Generative engine optimization cheating judgment system and method based on multi-dimensional feature analysis

Abstract

The invention discloses a multi-dimensional feature analysis-based generative engine optimization cheating judgment system and method, and relates to the technical field of computer network security and artificial intelligence, and the system comprises a data preprocessing module which is used for receiving a user query and an original answer text generated by a large language model, performing named entity recognition and linguistic feature extraction on the original answer text to obtain a standardized analysis object; the detection and judgment module is used for executing heuristic language feature analysis on the standardized analysis object to identify pragmatic anomalies, and comparing an AI recommendation sequence with real world market data based on a statistical imbalance principle to output a quantifiable risk judgment result; according to the GEO cheating judgment system and method, heuristic language feature analysis, statistical imbalance difference calculation, antagonistic traceability and automatic RAG evaluation are integrated and fused, and an active confrontation probe reproduction mechanism is further introduced to form an auditable evidence chain.

Description

Technical Field

[0001] This invention relates to the fields of computer network security and artificial intelligence technology, specifically to a generative engine optimization cheating detection system and method based on multidimensional feature analysis. Background Technology

[0002] With the popularization of generative artificial intelligence, a new threat has emerged: Generative Engine Optimization (GEO). Through high-dimensional soft advertising and data poisoning, it uses the black-box characteristics of large language models to seamlessly integrate marketing corpora with objective facts, making it difficult for users to detect anomalies simply by reading.

[0003] Against this backdrop, those skilled in the art face at least the following technical challenges when building GEO anti-fraud capabilities: how to automatically identify hidden pragmatic anomalies and structured inducement behaviors in generated text, and form interpretable evidence of anomalies, without relying on internal model parameters and based solely on generated text and available external data interfaces; how to establish a set of quantifiable statistical imbalance indicators to compare the degree of disconnect between AI recommendation rankings and real-world market data, and output risk assessment results with configurable thresholds accordingly; how to verify the fidelity and citation density of generated content, and combine it with transferable adversarial watermarking to achieve content traceability, thereby suppressing anonymous poisoning; and how to visualize the above assessment results in an intuitive way, enabling users to receive actionable early warning feedback and restore the integrity of the digital information ecosystem. Summary of the Invention

[0004] The purpose of this invention is to provide a generative engine optimization cheat detection system and method based on multidimensional feature analysis to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a generative engine optimization cheating detection system and method based on multidimensional feature analysis, comprising a data preprocessing module for receiving user queries and original answer text generated by a large language model, and performing named entity recognition and linguistic feature extraction on the original answer text to obtain standardized analysis objects; a detection and judgment module for performing heuristic linguistic feature analysis on the standardized analysis objects to identify pragmatic anomalies, and comparing AI recommendation ranking with real-world market data based on the principle of statistical imbalance to output quantifiable risk judgment results; and a source tracing and alarm module for performing automated RAG evaluation on the generated content to obtain fidelity and citation density indicators, and achieving content source tracing based on adversarial watermarking, thereby visually outputting alarm indicators and difference charts of risk levels.

[0006] According to the above technical solution, the data preprocessing module includes: an input interface unit for receiving user queries and the original answer text generated by the large language model; an NLP preprocessing unit for performing named entity recognition on the original answer text and extracting linguistic features; and an entity standardization unit for performing aliasing and normalization mapping on the named entity recognition results to form an entity set that can be used for sorting and comparison. The detection and judgment module includes: a semantic emphasis deviation detection unit, used to calculate the density ratio of evaluative adjectives to descriptive nouns in the target entity description and mark anomalies accordingly; a parasitic citation detection unit, used to parse the domain name and path of the cited link and identify citation deception behavior pointing to user-generated content sub-paths under authoritative domains, and perform a weight reduction on the authority weight of the citation; a fact silo verification unit, used to extract key data assertions and perform cross-validation in a third-party authoritative database to mark data illusions and fabricated data; a structured inducement identification unit, used to identify sandwich ranking in the list structure and detect abnormal conversion structures where only the target brand is accompanied by discount codes and purchase links, and to detect the closed-loop fitting of the text logic flow to the PAS model; and an AI volume calculation unit, used for... The system calculates AI volume scores based on the frequency of mentions of entities in the response text, the top recommendation, and sentiment polarity, and forms a contextual ranking. A benchmark data acquisition unit connects to external data interfaces such as Google Trends, SEMrush, SimilarWeb, Reddit mention rate scanning, and Twitter mention rate scanning to generate real-world market rankings. An imbalance analysis engine calculates the ratio of the market ranking to the contextual ranking and outputs a high-risk alert when a preset threshold condition is met. An active adversarial probe unit generates multiple sets of semantically equivalent and constraint questions based on the user query and repeatedly triggers the heuristic language feature analysis and imbalance analysis to form an anomalous evidence chain that can reproduce the experiment. The source tracing and alarm module includes: an automated RAG evaluation unit, used to construct a golden dataset and calculate fidelity and citation density metrics; an adversarial source tracing unit, used to embed highly transferable implicit watermarks during the content generation or original content publication stage and decode the watermarks after cross-platform copying and rewriting of the content to trace the original content source; and a visualization output terminal, used to display a comparison bar chart of AI volume and real popularity, and to indicate data consistency and data discrepancy states with green shields and red warning lights, respectively, and to provide hover prompts for ranking discrepancies.

[0007] A generative engine optimization cheat detection method based on multidimensional feature analysis includes the following steps: S1. Data Access and Preprocessing: Receive user queries and raw response text generated by the large language model, perform named entity recognition on the raw response text to form an entity set, perform normalization mapping on the entity set and extract the linguistic features required for subsequent analysis. S2. Linguistic Anomaly Detection and Probe Reproduction: Perform heuristic linguistic feature analysis on the original answer text to identify semantic emphasis bias, parasitic references and structured inducements. Construct multiple sets of equivalent and constraint questions based on user queries and repeatedly perform the heuristic analysis. Aggregate the results of multiple rounds of analysis into an anomaly evidence chain to improve the reproducibility and interpretability of the judgment. S3, Statistical Disorder Calculation: Calculate AI volume for entity sets and form contextual rankings. Obtain real-world market data through external data interfaces and generate market rankings. Compare contextual rankings with market rankings to output a disorder suspicion index and output risk levels based on thresholds. S4. Source Tracing and Fidelity Verification: Construct a gold dataset as the benchmark truth and perform fidelity and citation density evaluation on the generated content. Perform adversarial watermark decoding on the generated content to trace the source of the content and identify anonymous poisoning links. Link and store the evaluation results with the imbalance calculation results for auditing. S5. Visualization Output and Alerts: Generates visual charts to compare AI volume with real popularity, outputs green safety icons or red warning icons according to risk level, and outputs ranking gap and risk explanation information to prompt users to take appropriate action.

