GEO Methodology Whitepaper v1.0: From Being Indexed to Being Cited

This whitepaper gives a GEO implementation method for Chinese large models: turning claims into verifiable sources with the six-hop evidence chain, measuring citation status with a fixed question set and metric standards, and labeling evidence levels and scope honestly.

Abstract and version information

This is the first release of the whitepaper, aimed at teams and individuals who need to make content citable by generative engines. The document sells no services and evaluates no vendor; it answers only three questions: what GEO is, how to do it, and how to verify it. All methods are this site's working definitions, marked L2; they have no third-party validation and do not constitute an industry standard.

Five conclusions

1. Being indexed only means the page entered the index; being cited means a generative answer used the page as a basis. The evaluation targets differ and cannot replace each other.
2. A citation is not luck; it can be divided into six checkable stages: discoverable, openable, extractable, verifiable, attributable, and corroborable. If one hop breaks, the claim is downgraded from evidence to statement.
3. The key to extractability is answer-first writing. The first screen needs a conclusion under 60 characters that still stands out of context, and paragraphs must be self-contained.
4. Metrics must be defined separately. Citation rate, mention rate, primary answer rate, and ECI answer different questions; mixing them yields conclusions that look like growth but are effectively invalid.
5. Evidence levels must be stated honestly. Without an external anchor, an original method can be marked at most L2, with the words empirical method. Honesty is itself part of credibility.

Intended readers

Scope

The whitepaper applies to Simplified Chinese sites answering questions on Chinese large models (DeepSeek, Kimi, Doubao, ERNIE, Tencent Yuanbao, Qwen), where the content is publicly accessible without login. It does not cover paid media, private-domain operations, on-platform optimization for cross-border marketplaces, or content that requires login to view.

Problem background: why GEO is needed

The entry point for users seeking answers is expanding from the search results page to the generative answer. In products such as DeepSeek, Kimi, Doubao, ERNIE, Tencent Yuanbao, and Qwen, users ask a question directly, the model produces an integrated answer, and in some cases it provides clickable source links. For content owners, this means a new visibility position: not a result in a list, but a sentence inside an answer.

The change in position is not a simple increment but a change in evaluation. Search results appear as lists, and positions can be collected in bulk with tools; generative answers appear as sentences, and whether a citation happened must be judged one by one along with source and context. Treating ranking data directly as proof of GEO effect produces the false judgment of ranking well but never being cited.

In practice, three misjudgments are common and worth separating one by one.

Misjudgment one: being indexed equals being cited. Indexing only means the crawler fetched the page; a citation also requires the page to enter retrieval candidates, the paragraph to be extractable, and the source to be verifiable. Many pages are structurally complete and indexed normally, but the first screen is a long sentence that depends on earlier text, so the model cannot extract a conclusion and simply bypasses them.

Misjudgment two: more domains equal stronger evidence. Publishing the same content repeatedly under multiple domains is easily read as same-source duplication and actually dilutes the signal. What raises credibility is one entity, one standard, and verifiable sources, not the number of pages.

Misjudgment three: self-corroboration equals credibility. Citing only across owned domains at most counts as corroboration among owned materials and cannot serve as external endorsement. Without a third-party verifiable mention, a claim can only stay at L2.

From a mechanical view, citations happen at the generation stage and pass through at least four actions: candidates enter retrieval, paragraphs are extracted, sources are verified, and identity is attributed. The four actions correspond to different content conditions, which is the origin of the six-hop evidence chain. Understanding this prevents GEO from being reduced to writing more articles or adding more structured data.

GEO therefore works on evidence units, not keywords. A citable evidence unit usually has three parts: a conclusion that stands alone, a source that can be opened and verified, and a stable entity identity. Reworking scattered opinions on a page into such units is the foundation of all later actions.

One prerequisite is easily overlooked: the crawling and retrieval strategies of Chinese large models are not fully public and change with versions. Any method can only raise the chance of meeting the conditions for citation; it cannot control the result. The whitepaper therefore keeps a limitations note in every chapter instead of giving certain promises.

Mechanism: how Chinese large models discover and use web pages

To apply a method, you first need to understand how the other side works. This section describes the common processing path of Chinese large models across four stages, crawling, indexing, retrieval, and generation, and marks what content owners can and cannot control. The four stages are a working model inferred from public documentation, developer notes, and observable phenomena, not a vendor flow chart, so they serve to guide actions rather than predict results.

Stage one: crawling

The crawling stage answers whether the page can be obtained. Models usually acquire pages through their own crawlers or partner data; whether a page is crawled depends on the robots protocol, site reachability, rendering method, and crawl budget.

