So Outside Technology Security CMS
AnQiCMS
Online Demonstration

Only a little over 30% of the pages referenced in the AI overview come from the top ten, so should we still pay attention to the ranking?

📅 2026-11-01 👁️ 0

The pages referenced in AI Overview are only over 30% from the top ten, should we still focus on ranking? We should, but the focus has changed: the decline in overlap indicates that 'being in the top ten' is no longer a condition for being referenced. Ranking monitoring has changed from a conclusion to an entry signal, and the real factor determining whether a page can be referenced is whether there are paragraphs that can be extracted separately from the page, and whether these paragraphs can be consistently attributed to a certain issue.

The overlap is decreasing, which layer is being referred to?

The proportion of more than 30% comes from the analysis of more than 860,000 result pages, which compares the intersection of two things: the pages referenced by AI overview, and the top ten pages of natural results for the same query. One year ago, this intersection accounted for about three quarters, and now it is only a little more than 30%.

It should be noted that the decrease in overlap refers to 'the references are no longer concentrated at the top', not 'the quality of the top pages has deteriorated'. The ranking of natural results is still determined by the existing signals, and the generated answers will explore earlier to the pages after the top ten in selecting sources, as long as the content of that section can directly answer the question.

In fact, the impact on the site is that two things have been split apart: whether it can be clicked, it depends on the ranking; whether it can be cited, it depends on whether the paragraph is an independent answer.

What are the remaining functions of ranking monitoring?

Ranking is not without value, but its scope of action has narrowed.

The first is entry inspection. The sudden drop in ranking usually appears before the reduction in citations, and once the position falls to the second page, the chance of the page being read decreases, and such anomalies still need to be discovered through ranking monitoring. The second is coverage verification. Whether a new issue can enter the top ten after going online reflects the match between the content and the issue, which is one of the prerequisites for citation. The third is competitive analysis. Who rises in the same batch of queries often also appears in the same batch of answers' citations.

The function no longer applicable is 'using ranking instead of visibility'. A weekly report that only reports rankings would miss a common situation: no change in ranking, decrease in citations; and also miss another situation: ranking decline, citations increase instead.

Three things more worthy of attention than ranking

First is the list of citation matching issues. Group queries of the same topic into several questions, and see which ones have entered the answer and which have not, which is easier to locate gaps than page by page to view traffic.

The second is the extractability of paragraphs. The generated answers extract a self-contained piece of text, which is easier to quote directly when it contains premises, numbers, or steps; content that requires cross-paragraph supplementation to understand is often rewritten into a single sentence summary, and the source may not necessarily fall here.

The third is identifiability. Content under the same theme should be able to be consistently categorized: the keyword library determines which thematic collection these pages belong to, and the anchor text determines what to call when linking internally. When the口径 are inconsistent, the same concept within the site may have multiple names, and the signals read will also be scattered.

How to make content extractable in the site

In practice, it can be condensed into four steps.

First, a paragraph should answer only one question, state the conclusion at the beginning, and then provide the basis without jumping to another paragraph. Second, conditions, values, and time ranges should be written into the sentence itself, not inferred from the context. Third, the anchor text of the internal link should use the original word in the question, not terms like 'here' or 'details' that provide no information. Fourth, align the keywords in the keyword library with the words in the paragraph, and maintain the same writing style for the same concept in the title, body, and internal links.

These four steps' benefits are not only in generative retrieval. After the paragraph is self-contained, even manual readers can judge whether to click on the result page by looking at only one abstract, which is both sides of the same thing as the ranking click-through rate.

Common questions

The overlap is only a little over 30%, should we still do natural search optimization?Yes. The page of reference dependency can be read, and the natural results are still the main entry point, and both are two sections on the chain, not a substitute relationship.

Should we focus on ranking first or on references first?Keep an eye on the anomalies in the rankings to identify problems; then focus on which issues are hit by references to locate the gaps. Only looking at rankings will miss out on 'positions unchanged but references reduced', and only looking at references will miss out on the entry itself becoming ineffective.'