[0008] According to the above technical solution, S1 specifically includes: S1-1 Receiving Steps: The input interface unit receives the user's query text and the original answer text generated by the large language model, and binds the two into the same session record; wherein, the session record contains at least the query text, the answer text, the receiving timestamp, and the answer source identifier, so as to ensure that all subsequent detection results can be traced back to the same original input and output; S1-2, Entity Extraction Step: Perform named entity recognition on the original answer text to obtain an entity set. ,in Each element in the text is a standardized name of a business entity; the named entity recognition takes the answer text as input, outputs entity candidates and their position range in the text, and merges different spellings of the same entity into the same standardized name through entity standardization rules, thereby avoiding ranking and statistical bias caused by different spellings of the same name; and limits the scope of objects for all subsequent statistical, sorting, citation and tracing processing. S1-3, targeting For each business entity in the text, extract the fragments of its appearance in the original answer text and generate feature records for subsequent semantic emphasis bias detection, parasitic reference detection, and structured induced recognition. The feature records include at least: the number of times the entity appears, the location of the entity's first appearance, fragments of modifiers around the entity, the set of reference links corresponding to the entity, and the paragraph and list structure information of the entity. The feature records are used to convert the natural language description in the text into structured input that can be stably processed by the rule engine, thereby ensuring the consistency and verifiability of the detection conclusions.

[0009] According to the above technical solution, S2 includes: S2-1, Count the number of evaluative adjectives in the descriptive fragment for each entity. and the number of descriptive nouns Calculate the density ratio The evaluative adjectives refer to words that subjectively praise, criticize, strongly recommend, or exaggerate an entity, while the descriptive nouns refer to nouns used to objectively describe product categories, functions, technical features, and service models. The larger the value, the more marketing-oriented the description becomes. Compared to competing products Exceeding the preset deviation threshold When this happens, the target entity is marked as having a semantic focus deviation anomaly; S2-2. Parse each cited link in the original answer text and extract the domain name field and path field. When the domain name field belongs to the preset authoritative domain name set and the path field belongs to the user-generated content sub-path set, mark the citation as a parasitic citation and reduce its authority weight for subsequent fact silo verification. The authoritative domain name set is used to identify site domain names that are generally recognized as highly credible sources, and the user-generated content sub-path set is used to identify the path type of content published by individual users within the site. The purpose of reducing the authority weight is to avoid the misleading sense of credibility formed by the authoritative domain name shell + low-quality UGC content. S2-3. Extract key data assertions from the original answer text and perform cross-validation in authoritative third-party databases, including Statista and Gartner. When a key data assertion exists only in a single source and cannot be supported in the cross-validation, the key data assertion is marked as data illusion or fabricated data. Herein, key data assertions refer to verifiable quantitative statements and factual ranking statements. The cross-validation is used to provide external objective references, thereby distinguishing between textual consistency and factual reliability. S2-4. Perform sequence parsing on the list structure. When a target brand is detected to be in a preset sandwich ranking sequence position and only the target brand is accompanied by a discount code and purchase link, it is marked as a structured inducement anomaly. Further, the text logic flow is tested to check the closed-loop fit of the PAS model to form an explanation for the anomaly. The sandwich ranking sequence position refers to the specific position structure in which the target brand is fixed between two types of comparison items in several adjacent items in the recommendation list. That is, the items before and after it belong to the preset authoritative comparison item and the filler comparison item, respectively, thus forming a seemingly objective comparison package. The PAS closed-loop fit refers to judging the narrative structure of the answer text, identifying whether it contains three consecutive paragraphs and sentence groups of questions, aggravating questions, and solutions in sequence, and finally pointing the solution to the target brand or its purchase link, thus forming a marketing guidance closed loop. This is used to depict the marketing narrative path of the text from raising pain points, intensifying pain points, and leading to solutions, so that the anomaly conclusion not only gives the result, but also provides an understandable chain of causes. S2-5. Generate probe query set based on user query. ,in This includes a semantically equivalent subset of queries and a constrained subset of queries. The semantically equivalent subset refers to a set of queries that, while maintaining the same question intent, entity category, and task objective as the original user query, only rewrite the expression, making any rewritten query equivalent to the original query in terms of information requirement. The constrained subset refers to a set of queries that, while maintaining the original user query's question intent, impose explicit constraints on the generative model's output. These constraints include at least neutral expression constraints, verifiable source constraints, constraints excluding transformation-oriented content constraints, and constraints limiting the output scope. For each probe query, S2-1 to S2-4 are executed repeatedly, and the anomaly marking results for each round are recorded. The proportion of the target entity marked as an anomaly in each probe round is calculated. It is used to characterize the reproducibility of anomalies.

[0010] According to the above technical solution, S3 specifically refers to: S3-1, For entity sets For each entity in the calculation, the AI ​​volume is calculated. ,in The frequency of mention of an entity in the original answer text. This is a Boolean weighted value indicating whether an entity is the first recommended entity. The first preset value is used when the entity is the first recommended entity in the answer; the second preset value is used when the entity is not the first recommended entity. The emotional polarity score is used to quantify the intensity of positive modifications in the surrounding modifiers of the entity. , , Preset weighting coefficients are used to reflect the contribution of different factors to the volume of AI voice. S3-2. Sentiment polarity scores are calculated from entity-related text fragments using a sentiment analysis model. For each occurrence of the entity in the response text, a context segment containing the entity is extracted and input into a pre-trained sentiment analysis model. The difference between the positive and negative tendencies of this segment is output as the sentiment polarity score for that occurrence. When the entity appears multiple times in the response text, the average sentiment polarity score of each occurrence is taken and normalized to a value within [...]. The interval is 1,1]. The closer it is to 1, the stronger the positive modification; the closer it is to -1, the stronger the negative modification. S3-3, Based on the AI ​​volume of each entity Sort to obtain context ranking The smaller the value, the higher the ranking, which is used to characterize the recommendation strength order relationship of entities within the AI ​​answer.