What content owners can control includes: robots does not block body pages; the sitemap provides complete entries and is updated after publishing; body text is readable with scripts disabled; the server returns a normal status stably; and body text is not placed in containers that only expand on interaction.

What content owners cannot control includes: crawl frequency, crawl depth, and whether new pages are prioritized. These are decided by the other side; a site can only reduce the chance of being skipped, not require access on a given schedule.

A common misconception is treating submission for indexing as completed crawling. Submission only announces the address; crawling and indexing are still decided by the other side. Another misconception is using parameter pages to carry body text, since parameter pages are usually outside crawl priority.

Stage two: indexing

The indexing stage answers in what form a page enters the candidate library. Models de-duplicate, split, and structure crawled content, and a page may be split into several fragments rather than stored as a whole.

What content owners can control includes: one canonical address per piece of content; a clear heading hierarchy where second-level headings express a sub-question independently; key conclusions near the start of a paragraph; structured data consistent with body text; and no republishing of the same content across many domains.

What content owners cannot control includes: split granularity, de-duplication thresholds, and fragment retention policies. Different products may split the same page differently, which explains a phenomenon: of the same article, one model cites one paragraph, another model cites another paragraph, and a third model cites nothing at all.

Writing should therefore assume that every paragraph may be taken on its own. Paragraph self-containment is not a style preference but a hard requirement created by index splitting.

Stage three: retrieval

The retrieval stage answers which fragment will be selected. After a user asks a question, the model performs query understanding and rewriting, then retrieves relevant fragments from the candidate library, possibly using both keyword matching and semantic matching.

What content owners can control includes: pages covering the real wording of the question, not only industry jargon; paragraphs for the same question under different phrasings; fragments containing a clear subject and conclusion; and source information on the same page as the fragment for later verification.

What content owners cannot control includes: rewriting method, recall size, ranking weights, and whether web retrieval is enabled. The same question may get different candidate sets across products and times.

This leads to a practical conclusion: the question set should cover how users actually ask, not only internal terms. Writing only generative engine optimization without writing how to get my site into AI answers misses a large number of real questions.

Stage four: generation

The generation stage answers how the answer is organized and how sources are presented. The model organizes an answer based on selected fragments; some products give clickable sources, some list references only at the end, and some show no sources at all.

What content owners can control includes: conclusions and sources on the same page and matching each other; stable, openable source addresses; clear author and publication time on the page; and evidence levels and limitations so the content is not misread as a certain conclusion.

What content owners cannot control includes: whether sources are shown, how many are shown, their order, and whether conclusions are rewritten. These are product design decisions that sites cannot influence.

An important point: not showing sources does not mean the page was unused. Conversely, showing a link does not mean the conclusion came from that page. Judging a citation requires checking answer content against link targets item by item, which is why primary answers need manual judgment.

Three observable signals

The mechanism is opaque, but three signals can be observed to judge whether actions take effect.

Site access logs. Watch crawler visits from each model, the distribution of visited pages, and visit frequency. Logs show crawling, not citation.

On-site search and landing page data. Watch whether visits from question-answering products appear. This data is affected by privacy policies, is often incomplete, and can only be auxiliary.

Manual question records. Fix the question set, ask each question, take screenshots, and record links and conclusions. This is currently the most reliable way to judge and the source of the metric system.

The three signals cannot replace each other. Treating log growth as citation growth is one of the most common misjudgments.

Why the mechanism chapter must state uncertainty

Stating uncertainty prevents the method from being treated as a switch. What a method can change is the probability of meeting conditions, not the other side's product strategy. Acknowledging this allows reasonable fluctuation ranges in metric design and lets you find the real cause when results disappoint, instead of repeatedly rewriting titles.

Term and standard definitions

Unclear terms directly cause misplaced actions. This section gives the unified standards used by this site; other pages must not rewrite the conclusions when citing them. All definitions are working definitions, marked L2.

How the five concepts relate

GAO is the goal layer. It cares whether AI forms an accurate, stable, repeatable perception of a brand or person and mentions it proactively in recommendation questions. This is a long-term result that single-page optimization cannot deliver directly.

GEO is the main path to GAO. It cares whether a concrete answer contains a clickable source pointing to this site and whether that source is described correctly. A citation is a precondition for perception: without ever being cited, a model struggles to form a stable perception.

AEO is a technical form of GEO. It organizes content into shapes that answer a question directly, such as definition sentences, steps, comparison tables, and Q&A pairs. Good AEO makes a page easier to extract, but it does not solve source credibility.

