So Outside Technology Security CMS
AnQiCMS
Online Demonstration

AI abstracts let the click-through rate of the top results drop by more than half, what indicator should the content site switch to?

📅 October 26, 2026 👁️ 1

AI summaries have reduced the click-through rate of top results by more than half, the first metric that content sites should switch to is 'cited' instead of 'clicked'. Ahrefs' research on 300,000 keywords states that the average click-through rate with AI summaries is 58% lower, in the December 2025 sample, the click-through rate of top results is only 0.016, which is one order of magnitude lower than 0.073 in December 2023.Clicks are still happening, but a portion of them landed on the summary instead of your page, so just looking at PV and click-through rate will mix 'content is not well done' and 'presentation has changed' into one thing.

What are these numbers saying

The mechanism of the decline in click-through rate is very specific: users have read the answer on the results page and do not need to enter the page. The click-through rate of the top results has dropped to 0.016, which means that even if the position is in the front, the proportion converted to the site is also very low. This is not a content quality score, but the value of the display position has changed.

Pay attention to the scope boundary: This is a comparison of the same batch of keywords between the AI abstract and without AI abstract scenarios, not the year-on-year conclusion of the site's own traffic. Taking 58% as an inevitable decline of your own site will understate the category differences - transactional, navigational queries are less affected by the abstract, while definitional, comparative, and step-by-step queries are greatly affected.

Why click-through rate is no longer a good judge

Three reasons.

First, the numerator was cut off. The abstract takes on the answer, and the denominator (exposure) of the click-through rate does not decrease synchronously, the indicator naturally declines.

Second, the attribution chain becomes longer. Users may see your paragraph in the summary, then click into similar sites, or may first enter the cited source given in the summary. The site logs cannot distinguish between these two paths.

Third, the judgment of long-tail queries is amplified. Zero clicks are more common in the long tail, and the long tail is where content sites accumulate thematic weight. Using click-through rate to evaluate long-tail articles can easily lead to the wrong conclusion of cutting off the long tail.

The click-through rate should still be considered, but it should be downgraded to a secondary indicator and not be the primary basis for选题选题 and content production.

Which three indicators should be looked at next?

Indicator level What to look at Collection method Why is it more reliable than click-through rate?
Cited layer The number of times the brand name and specific paragraphs appear in AI summaries/answers. Sampled answer screenshots and source lists by topic. Directly measure whether the content is extracted.
Reach the quality level. Stay on answer pages, second visits, and form/contact actions. On-site behavior reports layered by page type. Clicks are fewer, but the intention is clearer.
Indexable layer. Timeliness of new content inclusion, sitemap, and push receipts. Link push and active push records, sitemap submission status. Reflects whether the machine can read the content in time

The citation layer is the closest to the essence of the problem, but the disadvantage is that the cost of statistics is high, so it should be made into a sample rather than a full amount. The quality layer uses a small number of high-intent page conversions to replace the total. The indexable layer is a prerequisite: if the content is not read, the first two layers have no basis.

Two things that can be done on the site side

First, write the answer in a paragraph that can be extracted. The beginning of each article gives the conclusion first, the paragraph is self-contained and can be retold without the context, the subtitle is written in the form of a question that the user will ask. The abstract captures paragraphs, not the whole page.

The second is to clearly explain the machine-readable entry. AnQiCMS supports the automatic generation of the site LLMs.txt, which is convenient for large language models to understand and index the website content;This is different in terms of responsibility - the former informs the model of the content structure, while the latter provides link lists and statements about crawling boundaries.Accelerate inclusion and go through link push management, push new content to search engines, and support Baidu/Bing active push. On the side of in-site keywords, there is also keyword library management and anchor text management to ensure that internal links and names for the same topic are unified, so that the extracted paragraphs can fall on the topic aggregation page.

These two things do not change the display form, what changes is whether 'you are selected' and whether 'you can be found after being selected'.

Common questions

Should we rewrite old articles specifically for AI summary?Prioritize rewriting answer-type, comparative-type, and step-type pages, which are most easily summarized and rewriting benefits are evident; news-style and time-sensitive articles can be kept as they are.

Can the citation of the brand be counted?It is possible to conduct a sample survey: fixed a group of core queries, record the sources and paragraphs appearing in the answers on a weekly basis, forming a trend rather than an absolute value.

