DEEPSEEK AUTHORITY LEVELS OF 6,060 WEBSITES
Popularity measures and source classification

Publisher:   rethink.expert (Dr. Michael Mohaupt)
Version:     3 Nov 2026 (measured on 29 Oct 2026)
Licence:     CC BY 4.0 for our data collection and classification; values
             taken from third-party sources remain under their own
             licences (see Sources)
Article:     https://www.rethink.expert/en/insights/deepseek-authority-platforms/
             German original:
             https://www.rethink.expert/insights/deepseek-autoritaet-plattformen/
Contact:     info@rethink.expert

This is a translation of the German README (deepseek-autoritaet-2026-10-
README.txt). In case of doubt, the German version is authoritative. Column
names and coded values in the CSV files are German; this file explains them.
The study plan and protocol (studienplan-und-protokoll.md) is available in
German only.


WHAT THESE FILES ARE
--------------------

Since October 2026, DeepSeek's Search API returns two values for every
result:

  authority      Authority level of the hostname. According to DeepSeek's
                 documentation 0 to 4, -1 = unknown; from 3 upwards a source
                 counts as "highly authoritative". We only observed 1 to 4.
  rerank_score   Ranking score 0 to 3 which, according to the
                 documentation, combines relevance, authority and freshness;
                 from 1.5 upwards a result counts as high quality.

We collected the authority level for 6,060 hostnames and linked it to
freely available measures of linking, usage, prominence and age. For 812
hosts we also classified who runs them (publisher type) and who writes the
content (content origin).

The hosts come from four sources:
  suche          DeepSeek's top 50 for 120 search queries (60 German,
                 60 US)
  google_hosts   hosts from Google's organic results for the same queries
                 that DeepSeek did not return itself; level obtained via a
                 "site:" query to DeepSeek
  plattformen    74 platforms fixed in advance (social networks, video,
                 blog hosts, press release portals, review sites etc.)
  marken         DAX, MDAX and 167 brands; all hosts found via
                 "site:<domain>"

Plan, hypotheses and rules were fixed before the measurement. The plan
including all later deviations is enclosed (studienplan-und-protokoll.md).


FILES
-----

Every table comes in two versions with identical content:

  <name>.csv            UTF-8, comma as separator, point as decimal mark.
                        For statistics software, Python, R.
  <name>-excel-de.csv   For German-language Excel: UTF-8 with BOM,
                        semicolon as separator, comma as decimal mark.

The files contain umlauts and titles in other languages (UTF-8).

  ...-hosts              6,060 hosts with level and all measures (core table)
  ...-treffer            6,000 search results (120 queries x 50)
  ...-anfragen           the 120 queries
  ...-site-abfragen      results of the "site:" queries, per target and host
  ...-nicht-im-index     targets for which DeepSeek returned nothing
  ...-markenindex        DAX, MDAX and brands with the domain queried
  ...-einordnung         publisher type and content origin for 812 hosts
  ...-kontrolle          second classification of a sample of 40 hosts
  ...-studienplan-und-protokoll.md
                         study plan (fixed before the measurement) with all
                         later deviations (German)
  ...-nachrechnen.py     Python script that recomputes the study's key
                         figures from these CSV files


COLUMNS: HOSTS (hosts)
----------------------

host, domain
  Hostname and registrable domain.

stufe (level)
  DeepSeek's authority value (1-4). Applies per hostname, not per page.

stufe_abweichend (level deviating)
  Number of the host's results with a level other than the most frequent
  one (greater than 0 for 2 of 6,060 hosts).

gefunden_in (found in)
  How we found the host:
    suche          among DeepSeek's results for our 120 queries
    google_hosts   via the "site:" query to DeepSeek for a host from
                   Google's results for the same queries
    plattformen    via the "site:" query for one of the 74 platforms fixed
                   in advance
    marken         via the "site:" query for a DAX, MDAX or brand domain
  Besides the target itself, a "site:" query often returns further hosts,
  mainly subdomains (e.g. individual blogs on a blog platform). This is why
  far more than 74 hosts carry the value "plattformen".
  A host may have been found in several ways; the values are then separated
  by semicolons (e.g. "suche;marken").
  Hosts per source: suche 3,673, google_hosts 545, plattformen 254,
  marken 1,671 (multiple sources possible; 6,060 hosts in total).

n_treffer (number of results)
  Number of results for the host across all requests.

crux_klasse_de, crux_klasse_us, crux_klasse_global
  Chrome UX Report 09/2026, rank bucket of the origin by usage in Germany,
  the US and worldwide (1000 = among the 1,000 most visited, 5000, 10000,
  50000, 100000, 500000, 1000000). Empty = not listed. Where there are
  several origins (http/https), the best bucket counts.

tranco_rang (Tranco rank)
  Rank of the domain in the Tranco list of the measurement day. Empty = not
  listed.