LLMO sits lower. It describes how corpus and entity signals affect internal representations. This layer cannot be observed directly and can only be inferred indirectly from mentions, citations, and description deviations on the output side.

SEO is the foundation. Crawling, indexing, canonical links, and internal links are also needed by GEO. The difference is in evaluation: SEO looks at indexing and clicks, GEO at citations and description accuracy.

A special emphasis: GAO is Generative Awareness Optimization, not Google AI Overviews optimization. The latter is only a sub-scenario of GEO within a specific ecosystem and cannot represent the full meaning of GAO.

Five common misuses

• Equating GEO with writing FAQs. A FAQ is only one answer format; without sources and entities it is still unverifiable.
• Treating a mention as a citation. A mention has only a name and no clickable source, so it cannot serve as source evidence.
• Treating structured data as an indexing switch. JSON-LD lowers comprehension cost and does not change whether content deserves citation.
• Treating llms.txt as an industry standard. It is a community proposal, and its crawling effect requires measurement.
• Treating multiple languages and domains as coverage. Same-source duplication is detected, and coverage does not equal credibility.

Three rules govern term usage: on first appearance in body text write the term with its English name, then use the standard term consistently; keep one page per term and put aliases into alternateName in structured data; definitions must trace back to standards, textbooks, policies, or public documents, and when they cannot, label them as this site's working definition. For the basic definition of GEO, see the glossary entry What Is GEO (Generative Engine Optimization).

Method system: the six-hop evidence chain and four supporting principles

The method system has one main line and four supporting principles. The main line is the six-hop evidence chain; the supporting principles are the three-click rule, answer-first writing, entity consistency, and the machine-readable layer. The five are not parallel tricks but different links in the same chain.

The six-hop evidence chain

The six hops move AI from what it can see to what it dares to cite. Each hop has a checkable criterion, and results should be reproducible: different people checking the same claim reach the same conclusion.

The six hops are not a score but a process of elimination. If any hop fails, the claim is downgraded in the citation chain: without a source it becomes a statement, without attribution it becomes unsourced content, and without corroboration the perception cannot be established. Writing them as fixed actions has one benefit: results can be reviewed rather than judged by feel.

For the specific checks and counterexamples for each hop, see the method page The Six-Hop Evidence Chain: Moving AI from Seeing to Trusting.

The three-click rule

The three-click rule constrains the click distance to sources: from any landing page, a primary source must be reachable in at most three clicks. Beyond three, treat it as broken.

The basis for this rule is direct. When verifying sources, a generative answer can follow links only a limited depth; people reviewing also dislike jumping five or six times. If a claim requires readers to dig through a reference list for the matching passage, the claim effectively has no usable source.

There are two ways to apply it: give a link that goes straight to the primary source on the method page, or delete conclusions whose sources cannot be reached within three clicks. Both are more honest than keeping a citation that looks present but cannot be found.

Many pages fail at this hop rather than at writing quality. Complete reference links that do not map to specific claims are the typical looks like it has a source case. The fix is to split sources to the claim level: one claim maps to one evidence block, with location and excerpt stated inside. For details, see the method page The Three-Click Rule: Every Claim Must Reach a Primary Source Within Three Clicks.

Answer-first

Answer-first solves extractability. It has three requirements: the first paragraph states the conclusion; the conclusion fits within 60 characters; and the conclusion still stands out of context.

Of the three, the third is most easily overlooked. Sentences containing it, this, or the plan lose their subject when taken alone, and the model cannot tell what is being discussed. A writing self-check is to send the sentence alone to someone who has not read the page and see whether that person can state the subject, the judgment, and the conclusion.

At the paragraph level, one more rule applies: one paragraph covers one point, and no paragraph opens with a pronoun. This way, even if only one paragraph is taken, the meaning is complete.

Usable sentence patterns include: X is ..., and the criterion is ...; doing X takes N steps, and the criterion for each step is ...; the difference between X and Y lies in goal, target, and standard. Patterns are not style requirements but a way to reduce information loss during extraction.

For writing details and positive and negative examples, see the method page Answer-First: Writing a Citable Conclusion Within 60 Characters.

Entity consistency

Entity consistency solves attribution. One person or organization keeps a single authoritative identifier site-wide, and other pages only reference it without restating complete fields.

There are three practical rules: authors and publishers use fixed structured identifiers, and pages do not write another set of titles and organizations; name spellings are unified site-wide, with no modifiers such as teacher or mentor appended in the name field; and roles are divided clearly between sites, with methods and standards carried by one site and services and cases by another, so the same fact is not written twice.

