Content rewriting method and system considering e-commerce GEO visibility improvement and user experience

Abstract

This invention discloses a content rewriting method and system that balances improved visibility and user experience in e-commerce generative search (GEO). The method includes the following steps: acquiring original content, a target query, and user feedback corpus; rewriting the original content to generate candidate rewritten content; constructing a candidate set based on the candidate rewritten content and interfering content, and inputting it into a generative search environment with the target query to obtain a generative response; calculating the visibility score of the candidate rewritten content in the corresponding generative response; calculating the user experience score of the candidate rewritten content; constructing a comprehensive objective function based on the visibility score and the user experience score; performing dual-objective constraint filtering based on the comprehensive objective function; selecting candidate rewritten versions that meet the constraints as the target rewritten content; and compensatingly rewriting candidate rewritten versions that do not meet the dual-objective constraints until a candidate rewritten version that meets the dual-objective constraints is obtained. This invention enables synergistic optimization between visibility and user experience in e-commerce generative search.

Description

Technical Field

[0001] This invention relates to the field of intelligent content rewriting technology, specifically to a content rewriting method and system that balances improved visibility for e-commerce GEOs with enhanced user experience. Background Technology

[0002] With the deep integration of large language models and information retrieval systems, the presentation of e-commerce search results is gradually shifting from traditional link lists to generative responses. In e-commerce scenarios, users obtain comprehensive answers such as product comparisons, functional interpretations, and applicable scenario analyses through natural language queries. Generative search engines directly output the integrated answers and annotate the sources of information. Unlike the visibility mechanism in traditional e-commerce search, which is centered on ranking position and click-through rate, in the generative search environment, product visibility is mainly reflected in whether the content is selected by the model, whether it is cited, the position of the citation in the response, the citation span, and the strength of the textual contribution to the final answer. These factors collectively determine whether product content can gain exposure. Therefore, the research focus on optimizing e-commerce generative search engines has shifted from traditional keyword and link ranking to optimizing the semantic organization, structured expression, evidence organization, terminology, and text style of product content.

[0003] However, for e-commerce products, visibility only solves the problem of being seen by users, while user experience truly determines the conversion rate that leads to user trust and purchase. From the perspective of classic user behavior models, the AIDA model summarizes the consumer decision-making process into four stages: attention, interest, desire, and action. In generative search, visibility primarily plays a role in stimulating attention and interest, enabling product content to be selected by the model and presented to the user. User experience, on the other hand, permeates the entire process of interest transforming into desire and action, directly determining whether users can obtain sufficient decision-making support information from the content, such as realistic attribute descriptions, usage scenario explanations, disclosure of limitations, multi-dimensional comparison clues, and credible details, thereby completing the purchase. If content is overly compressed or templated during optimization, even if it is frequently referenced by the model (i.e., achieving high visibility), it is difficult to generate sufficient user drive at the desire and memory levels, ultimately resulting in a damaged conversion rate.

[0004] Most existing generative search optimization methods primarily aim to improve machine-side visibility. These methods typically employ authoritative expressions, statistical enhancements, keyword enhancements, citation enhancements, technical terminology enhancements, fluency optimizations, and comprehensibility improvements to enhance the model's recognizability, comparability, and integrability of e-commerce content. While such rewriting methods can increase the probability of product content being cited and presented in generative responses (i.e., increasing exposure) in some scenarios, existing methods often assume that improved visibility equates to successful optimization. They lack simultaneous constraints on the end-user experience, easily leading to the compression of crucial user-reliant decision-making cues (such as applicable conditions, material details, usage limitations, and realistic comparisons) during content rewriting. This results in over-optimization where the model is easier to cite, but user experience is worse, and conversion rates decrease.

[0005] Furthermore, existing technologies typically evaluate performance only after rewriting is complete, lacking a technical solution to transform core user concerns (especially key information affecting conversion rates in e-commerce scenarios) into explicit constraints during the rewriting generation phase. Therefore, if the rewriting process excessively reinforces the structured, standardized, templated, or singular expression of model preferences, it may impair the comprehensiveness, readability, credibility, or decision support of the content while increasing visibility (exposure). This could cause optimization strategies to fall into a conflict zone where increased visibility leads to decreased user experience and reduced conversion rates.

