuckDB, straight from the Hub
SELECT host, rule_gptbot, rule_ccbot, content_signal_ai_train
FROM 'hf://datasets/Bayanat-Labs/arabic-web-rights-census/data/hosts/2026-10.parquet'
WHERE training_block_any;

The parser and fetcher are in the arabic-web-rights-census Python package (GitHub, tag v0.1.0):

pip install "arabic-web-rights-census @ git+https://github.com/Bayanat-Labs/arabic-web-rights-census@v0.1.0"
awrc robots robots.txt --bot GPTBot

Exclusion list

exports/exclusion_list_2026-10.csv lists the 1,511 hosts with reserved_any = true: a full robots.txt block for at least one AI-training crawler, Content-Signal ai-train=no, or a TDMRep reservation, with the signals found and the time they were read. Corpus builders can use it as one input when they honour machine-readable reservations of text and data mining rights, for example the reservations that Article 4(3) of the EU Directive 2019/790 provides for and that Article 53(1)(c) of the EU AI Act refers to. It covers only the hosts and signals in this census, at the time they were read; it is not a complete register of reservations and not legal advice. Hosts that were unreachable or not requested are not in it.

Configs and columns

Config	File	Rows	One row per
hosts	data/hosts/2026-10.parquet	36,190	listed host
bot_rules	data/bot_rules/2026-10.parquet	940,940	requested host × crawler token
aggregates	data/aggregates/2026-10.parquet (and .csv)	1,840	rate (table, population, group, metric, weighting)

Main hosts columns: host, country, sector, inclusion_source, tranco_bucket (top-10k / top-100k / top-1M, never the exact rank), cc_pages and cc_ar_primary_pages (pages in CC-MAIN-2026-43), cc_rank (rank by Arabic pages), fetched, robots_outcome and its detail, rule_<token> for each crawler (full_block, partial_block, allow, wildcard_only), wildcard_full_block, content_signal_*, cloudflare_managed, tdmrep_status, tdm_reserved, x_robots_noai, homepage_outcome, and the derived classes training_block_any, training_block_major4, ai_search_blocked, search_allowed, reserved_any. bot_rules adds matched_group (specific, wildcard or none), effective and allows_root per token. robots_sha256 is the SHA-256 of the robots.txt body as served; bodies are not released.

aggregates columns: table, population (reachable for rates; the denominators of TDMRep and noai are the hosts where that signal was observed), group_by, group, metric, weighting (hosts or pages), numerator, denominator, rate (rounded to 6 decimals), crawl_id, as_of.

How it is built

• Host list. Common Crawl columnar index of CC-MAIN-2026-43 (all 300 subset=warc files, HTTP range reads of six columns): hosts with at least 50 HTML pages whose dominant CLD2 language is Arabic and an Arabic share of at least 30%. Added: Tranco list PYK6J domains under the 22 Arab League ccTLDs and the 32 Arabic-script IDN TLDs, and 10 fixed baseline hosts. User-content platform subdomains (blog hosts, site builders, app hosting) and IP literals are removed. SQL, thresholds and file lists: universe/.
• Live requests. At most three requests per host (7 hosts may have received a fourth: 11 of the 1,313 requests made during local network outages ended in a timeout, connection or TLS error rather than at the DNS look-up and may have reached the host): GET /robots.txt, GET /.well-known/tdmrep.json and, only where robots.txt allows our own token BayanatLabs-OpenData to fetch /, one homepage request whose headers and <head> meta tags are read and whose body is dropped. User agent BayanatLabs-OpenData/1.0 (+https://bayanatlabs.com/open; hello@bayanatlabs.com); at least 5 seconds between requests to the same host, measured on the machine's monotonic clock (the ledger records wall-clock times, and 74 gaps read 4.92–5.00 s where the system clock was corrected backwards, at three instants); a host is not contacted again after a 401, 403 or 429 reply, a challenge page or a login redirect; no challenge is ever solved. Hosts whose first robots.txt request timed out or failed on DNS or connection were tried once more at least one hour later. The census used its own fetcher (awrc.fetch), not the polite-fetch library described on the crawler page. Where they differ: it requested /.well-known/tdmrep.json without checking robots.txt; it kept cookies a host set in memory for the run and sent them back to that host; it retried a 5xx robots.txt once after 30 seconds; it treated more than five robots.txt redirects as no robots.txt (14 hosts); it followed homepage redirects without reading the target host's robots.txt; and when a certificate check failed it repeated the request once without verification (1,708 requests).
• Parsing. RFC 9309: groups naming the same token are combined, the longest matching rule wins, Allow wins a tie, * and $ are supported, 500 KiB limit; a 4xx reply means no restrictions; a 5xx reply, timeout or connection failure means unreachable. A separate check against protego 0.7.0 agrees on 99.98% of 4,122,170 can-fetch comparisons over 31,709 hosts (qa/protego_crosscheck.json).
• Aggregates. awrc aggregate over the released host table; integer counts, rates rounded to 6 decimals.

Request ledger for every live request and every Common Crawl index file read for the host list: PROVENANCE.jsonl (no bodies; rows with ledger = outage_requests are the 1,313 requests made during three local network outages, kept apart from the request ledger and not used for any result). Method in full: docs/METHOD.md.