Entity fragmentation is worse than it seems. When one subject has several names and several organization descriptions, the model cannot merge these pages and attributes citations to a vague institution, or even to the wrong subject. At that point, even good content fails the attribution hop.

The multi-site strategy is bound by the same principle. Building more sites does not raise credibility, and same-source duplication is detected; the correct approach is to consolidate content into one canonical page, with other sites providing only summaries and links. For details, see the method page Entity Consistency: Helping AI Recognize That These Pages Describe the Same Person.

The machine-readable layer

The machine-readable layer includes three types of files: llms.txt at the site root, JSON-LD inside pages, and the evidence index file. They serve different purposes and none replaces the content itself.

The positioning of llms.txt must be clear: it is a community proposal, not a W3C or IETF standard. Treating it as an indexing switch creates wrong expectations. This site uses it for one reason: to lower the model's cost of understanding the site structure.

JSON-LD's value is expressing information from body text in stable fields, such as article type, author identifier, publication time, modification time, and parent site. Fields must match body text; when they do not, they create conflicting signals.

The evidence index writes which claim, at what level, and where the source is into a machine-readable list for batch verification and updates. It still requires sources to open; otherwise it only repackages broken links. For the writing method and templates, see the method page The Machine-Readable Layer: How to Write llms.txt, JSON-LD, and evidence.json.

Six typical broken chains and repairs

【Claim】A core claim must pass six stages, discoverable, openable, extractable, verifiable, attributable, and corroborable; any broken hop downgrades it to a statement. 【Basis】This site's method page The Six-Hop Evidence Chain (working definition). 【Publisher】GEOA (Xingge, one-person company). 【Date】Published 2026-11-09; review cycle 365 days. 【Location】Section 3, The six-hop evidence chain. 【Excerpt】The six hops are not a score but a process of elimination. Any failing hop downgrades the claim in the citation chain. 【Link】https://geoa.com.cn/method/evidence-chain/ | verified 2026-11-09

Metric system and calculation methods

The role of a metric is not to prove effectiveness but to make records from different times comparable. That requires three fixed elements: a fixed question set, a fixed model list, and fixed record fields. If any of the three changes, data cannot be compared directly.

Boundaries between citation rate and mention rate

The denominator of citation rate is the number of questions, not the number of answers. When one question is asked on several models, calculate per model and aggregate as needed.

Mention rate includes cases without a link. When a brand name appears in an answer without a clickable address to this site, it counts only as a mention, not a citation. Mixing the two misreads being mentioned as being used as a source.

An example illustrates the calculation (sample data, not real monitoring results): 30 questions times 6 models gives 180 answers, of which 89 contain the brand name and 34 contain a clickable address to this site. Aggregated by answer count, the mention rate is 49.4% and the citation rate is 18.9%. These two numbers cannot be added together or substituted for each other. For definitions and boundaries, see Citation Rate: Definition, Calculation, and Limitations.

Judging a primary answer

A primary answer needs manual judgment, and three criteria are recommended:

1. Whether the core conclusion of the answer matches the conclusion on this site's page, rather than merely containing a link.
2. Whether this site's page is the main basis, rather than one of several parallel references.
3. Whether a screenshot was kept, containing the full question and the full answer.

All three must hold to count as a primary answer. A link with a conclusion from elsewhere does not count. A matching conclusion without a link also does not count, because the source cannot be verified.

How ECI is sampled

ECI measures evidence chain completeness and is calculated on a quarterly sample. Sampling 30 core claims each quarter and walking each through the six hops, recording pass or fail, is recommended.

Sample calculation (sample data): sampling 30 claims, of which 18 pass all six hops, gives an ECI of 60%. The 12 failures should be grouped by broken-chain type to form the repair list for the next quarter. For calculation and sampling details, see ECI: How to Calculate Evidence Chain Integrity.

Standard drift and exception handling

The greatest risk for metrics is not a low value but a standard that quietly changes while still being compared with history. Common drift sources fall into four types: product redesigns that change entry points and display formats; model upgrades that change answer style; question wording that gets casually optimized; and added or removed record fields.

There are three countermeasures: record the model version or entry point, the original question, and the collection date together; when the standard changes, rerun the baseline once and note in the report that the standard has changed and is not directly comparable with the previous period; and keep historical records additive only, expressing revisions as appended notes.

Outliers also need a process. When citations in a given week suddenly drop to zero, check the recording step first: whether collection was missed, whether the entry point changed, whether screenshots are missing. After confirming the records, check the page side: whether it was redesigned, whether crawling anomalies appeared, whether blocking rules were added by mistake.

