For years, search engines told us authority mattered and refused to show us the score. DeepSeek just put one in an API response, on a scale from 0 to 4.
A recent piece on Search Engine Journal covered DeepSeek opening its in-house search engine to developers on October 17. Most coverage will focus on the price. I think the interesting part is the fields.
What DeepSeek actually exposed
Every result from the new Search API carries an authority tier from 0 to 4, with -1 meaning unknown. Pages at tier 3 or above are described as "highly authoritative". There is also a rerank score from 0 to 3 that mixes relevance, authority and freshness, where 1.5 or more counts as high quality.
The scale is serious. According to DeepSeek's changelog, the engine indexes hundreds of billions of pages and handles hundreds of millions of search requests a day. Search and Fetch each cost $0.75 per 1,000 successful requests, and a site: query can return up to 50 of your pages, each with its own tier.
Then there is the line that should make every marketer pause. The changelog says the engine filters SEO spam, low-quality AI-generated content and "GEO manipulation". It does not define the term or explain how detection works.
Authority was always the real variable
I have spent most of my career watching people try to fake authority. Link networks in the affiliate years, then guest post farms, then review schemes. Each wave worked until the engine learned to discount it.
Generative search is following the same arc, only faster. In May, Google said its spam policies also apply to generative AI responses in Search. Now a second major engine says, in writing, that it filters attempts to manipulate AI citations. The window for shortcuts is closing before most brands even opened it.
I wrote recently that buying AI citations is the new link farm. DeepSeek's changelog is the first time an engine has said so directly. Expect the others to follow, because a model that cites manipulated sources produces wrong answers, and wrong answers are the one thing AI companies cannot afford.
Three things the tier tells you
First, authority is assessed per page, not only per domain. A strong brand site can still have pages sitting at tier 1. That matches what I see in practice: the homepage gets trusted while the pricing page, the documentation and the comparison pages get ignored.
Second, freshness is part of the rerank. A page that was authoritative two years ago and never updated slowly loses its place. Content maintenance is no longer housekeeping. It is part of the ranking input.
Third, the model reads a different version of your page than people do. Each result returns a 100 to 200 character display snippet plus a 600 to 800 character summary written for a model. If your key facts are not in the first few hundred characters a machine can extract, they may never reach the answer.
Cheap search changes who builds answers
The price matters more than it looks. At $0.75 per 1,000 requests, any startup can now give its own AI product a full web index without building one. That means more assistants, more agents and more answer layers pulling from the same pool of sources.
Every one of those products inherits DeepSeek's view of who is authoritative. A low tier is no longer a problem in one engine. It travels into every tool built on top of that engine, most of which you will never hear about.
This is how infrastructure works. When Google sold almost all the search traffic, you optimized for Google. When ranking is sold wholesale through an API, your reputation gets syndicated, good or bad, into places you cannot see.
What I would not do with it
The temptation is obvious. Query your own domain, sort by tier, and start optimizing for the number. That is the exact behaviour the filter is designed to catch.
There are also open questions worth respecting. DeepSeek's privacy policy says it integrates third-party search APIs and shares query keywords with them, and the launch pages do not identify which crawler builds the index or what user agent Fetch uses. Before pushing sensitive queries through any external API, know where they go.
The better use is diagnostic. If one engine scores your pages, the pattern probably says something true about how all of them see you. Weak tiers on your most commercial pages point to thin content, missing facts or no independent references. Those are fixable, and fixing them helps with every engine at once.
The lesson I take is simple. AI engines are converging on the same question: does the rest of the web treat this source as trustworthy, and does the page itself contain the answer? You cannot game your way into a yes for long.
What you can do is track it. If you want to see where your brand actually stands across ChatGPT, Gemini, Perplexity and others, GEOflux.ai (geoflux.ai) is built specifically for this. It maps not just whether your brand is mentioned, but which sources drive the mention and what to do about the gaps. I covered the underlying problem in why AI search won't cite you.
The score was always there. Now it is starting to come with a number attached.