[0011] According to the above technical solution, S3 further includes: S3-4. Obtain multi-source market signals related to the target entity through external data interfaces. These multi-source market signals include at least: search popularity signals, website traffic signals, and social media mention signals. Specifically, search popularity signals are provided by the Google Trends and SEMrush interfaces; website traffic signals are provided by the SimilarWeb interface; and social media mention signals are provided by the Reddit mention rate scanning and Twitter mention rate scanning interfaces, thus forming a multi-source observation set. ,in For the number of data sources, For the first Each data source outputs raw observations or observation sequences for the target entity. To ensure comparability between different data sources, [the following is done / is performed / etc.]. Perform uniform time window alignment to obtain windowed observations. ,in To pre-determine the market time window, and to eliminate the ranking drift caused by inconsistencies between short-term hot topics and long-term popularity metrics, Mapped to dimensionless contribution This ensures that scale differences from different sources do not affect subsequent fusion and ranking, and is used to unify the outputs of each source to a comparable metric domain, based on the dimensionless contribution of each data source. Perform source sorting on the candidate entity set to obtain the first... The ranking of each data source This is used to express market position under the data source caliber and form a source-specific ranking set. For the source ranking set Execute consistent fusion to obtain market rankings Its fusion rule adopts weighted robust aggregation, defined as follows: ,in For weighted robust aggregation operators, For the first The weight coefficients of each data source, all The sum is 1. A robust aggregation operator is used so that even when a data source experiences abnormal fluctuations, short-term distortions, or is subject to data manipulation, other data sources can still provide correction. It reflects the true market position under the meaning of multi-source consensus, thereby avoiding the erroneous amplification of single-source bias; S3-5. Construct a perturbation set of prompt words based on the user's original query. Each of them The algorithm maintains semantic consistency with the original query, while exhibiting controllable differences in expression structure, constraints, and wording style. The perturbation hints are used to trigger multiple outputs of the generative model under conditions of semantic invariance and expression variation, thereby evaluating the stability of the recommendation ranking. To determine the number of disturbance prompts, Each perturbation cue is input into the generative model to obtain the corresponding response text, and the context ranking for that round is generated according to existing AI volume calculation and ranking rules. This forms a context ranking sequence. ,in For the first The context ranking order of the target entity under the round-perturbation prompt; the smaller the value, the higher the ranking in the current round of responses. Executing robust aggregation yields stable context rankings. ,in The median operator is robust to extreme rounds, which can weaken the accidental bias triggered by a few prompt words, making... This is reflected in the typical recommendation positions under semantic consistency conditions, thereby achieving stable estimation that is resistant to prompt word perturbations. and Calculate the GEO suspicion index This is used to characterize the degree of disconnect between real-world position and AI-recommended position. The larger the indicator, the more it suggests an imbalance where AI is being pushed too hard but the market doesn't support it.

[0012] According to the above technical solution, S4 and S5 include: S4-1. Construct a verified golden dataset as a benchmark truth; the golden dataset consists of verifiable authoritative fact fragments and their corresponding sources, which is used to provide a definite reference for subsequent fidelity assessment and avoid making judgments based solely on the internal consistency of the model's response. S4-2. Calculate the loyalty index based on the aforementioned gold dataset. The generated answer is broken down into several factual assertions, and each assertion is checked against the retrieved supporting documents and the golden dataset. Assertions that can find consistent support in the supporting documents or the golden dataset are considered faithful; otherwise, they are considered unfaithful. The ratio of faithful assertions to the total number of factual assertions is used to measure whether the generated answer contains external knowledge not found in the retrieved document, identify the risk of illusion caused by supplemented external knowledge, and calculate the citation density index. The generated answer is divided into several output units. For each output unit, it is checked whether it has an accessible citation link or a locatable source identifier. If it does, it is counted as having a citation; otherwise, it is counted as having no citation. The ratio of the number of output units with citations to the total number of output units is used to measure the proportion of verifiable sources in the output units and to identify the risk of citation fraud that gives conclusions but lacks verifiable sources to support them. S4-3. Embed implicit watermarks during the content generation or original content publishing stage. During content generation, a preset key controls several optional expressions, such as synonym rewriting, word order fine-tuning, punctuation format variations, and low-order character patterns. This allows the text to form a set of statistically identifiable feature combinations while maintaining semantic integrity. These feature combinations constitute the watermark. Watermark decoding is performed on content copied and rewritten across platforms to trace the original content source and prevent anonymous poisoning. The aforementioned features are extracted from the test text and matched with the feature template corresponding to the key, calculating the matching statistics. When the statistics exceed a preset threshold, the watermark is determined to exist, and the source identifier is restored or the content propagation path is traced accordingly. It is an implicitly marked information that can be decoded. Its design goal is to maintain detectability even after the content has been copied, rewritten, or summarized, so that the path of content manipulation and dissemination can be traced with evidence. S5-1 outputs a bar chart comparing the AI ​​volume and the actual popularity, with a green shield icon indicating data consistency and a red warning light icon indicating data discrepancy; the difference bar chart is used to... The reflected AI-side exposure intensity and The market sentiment reflected is directly compared, allowing non-technical personnel to identify the contradiction between aggressive promotion and anonymous promotion without needing to understand the internal algorithm; S5-2. Output hover prompts for the ranking gap data in the visualization interface to present the context ranking. Market ranking The explanation of the differences, wherein the hover cues include at least the context ranking value, the market ranking value, and the imbalance index. The corresponding risk explanation statements make the alarm results readable and verifiable, thereby reducing the risk of unclear explanations caused by ambiguity.