Do not fill missing data with zeros. Write not collected when it was not collected and write the reason for interruption when collection was interrupted. Filling zeros makes a trend chart look complete while creating fake signals.

Sample baseline table

The table below is sample data, not real monitoring results, and is used only to show the record format and calculation method. Real monitoring requires asking each question, taking screenshots, and judging according to a fixed process.

Read the table in three steps: read the total to judge the overall level; then read differences between models to judge whether content is needed for a given entry point; finally read primary answers to judge whether the content is only mentioned rather than used as a basis. Only reading all three prevents mentions from being misread as effect.

Implementation path: staged actions and checkpoints

The implementation path has five stages. The stages are not a timeline but dependencies: the output of one stage is the input of the next. Skipping stages usually surfaces problems during evaluation.

Pass criteria for each stage

1. Baseline stage. Criterion: the question list, model list, and record fields are fixed, and the first round sets no target values. Setting targets invites bias during recording.
2. Structure stage. Criterion: body pages remain fully readable with scripts disabled, URLs are stable, and the sitemap has entries.
3. Content stage. Criterion: each target question maps to one page, the first screen has a conclusion, and key claims carry evidence blocks.
4. Machine-readable stage. Criterion: structured data fields match body text and sources in the evidence index open.
5. Evaluation stage. Criterion: records of the same question are comparable across weeks, and gaps map to specific pages and actions.

Boundaries with paid media, SEO, and public relations

GEO is not an isolated action and needs clear boundaries with existing functions to avoid duplicated investment or mutual offset.

With clear boundaries, metrics cannot impersonate each other. Traffic growth from paid media must not be written as a citation rate increase, and media coverage from public relations must not be written as more primary answers.

Twelve publishing checks

Of the twelve, items 3, 4, and 7 are the floor. When a source is missing, delete the conclusion rather than filling it with the industry generally believes. The same applies to levels: when unsure, mark L2, and do not inflate the level to look authoritative.

Content rework: turning old articles into citable pages

Most sites do not lack content; they lack a citable format. This section gives a rework order for existing articles, section pages, and product description pages. The goal is not more words but the conditions for extraction and verification.

Rework order

The recommended order is: add conclusions, add sources, add entities, add structure, then change titles.

This order is counterintuitive. Most teams instinctively change titles and add keywords first, then adjust body text. The problem is that titles determine clicks while conclusions determine citations. No matter how good the title is, a body made of long sentences that depend on earlier text still cannot be extracted.

Adding conclusions comes first because it decides whether the page is worth further investment. If a page has no conclusion that can stand alone, it is a pile of material with limited rework value, and should be merged or rewritten.

Adding sources comes next because conclusions need support. A conclusion without a source is only an opinion; mark it L2, describe it as an empirical method, and do not use it as an external claim.

Adding entities and structure solves attribution and readability. Once author, publisher, and name spellings are unified, the page can be attributed correctly; once heading hierarchy, tables, and lists are in place, fragments are easier to understand.

Changing titles comes last because titles should serve content that is already formed, not the reverse. Setting the title first tends to turn the body into a footnote to the title, full of empty phrases.

Six operations

1. Find the question the page really answers. Criterion: it can be written as one interrogative sentence and does not duplicate another page on the site.
2. Add one conclusion at the start of the body. Criterion: within 60 characters, with no context-dependent pronouns, and readable on its own.
3. Add an evidence block for each key claim. Criterion: all seven items present and the link reaches the source within three clicks.
4. Unify entity information. Criterion: author and publisher identifiers are consistent site-wide and name spellings have no variants.
5. Adjust structure. Criterion: second-level headings stand as sections, comparisons and parameters use tables, and steps use ordered lists.
6. Adjust the title and summary last. Criterion: the title matches the H1, and the summary is 20 to 60 characters and matches the body conclusion.

Look at the metrics after the six steps. If citations still do not change, go back and check the extractable and verifiable hops item by item instead of continuing to edit titles.

Before and after

A common counterexample

Counterexample (an illustration, not a real site): a section page was reworked by changing the title first, adding phrases such as complete guide and in-depth analysis, while the body stayed unchanged. After publishing, indexing was normal but citations did not change.

The cause is not the title but the body: the opening still discusses background, core conclusions are scattered across sections three and five, and sources are only three links at the end that do not map to claims. Even if the model retrieves the page, it struggles to extract one complete conclusion.

The fix is to return to the start of the order: move the conclusion to the first paragraph, split sources down to the claim level, and only then handle the title. This order looks slower but actually reduces rework.

