Second choice for AI search: What DeepSeek's newly visible source ratings reveal about Instagram, YouTube and Reddit
English translation of the German original published by rethink.expert, Berlin.
How AI search engines weight sources has so far been guesswork. Since mid-October, DeepSeek lets you look it up for the first time: its in-house search engine assigns every source an authority level from 0 to 4, publicly accessible through its developer API. Only from level 3 upwards does DeepSeek consider a source highly authoritative. We measured 6,060 websites, including 74 platforms where companies publish. The result: most major platforms do not reach this threshold. And the level is associated with backlinks and prominence, but hardly with traffic.
Our own study via DeepSeek's Search API and Google, 29 October 2026: 120 search queries, 6,060 websites, 74 platforms.
- →For the first time, you can look up how an AI search engine weights sources: DeepSeek's search assigns every website an authority level from 0 to 4; from 3 upwards, DeepSeek calls a source highly authoritative. We measured 6,060 websites.
- →Most platforms where brands publish never reach that rating: 41 of the 67 that DeepSeek's search knows. Facebook sits at level 1; Instagram, TikTok, YouTube, Reddit and the German press release portals at 2. We found no X, Bluesky or Vimeo pages in the index at all.
- →The level is linked to ranking: no result at the lowest level reached DeepSeek's own high-quality ranking threshold in our measurement; at the highest level, 64% did.
- →High levels go to websites that are well linked and well known, not to those with many visitors: all else being equal, the odds of a higher level are about three times as high with a Wikidata entry, more than twice as high with ten times better linking, and even slightly lower with more visitors. Young domains are rated highly authoritative much less often than old ones.
- →Brand websites are solid but rarely first choice: almost all main sites of Germany's DAX and MDAX companies sit at level 3. For almost one in four brands, one of their own subdomains, for example for a blog, shop or press room, is rated lower than the main site.
- →Google and DeepSeek see sources differently: a third of the websites in Google's top 3 sit at level 2 or lower at DeepSeek. And US government websites are almost always at the top: 96% of their results sit at level 4, compared with 74% for German ones.
An AI search engine shows its ratings
On 17 October 2026, DeepSeek opened its search and fetch API. Developers can use it to access the same search engine that, according to DeepSeek, also powers the company's web search. Every result comes with two values: a ranking score from 0 to 3 which, according to the documentation, combines relevance, authority and freshness, and an authority level from 0 to 4. From level 3 upwards a source counts as highly authoritative; from a ranking score of 1.5 upwards a result counts as high quality. How the level is calculated, DeepSeek does not disclose. Other search APIs return relevance scores, but to our knowledge no AI search engine has made an explicit authority rating publicly accessible before.
“For GEO, generative engine optimization, this opens a new window: until now, you could only infer from finished answers which sources an AI search engine prefers. Here, a system itself shows how it rates a website as a source – before an answer is even generated.”Dr. Michael Mohaupt, rethink.expert
Important for everything that follows: the level does not apply to individual pages but to the whole website, more precisely to the hostname. A press release on a press release portal therefore carries the same level as everything else there, however good it is. And de.linkedin.com and www.linkedin.com count as two different websites.
This rating does not only affect DeepSeek's own users. It is built into every place where DeepSeek's search engine supplies the sources:
- in DeepSeek's web search and in every app that integrates DeepSeek's Search API,
- in Claude Code and Codex when you run them with DeepSeek as the provider instead of Anthropic or OpenAI: according to DeepSeek, their web search then runs through its Search API.
And DeepSeek is in demand among developers: on OpenRouter, the largest routing platform for AI models, it has been the most used model since mid-May; in the week from 26 October, almost 30% of all requests there went to DeepSeek models. Anyone using the model directly through DeepSeek's API gets the web search along with it. The more agents are built this way, the more answers rely on sources that have passed through DeepSeek's rating. Which sources an agent finds in its web search is therefore determined not by the model but by the search behind it (“Same model, different sources”, in German).
6,060 websites, 74 platforms
On 29 October, we sent 120 search queries to DeepSeek's Search API, 60 each in German and English on six topics from finance and health to travel, with German and US locations. Per topic and language there were five informational queries such as “cold symptoms” and five purchase queries such as “car insurance calculator”. We drew the queries by fixed rules from frequent Google searches rather than picking them by hand. We added the same queries on Google, plus 74 platforms where companies publish, and the websites of the companies in Germany's DAX and MDAX stock indices and of 167 brands. For all these websites we recorded the level and compared it with freely available measures of linking, visitor numbers, prominence and age. Plan, hypotheses and analysis rules were fixed before the measurement.