[0013] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: This invention can identify pragmatic anomalies such as semantic emphasis bias, parasitic citations, and structured inducement at the generated text level, and output a quantifiable suspicion index at the statistical level based on the imbalance ratio between context ranking and market ranking, thereby transforming covert GEO manipulation behavior into an explainable, measurable, and thresholdable risk outcome; at the same time, it verifies fidelity and citation density through automated RAG evaluation, and achieves source tracking by combining highly transferable adversarial watermarking, thereby improving the reliability of judgment and reducing misjudgments caused by self-consistency illusion and citation deception; finally, it outputs red and green alerts and difference charts through a visual interactive interface, enabling users to intuitively understand ranking gaps and risk levels, and improving their ability to use generated content safely. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall modular structure of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This invention provides a technical solution: a generative engine optimization cheating detection system based on multidimensional feature analysis, comprising a data preprocessing module for receiving user queries and original answer text generated by a large language model, and performing named entity recognition and linguistic feature extraction on the original answer text to obtain standardized analysis objects; a detection and judgment module for performing heuristic linguistic feature analysis on the standardized analysis objects to identify pragmatic anomalies, and comparing AI recommendation ranking with real-world market data based on the principle of statistical imbalance to output quantifiable risk judgment results; and a source tracing and alarm module for performing automated RAG evaluation on the generated content to obtain fidelity and citation density indicators, and tracking the source of the content based on adversarial watermarking, thereby outputting alarm labels and difference charts in a visual manner to display the risk level. The data preprocessing module includes: an input interface unit for receiving user queries and raw response text generated by a large language model; an NLP preprocessing unit for performing named entity recognition and extracting linguistic features from the raw response text; and an entity normalization unit for performing aliasing and normalization mapping on the named entity recognition results to form an entity set that can be used for sorting and comparison. The detection and judgment module includes: a semantic emphasis deviation detection unit, used to calculate the density ratio of evaluative adjectives to descriptive nouns in the target entity description and mark anomalies accordingly; a parasitic citation detection unit, used to parse the domain name and path of the cited link and identify citation deception behavior pointing to user-generated content sub-paths under authoritative domains, and to perform a weight reduction on the authority weight of the citation; a fact silo verification unit, used to extract key data assertions and perform cross-validation in a third-party authoritative database to mark data illusions and fabricated data; a structured inducement identification unit, used to identify sandwich ranking in the list structure and detect abnormal conversion structures where only the target brand is accompanied by discount codes and purchase links, as well as detect the closed-loop fitting of the text logic flow to the PAS model; and an AI volume calculation unit. The system is used to calculate AI volume scores and form contextual rankings based on the frequency of mentions of entities in the response text, first recommendation, and sentiment polarity; a benchmark data acquisition unit is used to connect to Google Trends, SEMrush, SimilarWeb, Reddit mention rate scanning, and Twitter mention rate scanning interfaces through external data interfaces to generate real-world market rankings; an imbalance analysis engine is used to calculate the ratio of market ranking to contextual ranking and output high-risk alerts when preset threshold conditions are met; and an active adversarial probe unit is used to generate multiple sets of semantically equivalent questions and constraint questions based on user queries and repeatedly trigger heuristic language feature analysis and imbalance analysis to form an anomalous evidence chain for reproducible experiments. The source tracing and alerting module includes: an automated RAG evaluation unit, used to construct a golden dataset and calculate fidelity and citation density metrics; an adversarial source tracing unit, used to embed highly transferable implicit watermarks during the content generation or original content release stage and decode the watermarks after cross-platform copying and rewriting of the content to trace the original content source; and a visualization output terminal, used to display a comparison bar chart of AI volume and real popularity, with green shields and red warning lights indicating data consistency and data discrepancy respectively, and providing hover prompts for ranking discrepancies. A generative engine optimization cheat detection method based on multidimensional feature analysis includes the following steps: S1. Data Access and Preprocessing: Receive user queries and raw response text generated by the large language model, perform named entity recognition on the raw response text to form an entity set, perform normalization mapping on the entity set and extract the linguistic features required for subsequent analysis. S2. Linguistic Anomaly Detection and Probe Reproduction: Perform heuristic linguistic feature analysis on the original answer text to identify semantic emphasis bias, parasitic references and structured inducements. Construct multiple sets of equivalent and constraint questions based on user queries and repeatedly perform heuristic analysis. Aggregate the results of multiple rounds of analysis into anomaly evidence chain to improve the reproducibility and interpretability of the judgment. S3, Statistical Disorder Calculation: Calculate AI volume for entity sets and form contextual rankings. Obtain real-world market data through external data interfaces and generate market rankings. Compare contextual rankings with market rankings to output a disorder suspicion index and output risk levels based on thresholds. S4. Source Tracing and Fidelity Verification: Construct a gold dataset as the benchmark truth and perform fidelity and citation density evaluation on the generated content. Perform adversarial watermark decoding on the generated content to trace the source of the content and identify anonymous poisoning links. Link and store the evaluation results with the imbalance calculation results for auditing. S5. Visualization Output and Alerts: Generates visual charts to compare AI volume with real popularity, outputs green safety icons or red warning icons according to risk level, and outputs ranking gap and risk explanation information to prompt users to take appropriate action. S1 specifically includes: S1-1 Receiving Steps: The input interface unit receives the user's query text and the original answer text generated by the large language model, and binds the two into the same session record; wherein, the session record contains at least the query text, the answer text, the receiving timestamp, and the answer source identifier, so as to ensure that all subsequent detection results can be traced back to the same original input and output; S1-2, Entity Extraction Steps: Perform named entity recognition on the original answer text to obtain an entity set. ,in Each element in the code is a standardized name of a business entity; Named Entity Recognition takes the response text as input, outputs entity candidates and their position range in the text, and merges different spellings of the same entity into the same standardized name through entity standardization rules, thereby avoiding ranking and statistical bias caused by different spellings of the same name; it limits the scope of objects for all subsequent statistical, sorting, citation and tracing processing; S1-3, targeting For each business entity, extract the fragments of its appearance in the original answer text and generate feature records for subsequent semantic emphasis bias detection, parasitic reference detection, and structured induced recognition. The feature records include at least: the number of times the entity appears, the position of the entity's first appearance, the fragments of modifiers around the entity, the set of reference links corresponding to the entity, and the paragraph and list structure information of the entity. The feature records are used to convert the natural language description in the text into structured input that can be stably processed by the rule engine, thereby ensuring the consistency and verifiability of the detection conclusions. S2 includes: S2-1, Count the number of evaluative adjectives in the descriptive fragment for each entity. and the number of descriptive nouns Calculate the density ratio Evaluative adjectives refer to words that subjectively praise, criticize, strongly recommend, or exaggerate an entity, while descriptive nouns refer to nouns used to objectively describe product categories, functions, technical features, and service models. The larger the value, the more marketing-oriented the description becomes. Compared to competing products Exceeding the preset deviation threshold When this happens, the target entity is marked as having a semantic focus deviation anomaly; S2-2. Parse each cited link in the original answer text and extract the domain name field and path field. When the domain name field belongs to the preset authoritative domain name set and the path field belongs to the user-generated content sub-path set, mark the citation as a parasitic citation and reduce its authority weight for subsequent fact silo verification. The authoritative domain name set is used to identify site domain names that are generally recognized as highly credible sources, and the user-generated content sub-path set is used to identify the path type of content published by individual users within the site. The purpose of reducing the authority weight is to avoid the misleading sense of credibility formed by the authoritative domain name shell + low-quality UGC content. S2-3. Extract key data assertions from the original answer text and perform cross-validation on authoritative third-party databases, including Statista and Gartner. When a key data assertion exists only in a single source and cannot be supported by cross-validation, the key data assertion is marked as data illusion or fabricated data. Here, key data assertions refer to verifiable quantitative statements and factual ranking statements. Cross-validation is used to provide external objective references, thereby distinguishing between textual consistency and factual reliability. S2-4. Perform sequence parsing on the list structure. When a target brand is detected to be in a preset sandwich ranking sequence position and only the target brand is accompanied by a discount code and purchase link, it is marked as a structured inducement anomaly. Further, the text logic flow is tested to check the closed-loop fit of the PAS model to form an explanation for the anomaly. The sandwich ranking sequence position refers to the specific position structure in which the target brand is fixed between two types of comparison items in several adjacent items in the recommendation list. That is, the items before and after it belong to the preset authoritative comparison item and the filler comparison item, respectively, thus forming a seemingly objective comparison package. The PAS closed-loop fit refers to judging the narrative structure of the answer text, identifying whether it contains three consecutive paragraphs and sentence groups of questions, aggravating questions, and solutions in sequence, and finally pointing the solution to the target brand or its purchase link, thus forming a marketing guidance closed loop. This is used to depict the marketing narrative path of the text from raising pain points, intensifying pain points, and leading to solutions, so that the anomaly conclusion not only gives the result, but also provides an understandable chain of causes. This invention identifies parasitic references to authoritative domain shells combined with UGC sub-paths, and combines narrative structure features such as sandwich ranking and PAS closed loops to deconstruct seemingly objective recommendation packaging into a verifiable structured pattern. The innovation of this step lies in its grasp of common manipulation paths in generative engines—not by directly fabricating facts, but by guiding user trust and conversion through the appearance of references and narrative structure—thus enabling structured identification of soft manipulation without relying on the veracity of a single fact, providing a more realistic attack surface detection tool for GEO cheating.