Boundaries of rework

Rework can improve the conditions for citation but cannot promise results. The following cases may see no citation for a long time even after the six steps: the page topic is too broad and lacks a clear question; the site overall lacks verifiable external mentions; the content differs widely from how users actually ask.

In such cases, adjust the topic selection and question coverage first rather than piling more content onto the same page. The longer the page and the more scattered the conclusions, the harder it is to extract.

Organization and cadence: who does what and how often

Whether a method lasts depends on division of labor and cadence. Treating GEO as a one-off project usually stops after the first recording round; treating it as routine work at a fixed cadence is what accumulates comparable data.

The three roles can be the same person, but responsibilities must be recorded separately. The reason is practical: when writing and checking are done by the same person, it is easy to treat what I remember as a source. At least keep the checking action as a separate step with a verification date.

The value of cadence is comparability. Only fixed weekly collection can separate random fluctuation from trend change; only fixed quarterly review prevents the standard from deforming unnoticed.

Bulk rework in one pass is not recommended. Small batches mean each batch maps to one recording round, and problems are easier to locate. Rework twenty pages at once and you cannot explain why results improved or worsened.

Compliance and ethics

GEO compliance follows traditional content marketing, with two extra risks: fabricated evidence and fabricated effect. Both destroy credibility directly and are hard to repair.

AI-assisted writing needs disclosure. The disclosure is not a disclaimer but information about how the content was produced, so readers can judge credibility. Paragraphs involving standards, policies, and data must be checked against originals by a person and cannot rely on model paraphrase.

Public evaluation must follow three rules: publish the method and sample; publish all records rather than selected results; and clearly label the difference between sample data and real data. When the third is impossible, publish only the method, not the numbers.

Evidence standards and grading

The purpose of evidence grading is not to label content but to tell readers how far a claim can be verified. This site uses five levels.

Why this site's original work defaults to L2

This site's methods come from practice summaries and the organization of public material, and they are reproducible: readers can follow the steps and use the checklists. But these methods have not been validated by a third-party institution, nor reproduced by external research, so they do not meet the conditions for L3 or L5.

Marking original methods L2 and stating empirical method on the page is not self-deprecation; it lets readers estimate evidence strength correctly. Conversely, marking original methods L3 or L5 damages the credibility of the whole site: once readers find a level that does not match reality, they doubt the labels on other pages too.

Mutual citation between owned domains counts at most as L4. This is often overlooked: cross-links within a site network look like multi-party corroboration but are still the same subject.

The seven evidence block items

Each core claim carries one evidence block with seven fixed items; if one is missing, delete the claim.

1. Claim: one sentence that can be cited independently.
2. Basis: the full title of the document or method page.
3. Publisher: the agency, organization, or author.
4. Date: the publication date, with effective or expiry dates when necessary.
5. Location: clause number, section, page number, or data field.
6. Excerpt: not over 200 characters, with the cited range noted.
7. Link: the primary source address and verification date.

【Claim】Original methods and whitepapers are labeled L2 by default and are not raised without third-party validation. 【Basis】This site's specification Content Production and Governance (working definition). 【Publisher】GEOA (Xingge, one-person company). 【Date】Published 2026-11-09; review cycle 180 days. 【Location】Section 6, Why this site's original work defaults to L2. 【Excerpt】Marking original methods L2 and stating empirical method on the page is not self-deprecation; it lets readers estimate evidence strength correctly. 【Link】https://geoa.com.cn/about/editorial/ | verified 2026-11-09

Upgrades and downgrades

There are only two upgrade conditions: adoption by a third-party institution or a public publication with the source noted; or verifiable reproduction by external research. Meeting either one allows the related claim to move from L2 to L5. Note that this applies to the related claim, not the whole article.

Downgrade conditions are easier to trigger: a source link fails, a standard version updates, a data definition changes, or entity information changes. When any occurs, downgrade and label first, then schedule the repair. Until the repair is done, the claim is no longer used as an external conclusion.

Limitations and unsuitable scenarios

A whitepaper must state what it cannot solve; otherwise readers treat the method as a promise. The following six are known limitations.

Model mechanisms are opaque. The crawl scope, indexing method, retrieval strategy, and ranking logic of Chinese large models are not fully public and change with versions. Any method can only raise the probability of meeting conditions, not control results.

Results are unstable. The same question asked at different times and entry points may get different answers. A single observation cannot support a conclusion, so a fixed question set and repeated records are needed.

Effect cannot be promised. Any claim that citations can be promised does not hold. What can be promised is only the process: a fixed question set, a fixed recording method, and review at a fixed cadence.

