So Outside Technology Security CMS
AnQiCMS
Online Demonstration

What do the terms 'references' and 'share' refer to when measuring the performance of AI search?

📅 2026-11-01 👁️ 0

What do the reference rate and share mean when measuring the performance of AI search? The share answer is whether the brand appears in the answer, according to the standards checked on the day, the model share refers to the proportion of the brand included in the AI answer. The reference rate answers how many of the sources provided in the answer point to the brand itself. One measures presence, and the other measures source construction, and the two sets of standards cannot be substituted for each other.

What are the two indicators asking for

The denominator of the model share is the sample size of a batch of questions, and the numerator is the number of responses that mentioned the brand. It is suitable for answering whether the brand is considered in the question and answer scenario, and it is a comparative indicator: how often does it appear compared to its peers under the same set of questions.

The denominator of the citation rate is the number of sources provided in the answers, and the numerator is the number of sources that belong to the brand. It is suitable for answering whether the content layer is sufficiently supplied: even if the answer mentions the brand, if all citations come from third parties, there is no benefit to page引流 and credibility construction.

According to the口径 checked on the day, most of the references come from the brand's own or nearby sources, the 86% proportion indicates that the main battlefield of source construction is in the own domain, rather than external转载.

How to calculate the share

Three things must be fixed to calculate the share: question set, statistical period, and criteria.

The question set should cover the real questions that users will ask, not just pick good words; the statistical period is usually aligned with the weekly or revision node; the criteria should be clear on what counts as “appearance”——brand name, product name, own domain, any one of them is different in口径, mixed use will make the trend distorted.

The denominator is also crucial. The sample size is the number of questions asked rather than the number of page views, and the magnitude difference between the two is very large, it must be written clearly in the report. When the share changes, first check whether the question set has changed, and then look at the content-side changes.

How is the citation rate calculated?

The citation rate needs to be classified by the source domain. There are three available splitting methods: pages under the self-owned domain, sites close to the same subject, and the rest of the third-party sources. The second layer is easy to miss, and although the nearby sources do not belong to the main site, they are still within the controllable range.

Judgment of being cited should be based on the traceable sources provided in the answer, not on the text generated by the model. Only looking at the description would include false positives and spellings similar to the target, leading to a虚高index.

When the same question is asked multiple times, the citation rate should be averaged by round. A notice from the day before explained the reason: only 30% of brands can remain visible between two consecutive answers, and the single-round sampling results are not sufficient to support the conclusion.

Data collection and verification path

Indicator Measurement object Denominator Questions suitable for answering
Model share Whether the brand appears in the answer Number of sample questions Whether the brand is considered in the question and answer scenario
Citation Rate Number of brand-related items in the source Total number of source items provided in the answer Does the self-owned content become a cited source
Stability of visibility Whether the adjacent rounds continue to appear The number of consecutive questions Is the performance stable or randomly hit

The indicator needs to fall on the interface or backend data before it can be reviewed. AnQiCMS publicly releases high-frequency API call capabilities covering document lists and details, categories and tags, uploading images, comments and replies, friend links, system settings and contact information. The accompanying interface documents can be checked against the returned fields according to actions; on the MCP side, the convergence tool surface of the intent directory has round-level approval for writing operations, and data collection actions and writing actions are layered, making it convenient to hang the statistical script on the read side.

The measurement on the site side still depends on the content surface data. A total of 14 background modules related to GEO and SEO, such as llms.txt, full-text search, and comment anti-spam, can provide the original data of content visibility and user feedback. The calculation of indicators directly depends on these verifiable details.

The two sets of specifications cannot be substituted for each other

A high share and low citation rate indicates that the brand is mentioned but the content is not cited, the problem lies in the extractability of the self-owned page; a low share and high citation rate indicates that a few pages are cited repeatedly but the coverage is narrow, it is necessary to expand the content corresponding to the problem set. Both indicators together can locate the gap.

Common questions

Can we only look at one indicator?Not recommended. The share reflects presence, the citation rate reflects source construction, and the causes of the gap are different.

Is the neighboring source considered an owned citation?It should be classified separately as a neighboring layer; including it in the owned domain scope will overestimate the contribution of the main site's content.