Is a decrease in PV a sign that the content strategy has failed?It is not necessarily so. First, make a layered judgment: whether the exposure is stable, whether the collection is timely, and whether the citation is rising. If all three are normal but the clicks drop, it belongs to the change of display form; if the exposure or collection declines at the same time, it is a problem on the site side.

Related articles

The cleanup of the garbage update will take two weeks, should the content team stop releasing during the update window?

The content team in the update window does not need to stop posting regular content. Taking Google's September garbage update as an example, it began on September 24, 2026, and ended on October 8, with a two-week rolling cycle on the system side, not a site exception. The delay is for template revision and structural changes, regular articles are still published but classified according to risk levels, and reviews are kept before going live with scheduled publishing and draft, pending publishing status.

2026-10-26

How much does it cost to self-host a Go language system for five years, and where do the bills differ from PHP stacks?

The main difference in the five-year self-hosting gap is not in the software license, but in the host slot usage decided at runtime, and who bears the maintenance hours. Third-party accounting provides a set of references: under the same VPS slot, the hardware bill for five years is about several hundred dollars, while the complete five-year accounting for maintenance sites can reach $2,000 to $4,000. The memory and concurrency surplus affects the lowest running slot, which is where the stacks truly differ.

2026-10-26

The statistics that 13,000 websites are hacked every day are saying who, and which layer the attack has entered.

The statistics show that about 13,000 sites are hacked every day, which refers to the existing risk of a large number of publicly available sites under batch scanning, not an occasional accident of a single site. What is more noteworthy is the level of attack entry: malicious code is injected into legitimate core, plugin, and theme files, bypassing the approach of finding problems by "abnormal files". The defense line should be shifted to change verification and recovery capabilities, rather than just focusing on login entries.

2026-10-26

The report shows that 90% of vulnerabilities come from plugins, which part of the site should be strengthened first.

The annual security report shows that 90% of the newly discovered vulnerabilities come from plugins and themes. This proportion points to the attack surface location, not the core quality. The sequence of site reinforcement should be based on the exposed surface rather than the number of components: first, converge the entry points and high-risk operations that can be triggered externally, then count the list of extensions, and finally prepare credentials and backups as recovery capabilities. Installing fewer plugins reduces follow-up costs, but does not equal security.

2026-10-26

llms.txt is called the sitemap of the AI era, how does it divide labor with the robots rules

llms.txt and the robots rule are not the same level of things: llms.txt is the content structure and summary entry for large language models, placed in the root directory or any path according to the proposal in Markdown; the robots rule declares the crawling boundary and the sitemap submits link lists. The three complement each other rather than replace each other, and interpreting llms.txt as a shutdown switch is a misreading.

2026-10-26

Can the content built by the page builder be exported completely, and which one should be checked before switching to the new system?

The content built by the page builder usually cannot be exported completely: blog article data has an export channel, the layout and components generated by the builder often need to be rebuilt in a new system, and most of the migration from platforms like Weebly relies on manual work. Before switching systems, one should first check whether the data layer can be exported structurally; then, consider the cost of reconstructing the structural layer and whether the link layer can use 301 redirects and replace the entire site to handle old addresses.

2026-10-26

How to arrange the upgrade plan for websites still using Joomla after the five branches stop issuing regular repair announcements.

In the Joomla official release records, the routine repair versions of the five branches 5.4.8, 5.4.9, 5.4.10, 6.1.3, and 6.1.4 have been released in one go, and the 6.2 announcement was also made on the same day, specifying that 5.4.10 is the last bugfix version in the Joomla 5.x series. Websites still using 5.x should arrange the upgrade plan in three steps: count the branches and extension dependencies, test the upgrade in the test environment, and determine the window and rollback points.

2026-10-26

How should operations use these few days to buffer the advance notice of security updates?

When a security announcement is released ahead of schedule, this buffer time should not be used to wait for the announcement. Taking the Drupal core security release preview as an example, the announcement was issued on October 12th, and the fix was released from 16:00 to 21:00 UTC on October 14th, with the highest risk rating being critical. The correct use of the buffer period is to first make recoverable backups and rollback points, clarify the list of affected branches, and then compare update records and arrange for upgrades at the designated time.

2026-10-26