I Tested DeepSeek’s New Search Api, Here’s How It Ranks And Scores Websites

DeepSeek has opened entry to its personal search engine by a public Search and Fetch API. It returns an authority rating for each end result and says it filters out web optimization spam, GEO manipulation, and low-quality AI content material.

That could be a vital launch for SEOs.

The announcement on Oct. 17, 2026 got here alongside an replace to its Flash mannequin. DeepSeek says the search engine already powers search on its web site and thru its API, indexes lots of of billions of pages and handles lots of of tens of millions of searches a day.

We are able to now question that system immediately, examine the sources it returns, and see a number of the scores hooked up to them. That offers us one thing concrete to research earlier than an AI mannequin turns these sources into a solution.

Naturally, I wished to see what it considered my web site. And whether or not the spam filtering held up.

I spent $0.08 discovering out.

My web site got here again with an authority rating of 1. Spam PDFs on authorities and college hosts got here again with authority 4. The Fetch API additionally offered itself as Firefox from Hong Kong and returned content material from URLs disallowed by robots.txt.

These are my preliminary findings.

What I Examined

I ran 110 calls on Oct. 17, 2026 from Dubai: 93 Search calls and 17 Fetch calls. Round 103 have been billable underneath the documentation’s billing rule, costing roughly $0.08.

The checks coated rating distributions, authority throughout completely different hosts, six queries associated to my very own instruments and analysis, localization, date filters, and fetch conduct.

I used the Search and Fetch documentation because the reference. The place the responses differed from the documentation, I’ve saved the noticed outcomes clear.

DeepSeek Provides You 2 Scores

Every search end result contains two fields value taking note of:

These definitions come from the Search API reference. The authority tier is DeepSeek’s personal evaluation, and the documentation describes it as a property of the web page.

Throughout 20 web optimization and GEO queries at rely=50, I collected 1,000 outcomes. Their rerank scores ranged from 0.75 to 2.70, with a median of 1.63. Outcomes have been persistently sorted by that rating, and 68% scored 1.5 or above.

Solely 196 distinct scores appeared throughout these 1,000 outcomes. The returned values seemed bucketed, though that alone doesn’t inform us whether or not the underlying scores are rounded or calculated in fastened steps.

Authority Appeared Tied To The Host. Mine Was 1

Throughout all 4,322 outcomes within the run, authority tiers -1 and 0 by no means appeared. The distribution seemed like this:

Larger authority tiers got here with larger median rerank scores on this pattern. That doesn’t isolate the impact of authority from the opposite rating alerts, however it makes the sphere value investigating.

I then ran 16 web site: probes throughout 16 hosts. On each host, each returned URL had the identical authority rating. Listed below are some examples:

So though the documentation calls this a web page authority tier, it behaved like a bunch rating in my checks.

The Reddit subdomains have been particularly helpful right here: qh.reddit.com scored 1, whereas www.reddit.com scored 2. That factors in the direction of the host being the related unit, slightly than your complete registrable area sharing one rating.

A seek for web site:suganthan.com crammed the 50 end result cap. It returned my homepage, /assets/, notes and weblog posts. DeepSeek was discovering content material throughout my web site. These pages all carried authority 1.

Being listed was clearly solely a part of the story.

I Misplaced My Personal Software Queries To Registries And App Shops (Lol)

I ran six goal queries with rely=50 from the UAE.

Package deal registries, app shops, and market listings did properly on these instrument queries. PyPI, npm, Apify, the Chrome Net Retailer, WordPress.org, and GitHub have been the sorts of hosts surfacing.

My Gemini article and FanoutFox each had authority 1. Their rerank scores of 1.14 and 1.37 additionally fell beneath DeepSeek’s documented 1.5 high-quality threshold.

For my web site, having content material immediately concerning the factor being looked for wasn’t sufficient to beat these hosts. The low authority tier is a believable a part of the reason, though these checks don’t inform us its actual weighting.

