GEO Citation Baseline Across Six Chinese Large Models (Sample Data)

This report records a sample baseline: 30 questions asked on 6 Chinese models, with mentions, citations, and primary answers recorded in fixed fields, and with the calculation method and limitations published.

Research questions

This report answers three questions in order to test whether the recording process is executable, not to deliver industry conclusions.

1. On the same question set, what distribution do mentions, citations, and primary answers for this site's content show across six Chinese models?
2. How large are the differences between models, and are they enough to support a judgment such as one model citing more easily?
3. Which type of question is harder to earn a citation for, and which part of the content format might explain it?

The report also serves as a method demonstration: it publishes the question set, record fields, judgment rules, and calculation method so readers can reproduce the flow. All figures are samples, not real monitoring results, and must not be used for external promotion or model capability comparison.

Sample and method

Question set

The question set has 30 items in four groups by content type. The purpose of grouping is not scoring but locating which type of question is easier to cite in the results.

Question wording stays fixed during the collection period and is not revised because an answer is disappointing. This is the precondition for comparability: once questions change, historical data cannot be compared directly.

A unified question template is used.

Please answer: <original question>. If you quote a web page, please provide the link.

The template contains no site name and does not prompt the model to mention a brand. Leading questions inflate the mention rate and destroy the observation value.

Models and entry points

All six models use their web interface, with login status and network environment kept consistent. The collection date is 2026-11-08.

Different entry points produce different results. Apps, mini programs, and web interfaces may differ in retrieval scope and display format, so the report compares data only within the same entry point and never extrapolates web results to other entry points.

Collection environment and operating constraints

Collection is done on the same device and the same network, with the browser in the same state outside incognito mode, to avoid login differences affecting retrieval scope. Before each model's collection, the session is cleared, and questions are then asked in number order.

Each collection session is kept within two hours to reduce interference from temporary product adjustments. If a login expires or a page misbehaves, record that item as not collected and do not fill in a guessed value. A gap is a gap; filling zeros distorts the ratio.

During collection, information visible in the product interface is also recorded, such as whether reference sources are shown, whether sources are collapsed, and whether sources sit at the start or the end of the answer. This information does not enter the fixed fields of the public data file but is written into notes, to help judge later whether the display format has changed.

Record fields

Each answer records nine fields, matching the public data file.

30 questions times 6 models equals 180 records. The file is UTF-8, and boolean values are written as true or false, never as yes or no, to avoid type errors on import.

Judgment flow

1. Ask each original question in order; ask the same question on the same model at most twice per day, and record the instance with the complete answer.
2. Screenshot the full question and the full answer, with screenshot numbers matching record rows.
3. Judge a mention: a brand or author name in the answer counts as true, with no link required.
4. Judge a citation: only a clickable address to this site in the answer counts as true; a name alone is false.
5. Judge a primary answer: the core conclusion matches a page on this site and that page is the main basis; both must hold for true.
6. Enter the records, including collection date and notes.
7. Review: check that the inclusion relationships among boolean fields hold, and check that addresses open.

One clarification: judgment in this report was done by a single person without a second review. This is a known limitation and follows from the nature of sample data. Real evaluation should use two independent judges and revisit screenshots where they disagree.

An example of one judgment

Suppose a model answers the question about the six-hop evidence chain by explaining the concept in two sentences, then giving a link to this site's method page, and then expanding on the six hop names in the next paragraph.

The judgment is: a mention is true because the site or author name appears; a citation is true because a clickable address to this site appears; a primary answer is true or false depending on whether the core conclusion matches the page. If the explanation matches the page conclusion and the page is used as the main basis, record true; if the page is only one of several parallel references, record false.

Conversely, if the answer restates the six hop names without a link, the mention is true, the citation is false, and the primary answer is false. This case accounts for a notable share of the sample data and is one direct source of the pattern of many mentions and few citations.

If the answer gives a link but the conclusion clearly differs from the page, for example explaining the six hops as the execution order of six steps, the citation remains true while the primary answer is false, and the wording problem is recorded in the notes and enters the content correction list. This example shows that the three fields are not redundant: they answer whether it was mentioned, whether it was treated as a source, and whether it was treated as the main basis.