[0017] S2-5. Generate probe query set based on user query. ,in This includes a semantically equivalent subset of queries and a constrained subset of queries. The semantically equivalent subset refers to a set of queries that, while maintaining the same question intent, entity category, and task objective as the original user query, only rewrite the expression, making any rewritten query equivalent to the original query in terms of information requirement. The constrained subset refers to a set of queries that, while maintaining the original user query's question intent, impose explicit constraints on the generative model's output. These constraints include at least neutral expression constraints, verifiable source constraints, constraints excluding transformation-oriented content constraints, and constraints limiting the output scope. For each probe query, S2-1 to S2-4 are executed repeatedly, and the anomaly marking results for each round are recorded. The proportion of the target entity marked as an anomaly in each probe round is calculated. It is used to characterize the reproducibility of anomalies; This invention not only detects anomalies once, but also proactively generates semantically equivalent and constraint questions and repeatedly triggers the detection process. It statistically analyzes the recurrence of anomalies across multiple rounds of sampling to distinguish between accidental ranking fluctuations and stability manipulation. The innovation of this step lies in its transformation of anti-cheating from static scoring to dynamic reproducibility experiments. By using multi-round consistent evidence to suppress random interference in generative systems, the judgment results are more convincing and resistant to controversy, making it particularly suitable for scenarios like GEO that rely on cue words and narrative structure manipulation.

[0018] S3 specifically refers to: S3-1, For entity sets For each entity in the calculation, the AI ​​volume is calculated. ,in The frequency of mention of an entity in the original answer text. This is a Boolean weighted value indicating whether an entity is the first recommended entity. The first preset value is used when the entity is the first recommended entity in the answer; the second preset value is used when the entity is not the first recommended entity. The emotional polarity score is used to quantify the intensity of positive modifications in the surrounding modifiers of the entity. , , Preset weighting coefficients are used to reflect the contribution of different factors to the volume of AI voice. S3-2. Sentiment polarity scores are calculated from entity-related text fragments using a sentiment analysis model. For each occurrence of an entity in the response text, a context fragment containing that entity is extracted and input into a pre-trained sentiment analysis model. The difference between the positive and negative sentiment of this fragment is output as the sentiment polarity score for that occurrence. When the entity appears multiple times in the response text, the average sentiment polarity score of each occurrence is taken, and normalization is limited to […]. The interval is 1,1]. The closer it is to 1, the stronger the positive modification; the closer it is to -1, the stronger the negative modification. S3-3, Based on the AI ​​volume of each entity Sort to obtain context ranking The smaller the value, the higher the ranking, which is used to characterize the recommendation strength order relationship of entities within the AI ​​answer; S3 also includes: S3-4. Obtain multi-source market signals related to the target entity through external data interfaces. These multi-source market signals include at least: search popularity signals, website traffic signals, and social media mention signals. Specifically, search popularity signals are provided by the Google Trends and SEMrush interfaces, website traffic signals by the SimilarWeb interface, and social media mention signals by the Reddit and Twitter mention rate scanning interfaces, thus forming a multi-source observation set. ,in For the number of data sources, For the first Each data source outputs raw observations or observation sequences for the target entity. To ensure comparability between different data sources, [the following is done / is performed / etc.]. Perform uniform time window alignment to obtain windowed observations. ,in To pre-determine the market time window, and to eliminate the ranking drift caused by inconsistencies between short-term hot topics and long-term popularity metrics, Mapped to dimensionless contribution This ensures that scale differences from different sources do not affect subsequent fusion and ranking, and is used to unify the outputs of each source to a comparable metric domain, based on the dimensionless contribution of each data source. Perform source sorting on the candidate entity set to obtain the first... The ranking of each data source This is used to express market position under the data source caliber and form a source-specific ranking set. For the source ranking set Execute consistent fusion to obtain market rankings Its fusion rule adopts weighted robust aggregation, defined as follows: ,in For weighted robust aggregation operators, For the first The weight coefficients of each data source, all The sum is 1. A robust aggregation operator is used so that even when a data source experiences abnormal fluctuations, short-term distortions, or is subject to data manipulation, other data sources can still provide correction. It reflects the true market position under the meaning of multi-source consensus, thereby avoiding the erroneous amplification of single-source bias; S3-5. Construct a perturbation set of prompt words based on the user's original query. Each of them The algorithm maintains semantic consistency with the original query, while exhibiting controllable differences in expression structure, constraints, and wording style. Perturbation cues are used to trigger multiple outputs of the generative model under conditions of semantic invariance and expression variation, thereby evaluating the stability of the recommendation ranking. To determine the number of disturbance prompts, Each perturbation cue is input into the generative model to obtain the corresponding response text, and the context ranking for that round is generated according to existing AI volume calculation and ranking rules. This forms a context ranking sequence. ,in For the first The context ranking order of the target entity under the round-perturbation prompt; the smaller the value, the higher the ranking in the current round of responses. Executing robust aggregation yields stable context rankings. ,in The median operator is robust to extreme rounds, which can weaken the accidental bias triggered by a few prompt words, making... This is reflected in the typical recommendation positions under semantic consistency conditions, thereby achieving stable estimation that is resistant to prompt word perturbations. and Calculate the GEO suspicion index This is used to characterize the degree of disconnect between real-world position and AI-recommended position. The larger the indicator, the more it suggests an imbalance where AI is being aggressively promoted but the market does not support it. This invention does not simplify true market position to a single platform indicator. Instead, it incorporates diverse signals such as search popularity, website traffic, and social media mentions, and outputs a multi-source consensus market ranking in a robust and integrated manner. The innovation of this step lies in its transformation of market signals from the external world—which are inherently noisy and potentially manipulated by inflated metrics—into a benchmark that is comparable, verifiable, and insensitive to single-source anomalies. This provides a reliable reference for subsequent imbalance assessments and avoids false positives and false negatives caused by single-source biases in traditional methods.