The Spam Filtering Nonetheless Let Some Questionable Stuff By means of

DeepSeek explicitly says its engine filters web optimization spam, GEO manipulation, and low high quality AI content material. I wished to see how that declare lined up with the outcomes.

Here’s what nonetheless appeared:

I additionally noticed self-published “frameworks” on zenodo.org obtain authority 4 and rank #one and #three for “generative engine optimization.” That’s one other statement about how authority behaved. Self-publication alone doesn’t make a web page spam.

The federal government and college examples are onerous to sq. with the promise of filtering these items out. These uploads nonetheless acquired the very best authority tier.

Because of this the obvious use of a bunch rating issues. A system that carries a bunch’s authority throughout its pages can find yourself giving parasite content material the identical remedy as respectable materials on that host.

The outcomes don’t inform us how a lot spam DeepSeek eliminated. They do present that these examples bought by, and {that a} excessive authority tier wasn’t a dependable judgement of the person web page’s high quality.

How To Run The Similar Sort Of Search

Here’s a request utilizing an API key saved in an surroundings variable:

curl -s https://api.deepseek.com/v1/instruments/search 
 -H "Authorization: Bearer $DEEPSEEK_API_KEY" 
 -H "Content material-Kind: software/json" 
 -d '{"question":"discover out who Gemini cites to your question","rely":50}'

A trimmed end result from that question:

{
 "url": "https://suganthan.com/weblog/gemini-citations/",
 "rerank_score": 1.14,
 "authority": 1
}

Search and Fetch every price $0.75 per 1,000 profitable requests. Underneath the billing guidelines, failed requests, empty search outcomes, and timeouts aren’t billed. Returning 50 outcomes prices the identical as returning fewer.

There’s a separate non permanent exemption for the server-side Search and Fetch calls already made when utilizing the DeepSeek API by Claude Code or Codex. These stay free till one November 2026 at 00:00 Beijing time, which is 31 October at 8:00 PM GST. Per request, billing begins then. My direct API checks have been billable.

3 Issues To Account For When Testing

1. Set Your Localization

Leaving out localization produced the identical outcomes as specifying a UAE IP throughout all six goal queries. I used to be calling from Dubai, which is per DeepSeek utilizing the caller’s IP because the default.

Set this area explicitly to an IP handle out there you need to take a look at. In any other case, your audit could replicate wherever the script occurs to run.

Utilizing a UK or U.S. IP modified the rerank scores of 26 to 41 of the 50 outcomes and swapped 4 to fifteen URLs. The imply rating delta was small, round ±0.05. It principally reshuffled the outcomes, with out considerably altering the rating scale.

A Chinese language IP, 114.114.114.114, behaved equally to the UK and US settings. I didn’t see a clearly separate Chinese language end result set in these checks, though this doesn’t set up how DeepSeek organizes its indexes.

2. The Unique Query Made Very Little Distinction

The ori_question area helps you to provide the consumer’s authentic query alongside the search question. DeepSeek says this could enhance question understanding.

Throughout my six queries, it modified the rerank rating for less than zero to 2 of the 50 outcomes and swapped at most one URL.

3. Date Filters Eliminated Each Undated Outcome

The after: and earlier than: operators produced zero date violations throughout 4 queries.

Additionally they eliminated each undated web page. 11% of the unfiltered outcomes had date: null, and people URLs disappeared when a date filter was utilized.

That issues in case your content material wants to seem in searches for latest info. Put an acceptable machine-readable date on the web page, then test that DeepSeek truly acknowledges it. A web page it treats as undated can disappear from a date-filtered search.

Each returned date that was current led to T16:00:00Z, which corresponds to midnight Beijing time. The filtering conduct pointed to Beijing calendar days in these checks.

I additionally discovered two variations between the documented and noticed textual content lengths:

That’s value understanding if you’re constructing a instrument round these responses. The summaries I acquired have been significantly longer than the documentation advised.