[0006] In summary, existing e-commerce generative search optimization technologies suffer from at least the following shortcomings: First, they lack a dual-objective constrained rewriting mechanism that balances machine-side visibility (exposure) and user-side experience (conversion rate); second, they lack technical means to transform frequently asked questions from user feedback into rewriting constraints and retain and compensate for them in candidate content; and third, they lack a closed-loop optimization process encompassing candidate rewriting generation, generative response evaluation, dual-objective screening, and compensatory rewriting. Therefore, there is an urgent need for a content rewriting technology that not only improves product visibility in e-commerce generative search but also maintains or enhances user experience, thereby achieving synergistic optimization of exposure and conversion rate. Summary of the Invention

[0007] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a content rewriting method and system that balances e-commerce GEO visibility improvement and user experience. Instead of simply setting the rewriting goal as maximizing visibility, this invention incorporates both the machine-side visibility improvement goal and the user-side experience preservation goal from e-commerce generative search into the rewriting process. Through user core concern extraction, candidate rewriting generation, generative response acquisition, visibility calculation, user experience calculation, dual-objective constraint filtering, and compensatory rewriting, it outputs rewritten content that balances visibility and experience. Furthermore, this invention does not rely solely on a single visibility metric for rewriting optimization. Instead, it combines explicit demand coverage analysis based on a core concern set and implicit user perception evaluation based on a large language model to jointly construct a user experience proxy metric, which serves as a constraint and optimization goal in the rewriting process. This achieves synergistic optimization between e-commerce generative search visibility and user experience.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a content rewriting method that balances improved e-commerce GEO visibility and user experience, comprising the following steps:

[0010] Obtain raw content, target queries, and user feedback corpora;

[0011] The original content is rewritten based on a preset set of rewriting actions to generate candidate rewritten content.

[0012] A candidate set is constructed based on the candidate rewritten content and the interference content, and then input into the generative search environment with the target query to obtain a generative response;

[0013] The generative response is segmented into multiple fragments, and the visibility score of the candidate rewritten content in the corresponding generative response is calculated.

[0014] The user feedback corpus is segmented and filtered to calculate the user experience score of the candidate rewritten content;

[0015] A comprehensive objective function is constructed based on visibility score and user experience score;

[0016] Based on the comprehensive objective function, a dual objective constraint of visibility and user experience is used for screening. Candidate rewrite versions that meet the dual objective constraints are used as the target rewrite content, while candidate rewrite versions that do not meet the dual objective constraints are rewritten compensatorily until a candidate rewrite version that meets the dual objective constraints is obtained.

[0017] As a preferred technical solution, the rewriting actions in the preset rewriting action set adopt multi-strategy rewriting actions based on controllable text generation, and generate a variety of candidate rewriting content by applying differentiated constraints to the original content.

[0018] As a preferred technical solution, the generative response is represented as follows:

[0019] ;

[0020] ;

[0021] in, Represents a generative response. This indicates a generative search environment. Indicates the target query. Indicates the randomness of context organization. This indicates that the model performs randomness. This represents a candidate set constructed based on candidate rewritten content and interfering content. Indicates the candidate rewrite content. This indicates interfering content.

[0022] As a preferred technical solution, a candidate set is constructed based on candidate rewritten content and interfering content, and input into a generative search environment with the target query to obtain a generative response, specifically including:

[0023] Generative search engines calculate the fit score of each candidate rewritten content based on its relevance to the target query, information completeness, and model preferences. They then rank the top n candidate rewritten contents from highest to lowest fit score and generate a generative response based on the ranked set of candidate rewritten contents.

[0024] As a preferred technical solution, the generative response is output in the form of an ordered list with source numbers, and each response entry references the number of the candidate rewritten content in the original candidate set.

[0025] As a preferred technical solution, the generative response is divided into multiple segments, and the visibility score of the candidate rewritten content in the corresponding generative response is calculated, specifically including:

[0026] Decompose the generative response into Given a set of ordered segments, extract the source ID, word count, and position of each segment. Assume the segment length is 1. The source of its attribution is The visibility contribution of candidate rewritten content in this response is calculated and expressed as:

[0027] ;

[0028] in, Indicates the position decay weight;

[0029] For the same rewrite action, responses are repeatedly generated under multiple random seeds, and candidate rewrite content is calculated for each rewrite action. The expected visibility score is:

[0030] ;

[0031] in, Indicates the randomness of context organization. This indicates that the model performs randomness. This represents the expectation of the visibility score for repeatedly generated responses under multiple random seeds, taking into account the randomness of context organization and model execution.

[0032] As a preferred technical solution, the user experience score of the candidate rewritten content is calculated, specifically including:

[0033] Based on the generative response, extract all sentence fragments of the corresponding candidate rewritten content, output multi-dimensional user perception effect scores through a large language model, and calculate the mean of the multi-dimensional scores;

[0034] Each text in the user feedback corpus is segmented according to sentence boundaries, and each text is divided into an independent sentence sequence. Based on a preset stop word list, the initial words in each sentence are filtered to build a candidate word set, and the candidate words in each sentence are deduplicated.