Limitations

• Signals are read as published at the time of the request; sites change their files. Each row carries fetched_at.
• One vantage point (United States). Some hosts block or time out from abroad; 57 of the 198 requested hosts under gov.sa or gov.om were unreachable or timed out from it, and are reported as such.
• Sector labels come from rules over host names and a list of news domains, and country labels from the TLD; neither was checked by a person.
• Page weights come from one Common Crawl crawl, which samples the web and does not collect pages from hosts that block CCBot.
• robots.txt rules are per host; for path-scoped baseline entries (bbc.com/arabic, dw.com/ar) the rule shown is the host's.
• Terms of use, licences and paywalls are not read. A missing signal says nothing about permission.

Claims

Each figure on this card is dated, states its unit and population, and is read by script from the released files; stats.json maps every figure to the row of data/aggregates/2026-10.parquet (or the file and column) it comes from.

Claim	Unit	Population	Source
2.4% block at least one AI-training crawler, as of 31 October 2026	hosts	33,985 hosts checked (reachable hosts), census 2026-10	aggregates: signals / reachable / training_block_any / hosts
2.6% page-weighted	Arabic pages in CC-MAIN-2026-43	same hosts	aggregates: signals / reachable / training_block_any / pages
4.4% reserve AI training by any signal	hosts	same hosts	aggregates: signals / reachable / reserved_any / hosts

Files

data/ (3 configs), exports/exclusion_list_2026-10.csv, bots.yaml (the 26 crawler tokens with operator, purpose and documentation link), universe/ (SQL, Common Crawl file lists, list ids, SHA-256 values), charts/ (SVG and PNG), qa/ (the separate check with protego, and the baseline reading and its comparison), stats.json, PROVENANCE.jsonl, MANIFEST.json, SHA256SUMS, CITATION.cff, LICENSE.

Licence

Data (all tables, the exclusion list, bots.yaml, universe/, PROVENANCE.jsonl): CC0 1.0. This card, stats.json, qa/ and the charts: CC BY 4.0. See LICENSE. Code: Apache-2.0 (GitHub).

Cite

Bayanat Labs (2026). Arabic Web Rights Census, census 2026-10 (version 0.1.0), as of 31 October 2026. https://huggingface.co/datasets/Bayanat-Labs/arabic-web-rights-census. CC0 1.0 (data), CC BY 4.0 (card text and charts).

Corrections and removal

Site operators can ask for a correction or for their host to be removed at hello@bayanatlabs.com (please give the host and the field).

From Bayanat Labs

This dataset is published by Bayanat Labs (بيانات لابز), a Riyadh-based Arabic AI data company. Release page: bayanatlabs.com/open/arabic-web-rights-census. For licensed or custom Arabic data, see Bayanat Search. Related: Arabic Dataset Rights Index (licences of Arabic datasets on the Hub). Questions, corrections and takedown requests: hello@bayanatlabs.com.

بالعربية

The Arabic text has not yet been reviewed by a native Arabic writer; the English governs.

إحصاء حقوق الويب العربي. يسجّل هذا الإحصاء إشارات استخدام الذكاء الاصطناعي المقروءة آليًا لدى 36,190 موقعًا في القائمة: مواقع باللغة العربية من Common Crawl، إضافةً إلى نطاقات شائعة تحت نطاقات المستوى الأعلى لرموز الدول العربية ولرموز الدول المكتوبة بالخط العربي. ويسجّل لكل موقع قاعدة ملف robots.txt لكل واحد من 26 زاحفًا، وقيم Content-Signal، وحجوزات TDMRep، وقيم noai. الإحصاء 2026-10، جُمع بين 2026-10-30 و2026-10-31 من نقطة وصول واحدة في الولايات المتحدة.

حتى 31 أكتوبر 2026، يحظر 2.4% من 33,985 موقعًا فُحصت زاحفًا واحدًا على الأقل من زواحف تدريب الذكاء الاصطناعي في ملف robots.txt (824 موقعًا). ولدى 31,653 موقعًا منها من عيّنة Common Crawl وحدها تبلغ النسبة 1.8% (555 موقعًا). وبالترجيح بعدد الصفحات العربية لكل موقع في زحف Common Crawl ‏CC-MAIN-2026-43 (وهو عيّنة) تبلغ النسبة 2.6%. ويحجز 4.4% (1,511 موقعًا) حق تدريب الذكاء الاصطناعي بإشارة واحدة على الأقل.

قاعدة "السماح" في robots.txt أو غياب الإشارة ليسا إذنًا بإعادة استخدام محتوى الموقع. يرصد الإحصاء الإشارات، لا التراخيص.

تُحسب النسب على المواقع التي أجابت (33,985 موقعًا). المواقع التي تعذّر الوصول إليها (1,170) أو انتهت مهلتها (650) أو أعادت صفحة تحقّق (385) تُذكر وحدها ولا تُعدّ سماحًا. وطُلبت كل المواقع المدرجة في القائمة (36,190 موقعًا).

قائمة الاستبعاد exports/exclusion_list_2026-10.csv تضم 1,511 موقعًا تحجز التدريب بإشارة واحدة على الأقل؛ وهي ليست سجلًا كاملًا ولا استشارة قانونية.

للتصحيح أو طلب الإزالة: hello@bayanatlabs.com