We are not the first to look at these values. Suganthan Mohanadasan analysed around 4,300 results in the first few days and showed that the level applies per hostname – and that spam files on government and university websites also get level 4. Hanns Kronenberg found the level stable across 100 identical requests, and Tom Brigl described first observations on platforms and on content loaded via JavaScript. OGTool saw a clear link between level and backlinks across 87 brand websites. Our study builds on this and adds three things:
- Four explanations compared: for every website we recorded linking, visitor numbers, prominence and age. Only then can you tell whether reach counts or linking.
- Real search results: the websites come from DeepSeek's answers to 120 search queries, that is, from what users actually get to see.
- The places where brands publish: 74 platforms from Instagram to press release portals, measured specifically.
Why the authority level matters
The level is not just a label. DeepSeek states that authority feeds into the ranking score of every result, and our data fit this: not a single one of 200 results at level 1 reached the ranking threshold above which DeepSeek calls a result high quality – at level 4, 64% did. Results at level 4 are also more often near the top: 40% of them landed among the first ten results of a search, twice as often as the average of all results (20%). At level 1 it was only 8%.
Note: this analysis was not planned in advance; we added it after reviewing DeepSeek's description of the ranking score.
At the same time, level 3 is the norm in DeepSeek's results – which already makes a source highly authoritative. Level 4 is the exception:
Websites at level 1 or 2 are therefore a minority, and their results almost never reach DeepSeek's high-quality ranking threshold. That is exactly where most platforms where brands publish sit.
Instagram, YouTube, Reddit: second choice
Of the 67 platforms that DeepSeek's search knows, 41 sit at level 1 or 2. These are almost all the places where companies can publish themselves without an editorial check. The remaining 7 of the 74 platforms we measured did not return a single result of their own for our “site:” query, among them X, Bluesky and Vimeo:
The exception is the major international newswires: PR Newswire, Business Wire and GlobeNewswire sit at level 3 or 4, the German press release portals at level 2. The same press release is therefore rated very differently depending on the distributor.
Two details stand out. LinkedIn is rated by subdomain: www.linkedin.com sits at level 3, country sites such as de.linkedin.com at level 2. LinkedIn serves German-language posts under de.linkedin.com; in our sample, DeepSeek listed 35 of 57 of them there and 22 under www. The same post can therefore carry level 2 or 3, depending on the address under which it was indexed. And on blog platforms every blog gets its own level: blogs on Blogspot mostly sit at level 1, on Substack and WordPress.com mostly at level 2.
“Many brands rely on profiles, press release portals and platforms for their most important statements. For DeepSeek, exactly these places are mostly second choice. The question is no longer only where a message gets reach, but where AI search engines take it seriously as a source.”Dr. Michael Mohaupt, rethink.expert
Well known beats well visited
What is the rating linked to? We compared it with freely available measures of linking, prominence, age and visitor numbers. The link is strongest with references from other websites and weakest with actual visitor numbers:
Linking: measure fixed in advance, from the open Common Crawl web graph. A second link index (Majestic) points in the same direction at 0.46 but covers only part of the websites.
Looking at all measures together, three keep weight of their own: with a Wikidata entry, the odds of a higher level are about three times as high; with ten times better linking, more than twice as high; per ten years of domain age, one and a half times as high. Higher visitor numbers, all else being equal, even go along with slightly lower odds:
Ordinal logistic model, 3,589 websites from DeepSeek's results; 1x = no difference. Association, not cause.
73% of the websites at level 4 belong to an organisation with a Wikidata entry; at level 1 it is 3%. The big platforms show that reach alone is not enough: YouTube, Instagram and TikTok are enormously visited and well linked – and still sit at level 2, Facebook even at level 1. An obvious explanation is that anyone can publish there; our data cannot prove it. A single case from our pilot study fits this: at MIT, the business magazine sloanreview.mit.edu sits at level 4, while the learning platform scratch.mit.edu, where children publish their own projects, sits at level 2 – even though Scratch has an estimated 1,000 times more visitors and even more backlinks.
Large German portals also sit low: Kleinanzeigen (classified ads, formerly eBay Kleinanzeigen) and gutefrage.net (Q&A), where users post content themselves, but also the comparison portal CHECK24. All three are among the most visited websites in Germany, are well linked and sit at level 2. The comparison portal idealo, by contrast, sits at level 3. What DeepSeek distinguishes them by, our data do not show.
