AI content ranking
How an AI system orders candidate sources before writing an answer, why it is a selection problem rather than a results page, and where each stage of the ordering fails.
Karl-Gustav Kallasmaa, Founder & CEOLast updated AI content ranking is the ordering an AI system applies to candidate sources before it writes an answer — deciding which passages enter the model's context and which of those get cited. It is a selection step inside a generation pipeline, not a results page a user sees.
The consequence of that difference is the useful part. A search results page shows ten ordered links and lets the reader choose. A generated answer shows a paragraph and, at most, a handful of supporting links. The winner-take-most shape is not a policy; it is what happens when the output is prose of a fixed length.
Eligibility and selection are different questions
Google separates the two explicitly. Its documentation on AI features states that there are no additional requirements to appear in AI Overviews or AI Mode and no special optimisations necessary, and that to appear as a supporting link a page must be indexed and eligible to be shown in Google Search with a snippet. It also says plainly that meeting every requirement does not mean the content will be crawled, indexed or served.
So eligibility is the ordinary index. Selection — which eligible thing is actually used — is where the ordering happens, and it is applied to a question, not to a topic. The same document notes that these systems display a wider and more diverse set of links than a classic web search, which is what you would expect from per-question passage selection rather than per-topic authority.
What actually gets ordered
In a retrieval pipeline the ranked unit is a passage. A corpus is chunked, each chunk is turned into a vector by an embedding model, the vectors nearest the query are retrieved, and the highest-scoring ones are placed in the context window alongside the question.
Four stages, four independent orderings:
- Candidate generation. A search or vector lookup produces the pool. A page absent here cannot be ranked, however good it is. This is where crawler access decides everything — see crawling and indexing.
- Relevance scoring. Vector similarity handles paraphrase well and exact identifiers badly. A part number or version string can be missed while a topically similar page scores higher.
- Budget selection. Only some retrieved passages fit. Anthropic's documentation notes that a simple factual query typically uses one to three searches while comparative research can use ten or more, and that results may be filtered by code before reaching context at all. More retrieval does not mean more of you.
- Citation choice. The model writes, then attributes. A passage can shape an answer and not be cited, and a cited passage need not be the one the sentence came from.
Treating these as one "ranking" is the most common analytical mistake in the field, because the fixes are different at each stage and only the first two are meaningfully influenced by publishing.
Failure modes
- Optimising the page instead of the passage. A well-structured page whose answer is spread over four sections loses to a narrower page that answers the question in one. The retrieved unit does not carry the rest of your article with it.
- Leaving the qualifier upstream. A caveat two paragraphs above the claim does not travel with the claim. State the limit in the same sentence as the number.
- Assuming rank transfers. Position on a results page is a different ordering from passage relevance to one question. High classic rank helps mainly by making you a candidate.
- Blocking preview and expecting citation. Google documents nosnippet, data-nosnippet, max-snippet and noindex as controls that limit what can be shown from a page in Search, and therefore in AI features. A page that forbids a snippet has opted out of being quoted.
- Chasing volume. Generating many pages to cover more query surface runs into Google's scaled content abuse policy, which targets pages generated primarily to manipulate rankings rather than to help users.
How to act on it
Write sections that are correct when lifted out of the page. Put the direct answer in the first sentence under a heading phrased as the question. Keep identifiers — versions, codes, model names, prices with their date — as literal text a lexical matcher can find. Make the page cheap and fast to fetch and render server-side, because a slow page can fall out of the candidate pool before relevance is ever computed. And check what is being said about you: a confidently wrong sentence sourced to your page is usually a passage you can rewrite. See AI Overviews and source citation.
Frequently asked questions
Is there a separate AI ranking system?
Google says no additional requirements or special optimisations exist; eligibility runs through ordinary indexing with a snippet.
What is ranked?
Passages, not pages. Chunks are embedded, scored against the query, and selected into a limited context budget.
Why is a smaller site cited over a bigger one?
Because the ordering is passage relevance to one question, not authority over a topic.
Can I opt out selectively?
Yes — snippet controls limit what may be shown from a page, and Google-Extended covers training and grounding in its other systems.
Terms related to AI content ranking
The architecture that retrieves documents at query time and has a model write from them, and the reason your page can be quoted without ever being trained on.
Google's AI-generated summary at the top of a results page, and the snippet controls that decide whether your page can appear inside one.
How an AI answer attributes what it says to the pages it read, and why a citation is a distinct outcome from a click or a mention.
Vectors of floating point numbers whose distance measures relatedness, the representation underneath semantic search and retrieval.