majestic_rang, majestic_refsubnets
  Rank and referring IP subnets of the domain in the Majestic Million of the
  measurement day. Empty = not listed.

commoncrawl_pagerank_rang, commoncrawl_harmonic_rang
  Rank of the host in the Common Crawl host graph cc-main-2026-aug-sep-oct
  (PageRank and harmonic centrality; 1 = best linked). Empty = not included.

wikidata_host, wikidata_domain
  Wikidata item whose official website (P856) points exactly to the host or
  to the domain. Empty = no match (a lower bound, see Limitations).

erster_archiveintrag_domain (first archive entry of the domain)
  Date of the domain's first entry in the Internet Archive (Wayback
  Machine). "gesperrt" = operator blocks the archive, "kein Eintrag" = no
  entry, "Fehler" = request failed.

absendertyp, inhaltsherkunft (publisher type, content origin)
  Only for the 812 classified hosts, otherwise empty. Codes and selection:
  see "Columns: classification and control".


COLUMNS: SEARCH RESULTS (treffer)
---------------------------------

messzeit_utc, anfrage_id, land, rang
  Time of the request (UTC), query ID (German and US versions share the
  same ID), country (de/us), position 1-50.

url, host, titel, datum_quelle (url, host, title, source date)
  As returned by DeepSeek. Campaign and recipient parameters were removed
  from the URLs (e.g. utm_*, gclid, newsletter IDs); otherwise unchanged.
  Text excerpts (snippets) are not included.

stufe, rerank_score
  authority and rerank_score as returned by DeepSeek (rerank_score rounded
  to 4 decimal places).

weiterleitung (redirect)
  "ja" (yes) for redirect URLs (e.g. linkedin.com/safety/go), otherwise
  "nein" (no). For these, DeepSeek apparently rates the target; these
  results are excluded from the analysis.

absendertyp, inhaltsherkunft
  Classification of the host, if available. All top-10 results of the
  informational queries are classified.


COLUMNS: QUERIES (anfragen)
---------------------------

anfrage_id, land, thema, typ, anfrage (ID, country, topic, type, query)
  Six topics (finance, health, travel, technology, energy and household,
  mobility), 5 informational ("Info") and 5 purchase-oriented ("Kauf")
  queries per topic and country. German queries were drawn by rule and at
  random from Google's related searches; the US version has the same
  intent, it is not a literal translation.


COLUMNS: SITE QUERIES, NOT IN INDEX, BRAND INDEX
------------------------------------------------

site-abfragen: modul, ziel, host, stufe, n_treffer
  (module, target, host, level, number of results)
  For every target ("site:<target>", up to 50 results) the hosts returned,
  with level and number of results. Deliberately without URLs: for
  platforms such as Instagram or LinkedIn these would be profiles of
  private individuals. A target may return hosts that do not belong to it.

nicht-im-index: modul, ziel, grund (module, target, reason)
  "keine Treffer" (no results) or "nur andere Hosts" (only other hosts:
  neither the host nor its www/apex variant appeared).

markenindex: gruppe, name, domain, quelle (group, name, domain, source)
  DAX and MDAX according to Wikipedia (table as of 22 Jun 2026, retrieved
  on the measurement day), website according to Wikidata P856, manually
  corrected where necessary. Plus 167 well-known brands with their domains,
  depending on the brand for Germany, Austria, Switzerland, the UK,
  Ireland, the US and/or a global domain (gruppe "Marke de", "Marke at",
  "Marke ch", "Marke uk", "Marke ie", "Marke us", "Marke global"). The list
  comes from another rethink.expert study on Google's AI Overviews in
  searches for brand names (German):
  https://www.rethink.expert/insights/ai-overviews-markensuche/

  The column quelle states where the domain comes from:
    "Wikidata P856"         official website according to Wikidata
                            (DAX/MDAX)
    "manuell korrigiert"    manually corrected: Wikidata returned no or an
                            unsuitable website (DAX/MDAX)
    "Markenliste ai-overviews-markensuche"
                            from the brand list of the study linked above


COLUMNS: CLASSIFICATION AND CONTROL (einordnung, kontrolle)
-----------------------------------------------------------

einordnung: host, absendertyp, inhaltsherkunft, stichprobe (sample)
  stichprobe "stufen_stichprobe" = all hosts at level 1 and 100 random
  hosts each at levels 2-4 from the search (445 hosts);
  "info_top10" = all remaining hosts among the top-10 results of the
  informational queries (367 hosts, added later, see protocol).
  Classified blind, without knowledge of the level, based on the legal
  notice, self-description and sample results.

  Why not all 6,060 hosts: classification is manual work and is only
  needed for two additional questions (publisher type at equal popularity;
  Germany/US comparison). Hypotheses H1 to H5 do not depend on it.