The age of the domain also carries weight of its own – not only because older domains are usually better known and better linked: all else being equal, the odds of a higher level are one and a half times as high per ten years. Overall, the difference is large: of the domains online only since 2020, just over half (53%) are rated highly authoritative; of the domains online before 2005, 91%. For Google, by its own account, domain age is not a ranking signal. Anyone launching a separate domain for a campaign or a new product is therefore likely to start low at DeepSeek. Whether a domain moves up over the years, we did not measure.
“The pattern is clear: weight goes to those who are linked and known. Optimising for reach alone misses this rating.”Dr. Michael Mohaupt, rethink.expert
Brand websites: solid, but rarely top
What this means for brands is shown by the companies in the DAX and MDAX, Germany's two leading stock indices. Of the 84 main sites that DeepSeek's search knows, none sits at level 1 or 2, but only two at level 4 (Bayer and Deutsche Börse) – almost all sit at level 3. Government websites with the same linking and prominence are at level 4 far more often, advice and comparison portals considerably less often than brands. For DeepSeek, a brand website is solid but not the first address. And not every address of your own is equal: for almost one in four brands on our list (39 of 163), we found at least one of their own subdomains, for example for a blog, shop or press room, that is rated lower than the corresponding main site; among the DAX companies, 3 of 39. Three MDAX companies we did not find in DeepSeek's search at all: for the domains of Aroundtown, LEG Immobilien and TKMS, the query did not return a single result.
If www.company.com sits at level 3 but the blog at blog.company.com at level 2, every blog post counts as less authoritative for DeepSeek than the website itself. At www.company.com/blog it would carry the level of the main site, because the level applies per hostname.
Whether DeepSeek rates a moved section that way right away, we did not test.
Country domains are a different story: in a small preliminary survey of 26 global brands that run both a .de and a .com domain, 23 sat at the same level on both, mostly level 3. German content under brand.com/de/ carried the level of the .com in all 26 cases we checked. Whether a brand appears in Germany under .de or .com therefore makes hardly any difference to DeepSeek.
A top spot on Google is no protection
For the same 120 queries, we also recorded the Google results and checked which level DeepSeek gives those websites. A third of the websites in Google's top 3 sit at level 2 or lower at DeepSeek, and DeepSeek's search does not know 6% of the Google results at all. In DeepSeek's own top 10, by contrast, only 7% of results sit at level 2 or lower:
The top level stands out most: at DeepSeek, more than one in four of the first ten results comes from a level-4 website; at Google, fewer than one in ten. Within Google's top 10, the DeepSeek level makes hardly any difference: websites that DeepSeek rates at level 1 or 2 rank fifth on average at Google, websites at level 3 or 4 only about half a position higher on average. And the overlap is small: of the 710 different websites in Google's top 10, 498 do not appear among DeepSeek's first 50 results for the same queries.
Ranking high on Google can therefore still mean being rated as not very authoritative at DeepSeek. This fits a finding we have already seen with other AI systems: YouTube and Reddit are among their most cited sources (“Almost one in three AI citations”, in German) – at DeepSeek, both sit at level 2. Every system has its own idea of which source counts.
US government websites almost always at the top
For the informational queries (such as “cold symptoms”), we compared the first ten results for German and US locations. In the US, 54% come from sources at the highest level, in Germany 42%. This difference is not statistically significant – our expectation, stated in advance, that German results sit at level 4 less often was therefore not confirmed. What is revealing is where the gap comes from. There are two possible reasons: either DeepSeek returns different kinds of sources in Germany, for example fewer government and more company websites. Or the same kinds of sources are rated differently. Our analysis shows: it is mainly the rating. The mix of sources is almost the same in both countries, with government websites supplying just under 40% of results each. What differs is the rating, and almost only for government websites:
Note: the breakdown by type of source was not planned in advance and is purely descriptive. Purchase queries are not included; there, only few results come from sources at the highest level anyway (6% in Germany, 10% in the US).
For companies and editorial media, both countries are close; there is no sign of a disadvantage for German company websites. Across all websites, too, US websites are not rated measurably higher with the same linking and prominence. The location of the search is not the reason: in a pre-test, the level of 40 websites was identical in all cases for requests without a location and from Germany, the US and China. It is the websites themselves that are rated. Why German government websites are less often at level 4, our data cannot answer: a clear effect of the domain ending (such as .gov versus .de) does not show once linking is taken into account.