The Fetch API Arrived As Firefox From Hong Kong

The Fetch requests I inspected didn’t determine themselves as DeepSeek. They used this consumer agent:

Mozilla/5.0 (X11; Linux x86_64; rv:149.0) Gecko/20100101 Firefox/149.0

The Settle for-Language header was:

zh-CN,zh;q=0.9

There was no From header, bot token, Signature or Signature-Enter. The requests offered themselves as Firefox, with no DeepSeek identifier to match within the consumer agent.

5 fetches got here from 5 completely different IPs, all recognized as AS136907, Huawei Cloud, Hong Kong:

121.91.172.54
183.87.40.160
119.8.239.43
27.106.106.73
159.138.131.203

Within the rendering take a look at, it loaded CSS. It did not run JavaScript, load photographs, or fireplace the XHR beacon. Content material that required JavaScript didn’t seem in these fetches.

These observations concern the Fetch API. They don’t set up how the crawler that builds DeepSeek’s search index behaves.

It Returned Content material From URLs Disallowed By Robots.Txt

The Fetch API returned full content material from three URLs disallowed for Person-agent: *:

1. httpbin.org/deny
2. Wikipedia’s Particular:Search
3. google.com/search

I couldn’t affirm whether or not it ever requested /robots.txt. What I may affirm was that the disallowed URLs nonetheless returned content material by the API. These guidelines didn’t forestall the fetches I examined.

What To Look For In Your Logs

In case your entry logs file Settle for-Language, this can assist floor requests matching the noticed headers:

grep -E 'Firefox/149.0' entry.log | grep 'zh-CN'

A extra particular filter combines the Firefox model, language header, and ASN. A Cloudflare customized rule in that form can be:

(ip.src.asnum eq 136907 and http.user_agent incorporates "Firefox/149.0" and any(http.request.headers["accept-language"]
 incorporates "zh-CN"))

Deal with that mixture as a fingerprint to research. It’s primarily based on what I noticed in these requests and will change.

There isn’t any DeepSeek-specific consumer agent right here, and blocking the Firefox string alone can catch respectable browsers. AS136907 can also be a normal Huawei Cloud community. Blocking the entire ASN can have an effect on different tenants, so don’t assume each request from it belongs to DeepSeek.

What I Would Do With This As An web optimization/GEO/AEO/AI web optimization

The helpful a part of this launch is that we will examine the search outcomes and scoring immediately.

Right here is the place I’d begin:

1. Run your precise goal queries towards POST https://api.deepseek.com/v1/instruments/search. Document rank, rerank rating, and authority collectively.
2. Verify your web site’s authority tier with web site:yourdomain.com, together with any related subdomains.
3. Set localization explicitly so repeated checks characterize the market you care about.
4. Verify your dates. Add acceptable machine-readable dates and confirm whether or not DeepSeek returns them. Evaluate outcomes with and with out date filters.
5. Take a look at related distribution to your instruments. PyPI, npm, Apify, the Chrome Net Retailer, WordPress.org, and GitHub carried out properly in these queries. In case your instrument belongs on a kind of platforms, a respectable itemizing is value testing alongside your personal web site.
6. Monitor the fetcher in your logs utilizing the mixed Firefox/149, zh-CN and AS136907 sample, whereas checking matches earlier than treating them as DeepSeek.

For me, probably the most fascinating discovering is the mixture of seen authority tiers, weak rankings for some authentic instrument content material, and apparent spam receiving the very best authority tier. It provides us particular questions we will now take a look at.

This was in the future, 110 calls, six goal queries, and one calling area, with extra probes for scores, hosts, filters, and fetching. I nonetheless couldn’t affirm whether or not the fetcher requested robots.txt. Deal with the numbers as directional, please.

Extra Assets:

This publish was initially printed on Suganthan.

Featured Picture: Fahmi_Ruddin_Hidayat/Shutterstock