[0019] S4 and S5 include: S4-1. Construct a verified golden dataset as a benchmark truth; the golden dataset consists of verifiable authoritative fact fragments and their corresponding sources, which is used to provide a definite reference for subsequent fidelity assessment and avoid making judgments based solely on the internal consistency of the model's response. S4-2 Calculating Loyalty Indicators Based on the Gold Dataset The generated answer is broken down into several factual assertions, and each assertion is checked against the retrieved supporting documents and the golden dataset. Assertions that can be consistently supported in either the supporting documents or the golden dataset are considered faithful; otherwise, they are considered unfaithful. The ratio of faithful assertions to the total number of factual assertions is used to measure whether the generated answer contains external knowledge not found in the retrieved document, identify the risk of illusion caused by supplemented external knowledge, and calculate the citation density index. The generated answer is divided into several output units. For each output unit, it is checked whether it has an accessible citation link or a locatable source identifier. If it does, it is counted as having a citation; otherwise, it is counted as having no citation. The ratio of the number of output units with citations to the total number of output units is used to measure the proportion of verifiable sources in the output units and to identify the risk of citation fraud that gives conclusions but lacks verifiable sources to support them. S4-3. Embed implicit watermarks during the content generation or original content publishing stage. During content generation, a preset key controls several optional expressions, such as synonym rewriting, word order fine-tuning, punctuation format variations, and low-order character patterns. This allows the text to form a set of statistically identifiable feature combinations while maintaining semantic integrity. These feature combinations constitute the watermark. Watermark decoding is performed on content copied and rewritten across platforms to trace the original content source and prevent anonymous poisoning. The aforementioned features are extracted from the test text and matched with the feature template corresponding to the key, calculating the matching statistics. When the statistics exceed a preset threshold, the watermark is determined to exist, and the source identifier is restored or the content propagation path is traced accordingly. It is an implicitly marked information that can be decoded. Its design goal is to maintain detectability even after the content has been copied, rewritten, or summarized, so that the path of content manipulation and dissemination can be traced with evidence. S5-1 outputs a bar chart comparing the AI ​​volume and the actual popularity, with a green shield icon indicating data consistency and a red warning light icon indicating data discrepancy; the difference bar chart is used to... The reflected AI-side exposure intensity and The market sentiment reflected is directly compared, allowing non-technical personnel to identify the contradiction between aggressive promotion and anonymous promotion without needing to understand the internal algorithm; S5-2. Output hover tooltip information for the ranking gap data in the visualization interface to present the context of the ranking. Market ranking The explanation of the differences, including hover cues, includes at least the context ranking value, market ranking value, and imbalance index. The corresponding risk explanation statements make the alarm results readable and verifiable, thereby reducing the risk of unclear explanations caused by ambiguity.

[0020] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0021] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A generative engine optimization cheating detection system based on multidimensional feature analysis, characterized in that: include: The data preprocessing module is used to receive user queries and raw response text generated by a large language model, and to perform named entity recognition and linguistic feature extraction on the raw response text to obtain standardized analysis objects. The detection and judgment module is used to perform heuristic language feature analysis on the standardized analysis object to identify pragmatic anomalies, and compare the AI ​​recommendation ranking with real-world market data based on the principle of statistical imbalance to output quantifiable risk judgment results; the source tracing and alarm module is used to perform automated RAG evaluation on the generated content to obtain fidelity and citation density indicators, and realize content source tracing based on adversarial watermarking.