How large should the sample size of the question be?Sample stratification by question category is more important than a total number; trends are not comparable when the sample scope changes.

Can the indicator interface with backend data?The read-side interface can return details such as documents, categories, and comments; the measurement script calculates according to the interface fields, and only retain verifiable records.

Related articles

Is it really easier for AI to cite statistical numbers in the main text?

According to the guidelines checked on the day, the visibility of content with statistical figures in AI answers is increased by 30% to 40%, but this is not an income that can be obtained simply by adding numbers. The numbers must have a scope and a verifiable source, and the paragraph should be able to stand independently; otherwise, isolated numbers will actually reduce credibility. The focus on content is determined by the keyword library on what questions to write, and content review ensures consistency in the scope of reference.

2026-11-01

AI answer the source of reference changes every round, what is left on the content side?

The change of reference source every round is the norm, not an exception. According to the criteria checked that day, only 30% of brands can remain visible between two consecutive answers, and 86% of the references come from brand-owned or nearby sources. There are still three handles on the content side: write paragraphs as independent extractable units, concentrate the pages worth reading in the owned domain and use llms.txt for navigation, maintain consistency and freshness with content review and scheduled publishing.

2026-11-01

After the garbage update, the fluctuation of rankings, first check which groups of site data

Rank fluctuation after garbage update, check the collection first and then the content. According to the October site owner report on the day, the garbage update in September is still fermenting in October, and at the same time, about a hundred publishers have obtained a small proportion of ad revenue in the artificial intelligence contribution pilot. The self-check order is: crawling and indexing status, invalid addresses in the site map, keyword library and anchor text distribution, and finally the ranking of specific word groups.

2026-11-01

Is the proportion of spiders blocked in training increasing, how should robots be grouped?

The proportion of blocked crawler training is rising. According to the samples checked on the day, the statistics of the robots.txt files of nearly five thousand top-level sites show that 14.3% are blocking GPTBot, and 19.2% are blocking at least one AI training crawler. It is recommended to divide the grouping writing by the purpose of crawling: search crawling, retrieval citation, and model training should be authorized separately, expressed by different user group rules, and llms.txt is only responsible for navigation without participating in authorization.

2026-11-01

When the form is flooded by robots, should we add human verification first or collect fields first?

When forms are flooded by bots, the field collection is done first followed by human verification. According to the samples checked that day, in the first eight days, the proportion of real people among the hundreds of contact form submissions was very low. The pricing of captcha service at a scale of thousands indicates that bypassing verification has already become industrialized, while edge-side condition filtering can block about three to five percentage points of the bottom flow. The order of processing is to tighten fields and condition filtering, then integrate reCAPTCHA, and finally leave a content review.

2026-11-01

When the system email cannot be sent out, should we first look at the bounce rate or the complaint rate

When the system email cannot be sent out, should we first look at the bounce rate, which directly reflects the quality of the address and whether the delivery is rejected; the complaint rate reflects whether the content has been judged as spam, usually after the delivery is successful. This article gives the troubleshooting sequence according to the delivery index threshold verified on the day and explains the division of labor in maintaining the email reminder and email subscription list triggered by the event

2026-11-01

How much of the operation and maintenance hours will be occupied by the patching frequency of self-hosted sites?

Based on the day's self-hosted operation standards, mainstream self-hosted content systems need to process patches from high to medium risk levels every six to eight weeks, accumulating eight to ten times a year. Each time, it involves announcement verification, assessment, backup, update and regression, plus on-site and backup drills, the hidden labor cost is often higher than the monthly fee of the host. This article provides the breakdown of working hours and the sections that can be saved.

2026-11-01

What order should be checked before upgrading the website when a planned-out-of-scope security version occurs in the running environment?

The release of a security version out of the normal release cycle for the operating environment indicates that the problem is not within the regular release period. Before upgrading, check in the following order: first, confirm that there is an available rollback point and complete the backup of static files; then, verify the current version with the affected range in the announcement; then, run the form, backend login, and key pages in the test environment; finally, peak load online for observation, and record the results in the version ledger.

2026-11-01