There is a ceiling on evidence levels. Without external anchors, this site's claims can be marked at most L2. This is a factual limit, not an effort problem.

Manual judgment carries subjectivity. Primary answers and wording accuracy require human judgment, and different people may disagree. The mitigation is fixed criteria and retained screenshots, not the pursuit of complete objectivity.

Sample size is limited. 30 questions cannot cover all user intents, and 6 models cannot represent all entry points. Conclusions should serve as trend references and not be extrapolated into industry conclusions.

For readers, a reasonable expectation is: after applying the method, pages meet the conditions for citation more often, records are more reviewable, and problems are easier to locate. An unreasonable expectation is: after applying the method, a citation by a given model within a given window is certain.

Common misconceptions and corrections

After a method has run for a while, what appears most often is not a technical problem but a shift in judgment. The following seven misconceptions recur in practice, each with a corrective action.

Looking only at whether there is a link, not whether the conclusion is right

A link is the easiest signal to observe and is therefore often used as the sole criterion. But a link in an answer does not mean the conclusion came from that page: the model may list the page as a reference while the conclusion comes from elsewhere.

The correction is to record three things together: whether there is a link, whether the conclusion matches the page, and whether the page is the main basis. Only by recording them separately can you separate a citation from a primary answer.

Doing structured data only, without changing body text

Structured data takes effect quickly and is concrete, so it is easily treated as the main lever. But machine readability only lowers comprehension cost and does not change whether content deserves citation.

The correction is to check first-screen conclusions and paragraph self-containment first, then add structured data. The reverse order usually gives a page with complete fields but nothing extractable.

Treating a mention as a citation

A brand name in an answer creates the illusion of being recognized. A mention only shows the name was stated, not that the content was used.

The correction is to split mention and citation into two columns in the record table, calculate them separately, and report them separately. Reporting them together overstates progress.

Rejecting the method after one uncited result

Citations are affected by all four stages, crawling, indexing, retrieval, and generation, so single observations fluctuate widely. Judging whether a method works from one question round gives an unreliable conclusion.

The correction is to fix the question set, record for at least four consecutive weeks, and then judge the trend. Single-week data is for finding anomalies, not for conclusions.

Changing the question set frequently

Once the question set changes, historical data cannot be compared. Some teams quietly remove hard questions that were not cited so that data looks better, which destroys comparability.

The correction is one question set version per quarter, additive only in between; when an adjustment is truly needed, keep the old version and record both tracks for one cycle.

Treating media mentions as citations

A press release or media report mentioning the brand is an external anchor and can strengthen corroboration, but it is not a citation in a generative answer.

The correction is to separate external mention and AI citation into two categories in the evidence ledger; the former raises evidence levels, the latter feeds the citation rate.

Ignoring reachability and loading

A page previews normally locally but cannot open online due to script rendering, blocking rules, or certificate problems. Such problems never appear in content review but invalidate all optimization.

The correction is to put reachability checks into the publishing flow: body text readable with scripts disabled, normal status codes, and mobile accessibility. One check takes a few minutes but prevents a whole batch of content from being wasted.

Conclusion

The goal of the whitepaper is not to give a template to copy but to offer a judging method that can be checked and reviewed. Methods can change, but judgment criteria should stay stable: whether a conclusion stands alone, whether a source is verifiable, whether attribution is clear, and whether metrics are comparable.

If only one piece of advice can be kept, it is to get the first-screen conclusion and the source right first, and then discuss structured data and machine-readable files. The former decides whether content deserves citation; the latter only lowers comprehension cost. In the reverse order, investment goes where it is visible but ineffective.

Citation format and license

Citation format

The author-year format is recommended, as it is easy to mark in body text and easy for machines to parse.

Xingge. (2026). GEO Methodology Whitepaper v1.0: From Being Indexed to Being Cited (v1.0). GEOA. https://geoa.com.cn/whitepaper/geo-method-v1/

In-text citation may be written as (Xingge, 2026), with the full entry in the reference list. When citing a specific chapter, note the chapter name, for example Section 4, Method system.

Excerpt requirements: a single excerpt must not exceed 200 characters; keep the original meaning without taking text out of context; attach the original link and version number. When data is involved, note sample data or the actual collection time.

Version rules

Version numbers and release dates must match in three places: the page, the downloadable file, and the changelog. Changing only one creates citation confusion: the version a reader has does not match what the page shows, so the reader cannot tell which version a conclusion belongs to.