Claim: “Websites with many visitors are also authorities for DeepSeek.”
Assessment: No. Visitor numbers from Chrome are barely related to the level (rank correlation 0.08); linking is related much more strongly (0.39). With equal linking, prominence and age, websites with more visitors even tend to sit slightly lower. That does not mean visitors hurt – but reach alone does not go hand in hand with a higher level.
Claim: “If you rank high on Google, DeepSeek trusts you too.”
Assessment: Not necessarily. 33% of the websites that Google placed in its top three for the same queries sit at level 2 or lower at DeepSeek, and DeepSeek's search does not know 6% of all Google results at all. In DeepSeek's own top 10, by contrast, 92% sit at level 3 or 4.
Claim: “The level rates individual pieces of content.”
Assessment: No. The level applies to the whole website (more precisely: the hostname) and is the same for all content on it, whether press release, profile or expert article. Across 6,060 websites it was inconsistent only twice, and all 778 websites from our pilot study a few days earlier had the same level.
Claim: “A Wikidata entry raises your level.”
Assessment: Our data do not show that. Websites with a Wikidata entry are more often at the top, but the entry is more likely a sign of prominence than its cause. Wikidata only accepts organisations with verifiable notability; entries without such evidence can be deleted. What makes sense is keeping an existing entry correct, for example with the right official website.
Claim: “DeepSeek's chat selects its sources by this level.”
Assessment: That has not been established. DeepSeek writes that its web search and the Search API are powered by the same in-house search engine. Whether and how the level is weighted in the chat, DeepSeek does not say. What is established is the link with the order of the search results themselves.
Check your own channels: the Search API returns the level for any website (a search with “site:” and the address is enough), at 0.75 US dollars per 1,000 requests. This shows how DeepSeek rates your own website, your subdomains such as blog.company.com or press.company.com, and the platforms you use. Subfolders such as company.com/blog carry the same level as the website itself. For almost one in four brands on our list, at least one of their own subdomains sat lower than the main site.
Publish key statements where they carry weight: studies, prices, terms and official statements belong on your established website rather than on new microsites and, where it fits, in media and trade portals at level 3 or 4. Profiles and platforms remain important for reach – as the only place for core statements, they carry less weight with DeepSeek.
Show prominence rather than reach: links from other websites, mentions in the media and a correct entry in open knowledge bases are associated with higher levels. These are not switches but the result of prominence – incorrect or missing information can, however, be checked and corrected with evidence.
Monitor Google and AI search separately: a good position on Google says little about how an AI search engine rates the source. If you measure AI visibility, look at the sources per system.
DeepSeek's authority level
A value from 0 to 4 that DeepSeek's Search API returns with every result (plus −1 for unknown); in our measurement only levels 1 to 4 occurred. From level 3 upwards, DeepSeek considers a source highly authoritative. The level applies to the whole hostname and was very stable in our measurement. How it is calculated, DeepSeek does not disclose.
41 of the 67 platforms that DeepSeek's search knows sit at level 1 or 2: Facebook at 1; Instagram, TikTok, YouTube, Reddit, Trustpilot, kununu, de.linkedin.com and the German press release portals at 2. At level 3 sit www.linkedin.com, Wikipedia, GitHub, guest post portals such as t3n and the international newswires.
Almost all main sites of the DAX and MDAX companies sit at level 3, only two at level 4 (Bayer and Deutsche Börse). For almost one in four brands on our list, at least one of their own subdomains such as blog.company.com sits lower than the main site. Subfolders such as company.com/blog, by contrast, carry the level of the main site.
Most strongly to linking and prominence: websites with many links, a Wikidata entry for the organisation and an older domain are more often at the top. The level is hardly related to visitor numbers. These are associations, not proven causes.
DeepSeek's own web search, apps using DeepSeek's Search API, and Claude Code and Codex when run with DeepSeek as the provider instead of Anthropic or OpenAI. Apps that only use the DeepSeek model and get their sources from another search engine are not affected. Whether the chat uses the level to select sources, DeepSeek does not say.
With a “site:” search plus the hostname via DeepSeek's Search API; every result includes the level. Since 1 November, the API costs 0.75 US dollars per 1,000 successful requests.
Want to know how AI search engines rate the places where you publish? Get in touch to talk about measuring and improving your visibility in AI answers.
Get in touch →