The End of Ranking

What this entry covers, and why it matters now

For roughly twenty five years, the working unit of online visibility was position: a page occupied a numbered slot in an ordered list, the slot was observable by anyone, and the slot could be tracked over time. "The end of ranking" is the claim that this unit no longer describes what is happening. AIOFacts treats the phrase as an overstatement of something real. Ranking has not ended. It has become plural, internal, and largely unobservable, and the single visible number that used to stand in for visibility now predicts less than it did.

The reason to write this down in 2026 rather than earlier is that the divergence is now measurable rather than merely argued. Ahrefs, analysing 863,000 keyword result pages and roughly 4 million AI Overview URLs, reported in March 2026 that 38 percent of pages cited in Google AI Overviews also ranked in the top 10 for the same query, against 76 percent in its earlier run of the same analysis. That is one company's dataset, not a law of nature, and it is discussed with its limits below. But a shift of that size in a metric that used to be the whole game is worth documenting carefully rather than sloganising about.

What ranking meant, precisely

Ranking, in the sense that is now under question, had four properties that made it useful. It was ordinal: results occupied positions one through ten and then onward. It was public: anyone could issue the query and see the order. It was stable enough to track: a position could be sampled daily and plotted. And it was singular: one query produced one list, so one number described a page's standing for that query.

Each of those properties is weakened on answer surfaces, and they are weakened for different reasons. The output of an AI answer is prose with attached links rather than an ordered list, so ordinality is at best implied. The output varies by user context, session, and model version, so publicness and stability both degrade. And, as the next section describes, one user query may become many machine queries, so singularity is the property that breaks first and hardest.

• Ordinal: a numbered position
• Public: observable by any third party issuing the same query
• Stable: samplable over time
• Singular: one query produced one list

What the platforms document about how sources are chosen

Google's Search Central documentation states that both AI Overviews and AI Mode may use a technique it calls "query fan-out", issuing multiple related searches across subtopics and data sources in order to develop a response, and that while responses are being generated its models identify further supporting web pages, so a wider and more diverse set of links can be shown than with a classic web search. That is the operator describing its own system, which is the strongest class of evidence available here, and it is enough on its own to establish that a single query no longer maps to a single retrieval.

Google's separate optimisation guide, published in 2026, is equally direct in the other direction: it states there are no additional requirements to appear in AI Overviews or AI Mode, no new machine readable files to create, and no special structured data to add, and that ordinary search best practice remains the relevant guidance. Both statements can be true at once. The selection process changed; the published advice about how to be selectable did not.

Other systems document different pipelines. Perplexity's help centre describes an approach that runs a web search for a query and then synthesises an answer from the retrieved material with inline citations. OpenAI documents distinct user agents for distinct purposes, with a separate crawler used for surfacing sites in ChatGPT's search features. Platform behaviour varies, and no publicly available documentation from any of these operators describes how a specific page is finally chosen for a specific answer.

What the measurements show

Three findings from independent datasets are worth recording, each with its provenance attached.

First, the weakening link between classic rank and AI Overview citation. Ahrefs' March 2026 analysis of 863,000 keyword result pages found 38 percent of cited pages ranking in the top 10, with the remainder split almost evenly between positions 11 to 100 and beyond position 100. Ahrefs attributes the shift to fan-out, which is consistent with Google's own documentation, though consistency is not proof of cause.

Second, the low overlap between what assistants cite and what search engines rank. In a separate Ahrefs study of 15,000 long-tail prompts, an average of roughly 12 percent of links cited by ChatGPT, Gemini, and Copilot appeared in Google's top 10 for the same prompt, with Perplexity a notable outlier at closer to one in three. Different systems retrieve differently, and a page's standing on one is a poor predictor of its standing on another.

Third, the collapse of position into traffic. Pew Research Center, tracking the browsing of over 900 consenting United States adults across 68,879 Google queries in March 2025, found users clicked a traditional result in 8 percent of searches that showed an AI summary against 15 percent of those that did not, and clicked a link inside the summary in about 1 percent of visits. Even being included, on that evidence, does not reliably return a visit.