License

This whitepaper uses the CC BY 4.0 license. Copying, reposting, adaptation, and commercial use are allowed under the following conditions:

1. Attribute the author and publisher and keep the original link.
2. Note the version number and citation date.
3. Adapted versions must describe the changes and must not imply that the original author endorses the adaptation.
4. Do not remove the evidence level labels or limitation notes from the original.

Disallowed uses include: removing the limitation notes and circulating the text as a certain conclusion; packaging an L2 method as an industry standard; and publishing unconfirmed content in this site's name. Errors or improper wording can be reported to the contact email published on the site; after verification they are revised under the correction process and logged in the changelog.

Appendix A: self-check list

The list is used by ticking each item, not by reading through and judging from impression. If any item is no, fix it before publishing; when it cannot be fixed, delete the related claim rather than keeping a half-finished conclusion.

Two ways to use the list

The list works as a publishing gate and as a review tool. The focus differs between the two.

As a publishing gate, judge each item yes or no, and do not publish if any is no. This suits the final check before a single page goes live, and the point is to stop obvious defects from shipping.

As a review tool, the point is not to make everything yes but to record which items chronically fail. When the same item fails for three consecutive months, the problem is not execution but process design, and the division of labor or templates need adjusting.

Both uses require records. Scoring without recording means the same problems repeat next time; recording without tracking means the issue list grows until nobody reads it.

Appendix B: FAQ

Will GEO replace SEO? No. They answer different questions: being found and being cited. Search remains an important discovery entry point, and generative answers are an added citation entry point, so the evaluation standards must stay separate.

Will adding JSON-LD get content cited? No. Structured data only lowers machine comprehension cost; whether content is cited depends on extractability and verifiable sources.

Is llms.txt an official standard? No. It is a community proposal, its crawling effect requires measurement, and it cannot serve as an indexing switch.

Will building more sites increase citations? Usually not. Same-source duplication is detected, so the correct approach is to consolidate into one canonical page and let other sites provide summaries and links.

How long until citations appear? There is no fixed cycle. It depends on crawling, indexing, retrieval, and content conditions, and any claim of a definite timeline is not credible.

Can the baseline data in the whitepaper be cited directly? Not as real monitoring results. The tables and examples are sample data used to show the record format and calculation method; real data should come from published evaluation reports.

Appendix C: quick glossary

Appendix D: twenty terms in brief

Appendix E: an example of a complete rework

This section uses a fictional example to show how the six-step rework reaches the sentence level. The example points to no real site and does not represent real results.

Before the rework, the page opened with: with the rapid development of generative AI, more and more companies are paying attention to GEO, but many confusions remain in practice, and this article will discuss it from several angles.

This passage has three problems: it opens with background and gives no conclusion; the subject is vague, since many confusions points nowhere; and the paragraph does not stand once out of context.

After the rework, the opening became: the GEO execution order is to add conclusions, add sources, add entities, add structure, and change the title last; a reversed order leaves optimization at the surface. This sentence is about 50 characters, has a clear subject, and still stands when read alone.

Before the rework, sources were three reference links at the end. After the rework, each key claim has an evidence block below it stating the basis name, publisher, date, location, excerpt, and link, so readers can reach the source within three clicks.

Before the rework, the structure was subheadings throughout with no tables. After the rework, the rework order, pass criteria, and common counterexamples are made into tables, and steps are ordered lists, so readers can follow the tables.

Before the rework, entity information appeared in three forms. After the rework, the whole site uses one name, author and publisher use fixed identifiers, and pages no longer redefine titles and organizations.

The self-check results after the rework are shown below.

It must be repeated: completing the rework only means the page meets the conditions for citation, not that it will be cited. To judge whether it worked, return to the periodic records of the fixed question set.

Appendix F: version record

The purpose of the version record is to make citations traceable. When readers cite a conclusion, they can confirm which version it belongs to and whether it has been revised. Revisions do not overwrite history: text adjustments are recorded in the modification time on the same path, while structural and conclusion changes open a new path, keep the old version, and label it as superseded.

The next version has three priorities: add real evaluation data with collection times; adjust metric standards and examples based on review results; and record actual progress on external anchors. Whether to move to v2.0 is evaluated after all three are done.

Related reading

• What Is GEO (Generative Engine Optimization)
• The Six-Hop Evidence Chain: Moving AI from Seeing to Trusting
• The Three-Click Rule: Every Claim Must Reach a Primary Source Within Three Clicks
• Answer-First: Writing a Citable Conclusion Within 60 Characters
• Entity Consistency: Helping AI Recognize That These Pages Describe the Same Person
• The Machine-Readable Layer: How to Write llms.txt, JSON-LD, and evidence.json
• ECI: How to Calculate Evidence Chain Integrity