How often should the keyword library and anchor text be verified?Verifying by topic is more practical: after adding a batch of content, check if its vocabulary falls into the existing topic set; merge the same concept immediately when two different writing styles are found, and do not leave it until the end of the quarter for集中处理.

Related articles

The garbage station update in September took nearly fourteen days to end, when is it suitable for review?

This round of garbage content update in September started in the late month and was marked completed in the early October, lasting for more than thirteen days. Before the update was completed, the ranking reading was the intermediate state within the rolling period, which cannot be taken as a conclusion; the time suitable for review should meet two conditions: the official panel is marked completed, and the data after the completion day returns to a stable interval. The waiting period should first check basic items such as robots and site maps.

2026-11-01

The window for the secure version and the major version release are scheduled in the same week, which one should the site follow first?

When the fixed secure version window and the annual major version are scheduled in the same week, follow the secure version first. The secure version closes known exploitable entries, while the major version changes data structures and template behaviors, with a higher verification cost. This article explains the differences in the nature of the two types of updates, how to evaluate the cost of delaying the major version by one week, and the two-step verification and rollback preparation to be reserved in the scheduling.

2026-11-01

Components that cannot be found in the vulnerability library, can they be considered risk-free?

A component that cannot be found in the public vulnerability library cannot be considered risk-free, as it is often just that the scope of collection and naming criteria do not match. The vulnerability library categorizes entries by core, extension, and theme, and verifies them one by one. Self-built and commercial components are often not included. This article provides three reasons for not finding them, the handling criteria for the worst-case scenario, and how to turn the asset list into an executable mitigation action sequence.

2026-11-01

Allowing customization of the position of the style, why does the style content also need to be filtered?

The reason for filtering the style content at the position where custom styles are allowed is that styles are not just a presentation layer: attribute names and values can be used to change the page display or make external requests. Filtering needs to cover the three segments of selectors, attribute names, and values, and be coordinated with sensitive word filtering, content review, and background role permissions. This article uses the bypass of the style attribute filter in the official repair list as an entry point to explain the criteria.

2026-11-01

Most llms.txt files have never been requested, and is it still worth maintaining them?

The analysis of logs from more than 100,000 sites shows that 97% of the llms.txt files have never been requested, with reading concentrated in a few sites. This result does not mean that the file is useless, but rather that it does not solve the two issues of being discovered and intercepted. Whether to continue maintaining it should be judged by looking at whether there are corresponding requests in the access logs of your own site, rather than by looking at whether the file has been uploaded.

2026-11-01

The adoption rate statistics of the two llms.txt files are very different, which decision口径 should be adopted?

Two studies that were checked on the same day are very different: The adoption rate is 8.3% when statistics are taken from the top ten thousand domain names, and it reaches 49.2% when statistics are taken from more than two thousand sites. The difference does not come from who made a mistake in the calculation, but from the denominator range and the method of judgment. When making decisions, do not take the average of the two numbers, but first see which category your site falls into among the samples, and then confirm whether the口径 can be used by checking the three-step verification method.

2026-11-01

The training crawler was blocked in robots, why can't the content still be retrieved

There are tests from more than 2,300 sites showing that only 3.6% of the sites block training crawlers in robots, but when accessing the homepage directly by real identity, 18.8% cannot access usable content. There is a gap between declarations and actual usability: writing rules does not mean they take effect, and there are also cases where the side that grants access cannot access the text. Self-inspection should be in order of four types of signals, not just whether that line of rules is in the text.

2026-11-01

Which type of spam submissions do the honeytrap field and captcha each block, can only one be left

The honey pot field blocks script submissions that do not parse styles, there is no visible challenge for real visitors; the scoring captcha is scored according to the possibility of robots, at the cost of mistakenly harming some real users. The two methods do not face each other, and whether only one can stand depends on the composition of garbage: when the amount of scripts is mainly used, the honey pot is enough, but when there are submissions with changed addresses and natural content, only the honey pot is not enough.

2026-11-01