Metric definitions and formulas

The three boolean fields map to three ratios, each with a denominator of 30 questions; aggregation uses 180 records.

Judgment rules have three boundaries that prevent double counting and logical contradiction.

1. If a citation holds, a mention necessarily holds. A clickable address means the brand or site appears in the answer.
2. If a primary answer holds, a citation necessarily holds. A core conclusion from this site means the page was used as a basis.
3. The reverse does not hold. A mention does not imply a citation, and a citation does not imply a primary answer.

ECI differs from the three ratios: it measures evidence chain completeness on the content side, not citation status on the answer side. Calculation samples 30 core claims, walks each through the six hops, and counts only those that pass all six. The example sets 18 as passing, giving an ECI of 60%. This figure only demonstrates the algorithm and does not represent this site's actual completeness.

Results

All results below are sample data, not real monitoring results. Figures are listed to show the table structure and how to read it, not as proof of effect.

【Claim】A public evaluation must first state the nature of its data, and sample data must not be used as real monitoring results. 【Basis】Section 5, Metric system and calculation methods, of this site's whitepaper GEO Methodology Whitepaper v1.0 (working definition). 【Publisher】GEOA (Xingge, one-person company). 【Date】Published 2026-11-09; review cycle 180 days. 【Location】Section 5, Metric system and calculation methods. 【Excerpt】The role of a metric is not to prove effectiveness but to make records from different times comparable. 【Link】https://geoa.com.cn/whitepaper/geo-method-v1/ | verified 2026-11-09

Distribution by model

Read the total first, then the differences. At the total level, the mention rate is near half, the citation rate is under one fifth, and the primary answer rate is under one tenth. This structure shows that even when content is mentioned, most answers still do not treat this site as a clickable source, and primary answers are even fewer.

Across models, mentions range from 14 to 16 and citations from 5 to 6, a gap of only one or two. With 30 questions as the denominator, one record equals 3.3 percentage points. A difference of that size cannot support a judgment that one model cites more willingly; it more likely reflects single-observation fluctuation.

Distribution by question type

Metric-type questions show a somewhat higher citation share, while service and judgment questions are clearly lower. The latter can be explained by content format: such questions usually expect advice and judgment, and answers tend to give general principles with fewer references to specific sources.

A reminder: these share differences rest on a very small sample, with only a dozen records per cell. Such differences are suitable for raising hypotheses, not for drawing conclusions.

Questions without citations

Of the 30 questions, 24 earned a citation on at least one model, and 6 earned no citation on any model.

This table is more useful than the overall numbers. Overall numbers only show the level; the gap list shows what to do next. After reworking according to the actions here, the next collection round can observe whether these 6 questions change.

Primary answer distribution

16 primary answers are concentrated on 9 questions, distributed as follows.

The distribution shows primary answers appear more on metric and method questions. A possible explanation is that these answers have a clear structure, with formulas, steps, or tables on the page, making conclusions easy to extract directly. This is only a hypothesis and needs more rounds of data.

Discussion

Observation one is the structural gap of many mentions and few citations. With a mention rate near half and a citation rate under one fifth, a name appearing is far easier than a source being adopted. The cause may lie in two places: the name mentioned has no matching clickable address, or the address exists but is not used as a basis. The fixes differ: the first requires checking whether the page is retrieved, the second whether the source is verifiable.

Observation two is that model differences are unstable. Citations across the six models cluster between 5 and 6, a gap of one or two. At this sample size, any ordering lacks statistical meaning. The report therefore lists distributions rather than model rankings. This is also why this site does not publish model leaderboards: conclusions date quickly and mislead readers into treating random differences as capability differences.

Observation three is that gaps concentrate on comparison and operational questions. Q02 lacks a comparison table, and Q15 and Q16 lack steps and pass criteria. Comparison tables and steps happen to be highly extractable structures, which suggests the gap is not whether something was written but whether it was written in an extractable way.