  Caution for your own analyses: the "stufen_stichprobe" is stratified by
  level, and level 1 is heavily over-represented in it (all 145 instead of
  roughly 4 % of all hosts). Shares such as "x % of websites are brands"
  cannot be read from it directly; this would require weighting by level.
  The "info_top10" hosts, by contrast, fully cover all top-10 results of
  the informational queries.

kontrolle: host, stufe, absendertyp_ki, absendertyp_mensch,
           inhaltsherkunft_ki, inhaltsherkunft_mensch
  The first classification was done by an AI (Claude, "ki"), the second by
  a human ("mensch"), both blind. Agreement on publisher type 31 of 40
  (78 %, Cohen's kappa 0.73), on content origin 85 % (kappa 0.29; almost
  all hosts RED, so of limited value).

Publisher type (absendertyp)
  OEF   public authority, university, public research institution
  OERM  public service broadcaster / public media
  PRM   private editorial media (publisher, trade magazine)
  MAR   brand, manufacturer, provider, retailer (own products/services)
  VGL   comparison, test or affiliate portal, directories of third-party
        offers
  COM   community, forum, Q&A, video platform
  RAT   other private advice and information portal
  WIKI  open encyclopedia
  MIX   mixed or unclear (associations, chambers, foundations, statutory
        health insurers, commissioned portals, unclear legal notice)

Content origin (inhaltsherkunft)
  RED     operator or editorial team
  NUTZER  mainly user content (including press release portals, guest
          posts)
  GEM     mixed, including archives and aggregators of third-party copies

Full rules and clarifications: studienplan-und-protokoll.md (German).


LIMITATIONS
-----------

One system, one measurement day
  All values apply to DeepSeek's Search API on 29 Oct 2026. DeepSeek can
  change the signal at any time. Whether DeepSeek's chat uses the same
  values to select sources has not been established.

Association, not cause
  The popularity measures show what the level is associated with, not how
  DeepSeek computes it.

Coarse and incomplete measures
  CrUX only measures Chrome usage, in coarse buckets. Tranco, Majestic and
  the archive date apply per domain; subdomains inherit the value. Matching
  Wikidata via the official website misses organisations. The first
  archive entry is not the registration date.

Selection
  The hosts from the search are DeepSeek's own selection for 120 queries,
  not a random sample of the web. The platform and brand lists were fixed
  in advance, but chosen deliberately.

Listing is not rating
  The tables contain all hosts measured, including spam and advertising
  sites. Listing a host is neither a rating nor a recommendation.

Not included
  Google's result lists and the Google traffic and backlink values we used
  for cross-checks come from a commercial data provider and are not part
  of this dataset. The results of these cross-checks are reported in the
  article and in the protocol.
  Also not included are the data of the pilot study and the pre-tests
  (location, country domains) that the article mentions in addition; they
  are described in the protocol and available on request.


SOURCES AND ATTRIBUTION
-----------------------

DeepSeek Search API, documentation "Search & Fetch":
  https://api-docs.deepseek.com/guides/search (retrieved 1 Nov 2026)
Chrome UX Report (Google), CC BY 4.0:
  https://developer.chrome.com/docs/crux
Tranco: Le Pochat et al. (2019), Tranco: A Research-Oriented Top Sites
  Ranking Hardened Against Manipulation, NDSS. https://tranco-list.eu
Majestic Million, CC BY 3.0: https://majestic.com/reports/majestic-million
Common Crawl Web Graph cc-main-2026-aug-sep-oct:
  https://commoncrawl.org/web-graphs
Wikidata, CC0: https://www.wikidata.org
Internet Archive Wayback Machine: https://web.archive.org


RECOMPUTING
-----------

  python deepseek-autoritaet-2026-10-nachrechnen.py

The script only reads the CSV files in this package and outputs the key
figures for hypotheses H1 to H5, the ordinal model (F1), the Germany/US
decomposition (F5) and the additional analysis of the ranking score. Same
method and same random seed as the original analysis; the results match
the article. The ordinal model requires numpy and statsmodels. Output
labels are German. The scripts for the data collection itself (requests to
DeepSeek) are available on request.


SUGGESTED CITATION
------------------

Dataset:
  rethink.expert (2026). DeepSeek-Autoritätsstufen von 6.060
  Websites: Popularitätsmaße und Einordnung der Absender [DeepSeek
  authority levels of 6,060 websites: popularity measures and source
  classification; dataset]. Measured 29 Oct 2026.
  https://www.rethink.expert/daten/deepseek-autoritaet-2026-10.zip
  Licence: CC BY 4.0.

Related article:
  rethink.expert (2026). Second choice for AI search: what DeepSeek's newly
  visible source ratings reveal about Instagram, YouTube and Reddit.
  https://www.rethink.expert/en/insights/deepseek-authority-platforms/
  German original:
  https://www.rethink.expert/insights/deepseek-autoritaet-plattformen/