Ranking did not end: it moved inside the pipeline

The more accurate description is that one visible ranking has been replaced by several invisible ones. A fan-out produces multiple sub-queries, each of which requires its own retrieval and its own ordering. Candidate documents are then reranked against the synthesis task rather than against the original query. Some subset is placed into a model's input context. Some subset of that is finally attributed in the visible answer. Each of those stages is an ordering operation, and none of them is published.

There is peer-reviewed reason to think that ordering inside the context matters and not only ordering during retrieval. Liu et al., in the Transactions of the Association for Computational Linguistics, documented that language model performance is often highest when relevant information sits at the beginning or the end of an input context and degrades when the model must use information in the middle. That is a finding about language models under controlled conditions, not a description of any production search system, and no operator has disclosed whether or how it applies to theirs. It is included here because it shows position did not stop mattering: it stopped being something an outsider can see.

The academic literature also treats visibility on these surfaces as a definable and optimisable quantity. Aggarwal et al., in the KDD 2024 paper introducing the term generative engine optimization, proposed a black-box framework and explicit visibility metrics for generative engine responses, which is a different measurement object from a rank position.

Why position is a weakening unit of measurement

The practical consequence is a measurement problem rather than a tactical one. If a user's question becomes several machine questions, then a rank check against the user's original phrasing samples only one of the retrievals that actually determined the outcome. If the answer names an entity without linking to it, the appearance leaves no trace in any analytics system at all. If the same prompt produces different sources for different users, a single observation is a sample rather than a measurement.

None of this makes rank worthless. On the Ahrefs data, top 10 pages still supply the largest single share of AI Overview citations, and pages that rank at all are plainly better represented than pages that do not. A defensible reading is that ranking remains a strong input and a poor summary: useful as a diagnostic of whether a page is retrievable, unreliable as a report of whether an entity is being represented.

What practitioners are doing, described rather than prescribed

Reported practice has moved from tracking positions for keywords toward tracking presence across prompts: asking a set of assistants a set of real buyer questions on a schedule and recording whether the entity is named, how it is described, and whether the description is accurate. This is a measurement change, not a technique that has been shown to cause inclusion.

On the production side, the publicly documented guidance has not shifted much. Google's optimisation guide continues to point at unique, expert-led, people-first content and at standard technical accessibility, explicitly stating that no new file formats or schema types are required. Where AIOFacts would add anything, it would add it as a position rather than a finding: where an entity is described inconsistently across the sources a retrieval step is likely to reach, the ambiguity has to be resolved somewhere, and resolving it in your own published material is the only part of that process you control.

• Track presence and accuracy across prompts, not position against keywords
• Treat rank as an input signal and a retrievability check, not as a visibility report
• Confirm that the crawlers used for AI search surfaces are not blocked, and record what was allowed and when
• Keep entity descriptions identical across the properties you own

The limits of this evidence

The two studies carrying most of the weight here come from a single commercial vendor whose products measure exactly what the studies find to be important. That is not a reason to discard them: the sample sizes are large and the methods are described. It is a reason not to treat them as settled. Neither has been independently replicated, and the earlier and later Ahrefs runs are separated by changes both in method and in the underlying systems, so the move from 76 percent to 38 percent conflates real change with measurement change to an unknown degree.

Nothing in this entry establishes causation. No source cited here observes a selection decision. Every account of what happens between a query and a citation, including Google's own, is a description of a process at a level of abstraction that leaves the decisive step undescribed. Claims that a particular property of a page causes inclusion in an AI answer are, on the current public record, inferences from correlation.

Finally, this is a snapshot. Google reported in January 2026 that AI Overviews moved to a newer model generation, and measurements taken before and after such a change are not directly comparable. Any number in this entry should be read with its date attached.