Observation four is the absence of the whitepaper. Q29 earned no citation at collection time for a simple reason: the relevant page was not live yet. This kind of gap needs no content revision, only the page, followed by re-collection. It reminds us that the question set should align with the content publishing schedule; otherwise not yet done is misread as done but useless.

From a method perspective, the mapping between results and the six hops can be read this way: mention maps to discoverable and openable, citation maps to extractable and verifiable, and primary answer maps to attributable and corroborable. The three metrics cover the progression from the first four hops to the last two. Mapping metrics to stages gives the gap analysis a place to land.

A fifth observation concerns the link between metrics and actions. The report lists three ratios, but ratios alone produce no action. What produces action is the gap list: which question type has no citation, whether the cause is missing structure or missing sources, and which page to change next. Making this step a fixed routine keeps reviews from stalling on whether the numbers look good.

Another phenomenon worth recording is that the same page performs differently across questions. The same method page may be cited for how do I get AI to cite my site, yet not appear for how do I rework an old article. The cause is usually different question phrasing: the former matches the page conclusion, while the latter needs another set of steps. Whether to add a page or add a paragraph depends on whether the difference repeats consistently.

Limitations

The sample is limited. 30 questions cannot cover all user intents, and 180 records are a single-day observation. Each question has only one valid record per model, so the fluctuation range of repeated asking cannot be estimated.

Model versions drift. Specific version numbers were not recorded at collection time, and answers to the same question may change after a product update. Results correspond only to the collection date and cannot be extrapolated to other times.

Parts are not reproducible. Answers may be affected by login status, region, time, and randomness; screenshots are only evidence and cannot fully reproduce the generation process. Even the same person rerunning the same flow may get different results.

The subjectivity of manual judgment. Primary answers require judging whether the core conclusion comes from this site, and different people may disagree. This report was judged by one person with no paired review, so consistency cannot be verified.

The nature of sample data. All figures in this report are samples, not real monitoring results, and must not be used for external promotion, commercial proposals, or model capability comparison. Public data exists to demonstrate fields and algorithms, not to prove effect.

It is not a model evaluation. The report makes no judgment about any model's capability, tendency, or quality. Model differences at this sample size have no explanatory power.

Judgment granularity also needs explanation. This report records by item and does not distinguish the number or position of links in an answer. An answer with two addresses to this site and an answer with one are recorded as the same boolean. For finer analysis, add fields for link count, link position, and cited paragraph, rather than inferring from existing fields.

Data download and field description

The public data file path is /en/research/geo-citation-baseline.csv, with the full address https://geoa.com.cn/en/research/geo-citation-baseline.csv.

The file is UTF-8 comma-separated text with one header row and 180 data rows. The header is fixed as: questionid, questiontext, model, mentioned, cited, primaryanswer, citedurls, recorded_at, note.

Fields match the body text, boolean values are written as true or false, cited_urls is an empty string when empty, and the note column is uniformly labeled sample data, not real monitoring results.

Reuse requires three things: keep the note column when reposting; do not remove the sample label and circulate it as real data; when doing secondary analysis on the file, state the limitations of the sample size and judgment method.

Citation format

Xingge. (2026). GEO Citation Baseline Across Six Chinese Large Models (Sample Data) (v1.0). GEOA. https://geoa.com.cn/research/geo-citation-baseline/

In-text citation may be written as (Xingge, 2026). When citing a specific table, note the table name and collection date so readers can confirm the data range.

Version rules: text revisions update on the same path and change the date; when the sample, question set, or conclusion changes, open a new path, keep the old version, and label it as superseded.

Related reading

• The Six-Hop Evidence Chain: Moving AI from Seeing to Trusting
• Answer-First: Writing a Citable Conclusion Within 60 Characters
• Citation Rate: Definition, Calculation, and Limitations
• Mention Rate: Definition, Calculation, and Limitations
• Primary Answer Rate: Judgment Method and Limitations
• ECI: How to Calculate Evidence Chain Integrity
• The Difference Between an AI Citation and an AI Mention