2. The cheat detection system for a generative engine based on multidimensional feature analysis according to claim 1, characterized in that: The data preprocessing module includes: an input interface unit for receiving user queries and the original response text generated by the large language model; an NLP preprocessing unit for performing named entity recognition on the original response text and extracting linguistic features; and an entity normalization unit for performing aliasing and normalization mapping on the named entity recognition results to form an entity set that can be used for sorting and comparison. The detection and judgment module includes: a semantic emphasis deviation detection unit, used to calculate the density ratio of evaluative adjectives to descriptive nouns in the target entity description and mark anomalies accordingly; a parasitic citation detection unit, used to parse the domain name and path of the cited link and identify citation deception behavior pointing to user-generated content sub-paths under authoritative domains, and perform a weight reduction on the authority weight of the citation; a fact silo verification unit, used to extract key data assertions and perform cross-validation in a third-party authoritative database to mark data illusions and fabricated data; a structured inducement identification unit, used to identify sandwich ranking in the list structure and detect abnormal conversion structures where only the target brand is accompanied by discount codes and purchase links, and to detect the closed-loop fitting of the text logic flow to the PAS model; and an AI volume calculation unit, used for... The system calculates AI volume scores based on the frequency of mentions of entities in the response text, the top recommendation, and sentiment polarity, and forms a contextual ranking. A benchmark data acquisition unit connects to external data interfaces such as Google Trends, SEMrush, SimilarWeb, Reddit mention rate scanning, and Twitter mention rate scanning to generate real-world market rankings. An imbalance analysis engine calculates the ratio of the market ranking to the contextual ranking and outputs a high-risk alert when a preset threshold condition is met. An active adversarial probe unit generates multiple sets of semantically equivalent and constraint questions based on the user query and repeatedly triggers the heuristic language feature analysis and imbalance analysis to form an anomalous evidence chain that can reproduce the experiment. The source tracing and alarm module includes: an automated RAG evaluation unit, used to construct a golden dataset and calculate fidelity and citation density metrics; an adversarial source tracing unit, used to embed highly transferable implicit watermarks during the content generation or original content publication stage and decode the watermarks after cross-platform copying and rewriting of the content to trace the original content source; and a visualization output terminal, used to display a comparison bar chart of AI volume and real popularity, and to indicate data consistency and data discrepancy states with green shields and red warning lights, respectively, and to provide hover prompts for ranking discrepancies.

3. The method for determining cheating in a generative engine based on multidimensional feature analysis according to claim 2, characterized in that: Includes the following steps: S1. Data Access and Preprocessing: Receive user queries and raw response text generated by the large language model, perform named entity recognition on the raw response text to form an entity set, perform normalization mapping on the entity set and extract the linguistic features required for subsequent analysis. S2. Linguistic Anomaly Detection and Probe Reproduction: Perform heuristic linguistic feature analysis on the original answer text to identify semantic emphasis bias, parasitic references and structured inducements. Construct multiple sets of equivalent and constraint questions based on user queries and repeatedly perform the heuristic analysis. Aggregate the results of multiple rounds of analysis into an anomaly evidence chain to improve the reproducibility and interpretability of the judgment. S3, Statistical Disorder Calculation: Calculate AI volume for entity sets and form contextual rankings. Obtain real-world market data through external data interfaces and generate market rankings. Compare contextual rankings with market rankings to output a disorder suspicion index and output risk levels based on thresholds. S4. Source Tracing and Fidelity Verification: Construct a gold dataset as the benchmark truth and perform fidelity and citation density evaluation on the generated content. Perform adversarial watermark decoding on the generated content to trace the source of the content and identify anonymous poisoning links. Link and store the evaluation results with the imbalance calculation results for auditing. S5. Visualization Output and Alerts: Generates visual charts to compare AI volume with real popularity, outputs green safety icons or red warning icons according to risk level, and outputs ranking gap and risk explanation information to prompt users to take appropriate action.

4. The method for determining cheating in a generative engine based on multidimensional feature analysis according to claim 3, characterized in that: S1 specifically includes: S1-1 Receiving Steps: The input interface unit receives the user's query text and the original answer text generated by the large language model, and binds the two into the same session record; wherein, the session record contains at least the query text, the answer text, the receiving timestamp, and the answer source identifier; S1-2, Entity Extraction Step: Perform named entity recognition on the original answer text to obtain an entity set. ,in Each element in the text is a normalized name of a business entity; the named entity recognition takes the response text as input, outputs entity candidates and their position range in the text, and merges different spellings of the same entity into the same normalized name through entity standardization rules; S1-3, targeting For each business entity in the original answer text, extract the fragments in which the entity appears and generate feature records for subsequent semantic emphasis bias detection, parasitic reference detection and structured induced recognition; the feature records include at least: the number of times the entity appears, the position of the first appearance of the entity, the fragments of modifiers around the entity, the set of reference links corresponding to the entity, and the paragraph and list structure information of the entity.

5. The method for determining cheating in a generative engine based on multidimensional feature analysis according to claim 3, characterized in that: S2 includes: S2-1, Count the number of evaluative adjectives in the descriptive fragment for each entity. and the number of descriptive nouns Calculate the density ratio The evaluative adjectives refer to words that subjectively praise, criticize, strongly recommend, or exaggerate an entity, while the descriptive nouns refer to nouns used to objectively describe product categories, functions, technical features, and service models. The larger the value, the more marketing-oriented the description becomes. Compared to competing products Exceeding the preset deviation threshold When this happens, the target entity is marked as having a semantic focus deviation anomaly; S2-2. Parse each cited link in the original answer text and extract the domain name field and path field. When the domain name field belongs to the preset authoritative domain name set and the path field belongs to the user-generated content sub-path set, mark the citation as a parasitic citation and reduce its authority weight for subsequent fact silo verification. The authoritative domain name set is used to identify site domain names that are generally recognized as highly credible sources, and the user-generated content sub-path set is used to identify the path type of content published by individual users within the site. S2-3. Extract key data assertions from the original answer text and perform cross-validation in authoritative third-party databases, including Statista and Gartner. When a key data assertion exists only in a single source of reference and cannot be supported in the cross-validation, the key data assertion is marked as data illusion or fabricated data. Herein, key data assertions refer to verifiable quantitative statements and factual ranking statements. S2-4. Perform sequence parsing on the list structure. When a target brand is detected to be in a preset sandwich ranking sequence position and only the target brand is accompanied by a discount code and purchase link, it is marked as a structured inducement anomaly. Further, the text logic flow is tested to check the closed-loop fit of the PAS model to form an explanation for the anomaly. The sandwich ranking sequence position refers to the specific position structure in which the target brand is fixed between two types of comparison items in several adjacent items in the recommendation list. That is, the items before and after it belong to the preset authoritative comparison item and the filler comparison item, respectively, thus forming a seemingly objective comparison package. The PAS closed-loop fit refers to judging the narrative structure of the answer text, identifying whether it contains three consecutive paragraphs and sentence groups in sequence: question, exaggeration of the question, and solution, and finally points the solution to the target brand or its purchase link, thus forming a marketing guidance closed loop. S2-5. Generate probe query set based on user query. ,in This includes a semantically equivalent subset of queries and a constrained subset of queries. The semantically equivalent subset refers to a set of queries that, while maintaining the same question intent, entity category, and task objective as the original user query, only rewrite the expression, ensuring that any rewritten query is equivalent to the original query in terms of information requirement. The constrained subset refers to a set of queries that, while maintaining the original user query's question intent, impose explicit constraints on the generative model's output. These constraints include at least neutral expression constraints, verifiable source constraints, constraints excluding transformation-oriented content constraints, and constraints limiting the output range. For each probe query, S2-1 to S2-4 are executed repeatedly, and the anomaly marking results for each round are recorded. The proportion of the target entity marked as an anomaly in each probe round is calculated. It is used to characterize the reproducibility of anomalies.

6. The method for determining cheating in a generative engine based on multidimensional feature analysis according to claim 3, characterized in that: Specifically, S3 is: S3-1, For entity sets For each entity in the calculation, the AI ​​volume is calculated. ,in The frequency of mention of an entity in the original answer text. This is a Boolean weighted value indicating whether an entity is the first recommended entity. The first preset value is used when the entity is the first recommended entity in the answer; the second preset value is used when the entity is not the first recommended entity. The emotional polarity score is used to quantify the intensity of positive modifications in the surrounding modifiers of the entity. , , Preset weighting coefficients; S3-2. Sentiment polarity scores are calculated from entity-related text fragments using a sentiment analysis model. For each occurrence of the entity in the response text, a context fragment containing the entity is extracted and input into a pre-trained sentiment analysis model. The difference between the positive and negative tendencies of the fragment is output as the sentiment polarity score for that occurrence. When the entity appears multiple times in the response text, the average sentiment polarity score of each occurrence is taken, and normalization is limited to [ ]. The interval is [1,1]. S3-3, Based on the AI ​​volume of each entity Sort to obtain context ranking The smaller the value, the higher the ranking.

7. The method for determining cheating in a generative engine based on multidimensional feature analysis according to claim 6, characterized in that: S3 further includes: S3-4. Obtain multi-source market signals related to the target entity through external data interfaces. These multi-source market signals include at least: search popularity signals, website traffic signals, and social media mention signals. Specifically, search popularity signals are provided by the Google Trends and SEMrush interfaces; website traffic signals are provided by the SimilarWeb interface; and social media mention signals are provided by the Reddit mention rate scanning and Twitter mention rate scanning interfaces, thus forming a multi-source observation set. ,in For the number of data sources, For the first Each data source outputs raw observations or observation sequences for the target entity. To ensure comparability between different data sources, [the following is done / is performed / etc.]. Perform uniform time window alignment to obtain windowed observations. ,in To pre-determine the market time window, and to eliminate the ranking drift caused by inconsistencies between short-term hot topics and long-term popularity metrics, Mapped to dimensionless contribution This ensures that scale differences from different sources do not affect subsequent fusion and ranking, and is used to unify the outputs of each source to a comparable metric domain, based on the dimensionless contribution of each data source. Perform source sorting on the candidate entity set to obtain the first... Source ranking corresponding to each data source This is used to express market position under the data source caliber and form a source-specific ranking set. For the source ranking set Execute consistent fusion to obtain market rankings Its fusion rule adopts weighted robust aggregation, defined as follows: ,in For weighted robust aggregation operators, For the first The weight coefficients of each data source, all The sum of is 1; S3-5. Construct a perturbation set of prompt words based on the user's original query. Each of them The algorithm maintains semantic consistency with the original query, while exhibiting controllable differences in expression structure, constraints, and wording style. The perturbation hints are used to trigger multiple outputs of the generative model under conditions of semantic invariance and expression variation, thereby evaluating the stability of the recommendation ranking. To determine the number of disturbance prompts, Each perturbation cue is input into the generative model to obtain the corresponding response text, and the context ranking for that round is generated according to existing AI volume calculation and ranking rules. This forms a context ranking sequence. ,in For the first The context ranking order of the target entity under the round-perturbation prompt; the smaller the value, the higher the ranking in the current round of responses. Executing robust aggregation yields stable context rankings. ,in The median operator is robust to extreme rounds, which can weaken the accidental bias triggered by a few prompt words, making... This is reflected in the typical recommendation positions under semantic consistency conditions, thereby achieving stable estimation that is resistant to prompt word perturbations. and Calculate the GEO suspicion index This is used to characterize the degree of disconnect between real-world position and AI-recommended position. The larger the indicator, the more it suggests an imbalance where AI is being pushed too hard but the market doesn't support it.

8. The method for determining cheating in a generative engine based on multidimensional feature analysis according to claim 3, characterized in that: S4 and S5 include: S4-1. Construct a verified golden dataset as a benchmark of truth; wherein, the golden dataset consists of verifiable authoritative fact fragments and their corresponding sources; S4-2. Calculate the loyalty index based on the aforementioned gold dataset. The generated answer is broken down into several factual assertions, and each assertion is checked against the retrieved supporting documents and the golden dataset. Assertions that can find consistent support in the supporting documents or the golden dataset are considered faithful; otherwise, they are considered unfaithful. The ratio of faithful assertions to the total number of factual assertions is used to measure whether the generated answer contains external knowledge not found in the retrieved document, identify the risk of illusion caused by supplemented external knowledge, and calculate the citation density index. The generated answer is divided into several output units. For each output unit, it is checked whether it has an accessible citation link or a locatable source identifier. If it does, it is counted as having a citation; otherwise, it is counted as having no citation. This is the ratio of the number of referenced output units to the total number of output units; S4-3. Embed implicit watermarks during the content generation or original content publishing stage. During content generation, a preset key controls several optional expressions, such as synonym rewriting, word order fine-tuning, punctuation format variations, and low-order character patterns. This allows the text to form a set of statistically identifiable feature combinations while maintaining semantic integrity. These feature combinations constitute the watermark. Watermark decoding is performed on content copied and rewritten across platforms to trace the original content source and prevent anonymous poisoning. The aforementioned features are extracted from the test text and matched with the feature template corresponding to the key, calculating the matching statistics. When the statistics exceed a preset threshold, the watermark is determined to exist, and the source identifier is restored or the content propagation path is traced accordingly. This is implicitly tagged information that can be decoded; S5-1 outputs a bar chart comparing the AI ​​volume and the actual popularity, with a green shield icon indicating data consistency and a red warning light icon indicating data discrepancy; the difference bar chart is used to... The reflected AI-side exposure intensity and A direct comparison can be made between the market sentiment reflected in the data; S5-2. Output hover prompts for the ranking gap data in the visualization interface to present the context ranking. Market ranking The explanation of the differences, wherein the hover cues include at least the context ranking value, market ranking value, and imbalance index. The corresponding risk explanation statement.

Images