[0035] Two types of candidate units are constructed, including single candidate words and bigrams consisting of two consecutive candidate words;

[0036] The comprehensive metric is calculated and expressed as follows:

[0037] ;

[0038] ;

[0039] in, This represents a comprehensive metric. Represents point mutual information, Indicates the proportion of positive correlation. To perform a logarithmic transformation on the number of sentences, This indicates the number of sentences appearing in the candidate unit. This represents the number of sentences that co-occur with the positive seed word, and T represents the total number of sentences in the user feedback corpus. The total number of sentences in the entire corpus that contain at least one positive seed word;

[0040] Candidate units are sorted from highest to lowest based on their comprehensive metrics. Core concerns are selected, and the number of sentences containing each core concern is counted to calculate its weight. :

[0041] ;

[0042] in, Indicate core concerns, This indicates the number of sentences containing each core concern item;

[0043] Calculate the weighted hit rate of core concerns , is represented as:

[0044] ;

[0045] in, Indicates candidate rewrite content For core concerns The hit indicator function;

[0046] The original experience score is obtained by weighting and fusing the mean of the multidimensional scores with the hit rate of core concerns. The original experience score is then standardized and linearly aligned with the visibility distribution to obtain the user experience score.

[0047] As a preferred technical solution, a comprehensive objective function is constructed based on visibility score and user experience score, expressed as:

[0048] ;

[0049] in, ,and , This represents the change in visibility score. This indicates the change in user experience score;

[0050] Candidate rewrite versions that satisfy both visibility and user experience objectives are selected. These two objectives are specifically expressed as follows:

[0051] ;

[0052] ;

[0053] ;

[0054] in, Raise the threshold to minimize visibility. To allow for a decrease in the tolerance threshold for user experience, Hitting the threshold for core concerns;

[0055] Among the candidate rewrite versions that satisfy the constraints, select the one that makes... The largest rewrite version is used as the target rewrite content.

[0056] As a preferred technical solution, candidate rewrite versions that do not meet the dual-objective constraint are rewritten compensatorily, specifically including:

[0057] When no candidate rewrite version satisfies the dual objective constraints, a missing concern set is constructed based on the missing core concern items. Compensation prompt words are constructed based on the uncovered core concern terms recorded in the missing concern set. The current candidate rewrite version is then rewritten compensatorily based on the large language model, and a new candidate rewrite version is output.

[0058] This invention also provides a content rewriting system that balances e-commerce GEO visibility improvement and user experience, used to implement the above-mentioned content rewriting method that balances e-commerce GEO visibility improvement and user experience, including: a data acquisition module, a candidate rewritten content generation module, a response generation module, a visibility score calculation module, a user experience score calculation module, a comprehensive objective function construction module, a dual-objective decision-making module, a compensatory rewriting module, and a result output module;

[0059] The data acquisition module is used to acquire raw content, target queries, and user feedback corpora.

[0060] The candidate rewritten content generation module is used to rewrite the original content based on a preset set of rewriting actions to generate candidate rewritten content.

[0061] The response generation module is used to construct a candidate set based on the candidate rewritten content and the interference content, and input the target query into the generative search environment to obtain a generative response;

[0062] The visibility score calculation module is used to divide the generative response into multiple segments and calculate the visibility score of the candidate rewritten content in the corresponding generative response.

[0063] The user experience score calculation module is used to calculate the user experience score of the candidate rewritten content;

[0064] The comprehensive objective function construction module is used to construct a comprehensive objective function based on visibility score and user experience score.

[0065] The dual-objective decision module is used to filter candidate rewrite versions that meet the dual-objective constraints of visibility and user experience based on a comprehensive objective function.

[0066] The compensatory rewriting module is used to perform compensatory rewriting on candidate rewrite versions that do not meet the dual objective constraints until a candidate rewrite version that meets the dual objective constraints is obtained.

[0067] The result output module is used to output the candidate rewrite version that satisfies the dual objective constraint as the target rewrite content.

[0068] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0069] (1) This invention transforms the generative search optimization from a single-objective visibility maximization problem into a dual-objective content rewriting problem. It introduces both visibility enhancement and user experience maintenance objectives during the rewriting generation stage, which can effectively reduce the risk of over-optimization in generative search.

[0070] (2) The present invention is based on the technical solution of extracting the core concern set from user feedback corpus and using it for rewriting constraints and compensatory rewriting, so that the information clues that users really rely on can be preserved, supplemented and strengthened in the content rewriting process.

[0071] (3) This invention comprehensively utilizes the AIDA model, the point mutual information word association theory in computational linguistics, and the controllable text generation and generative search optimization (GEO) methodology in the era of large language models. It forms a complete closed-loop process of candidate rewriting generation, generative response acquisition, visibility calculation, user experience calculation, dual-objective screening, and compensatory rewriting. It no longer relies on the ex-post judgment of a single indicator, but can dynamically suppress conditional failures during the rewriting process, and realize content rewriting that takes into account both machine-side visibility and user-side experience. It provides a scalable solution for content optimization problems in the generative e-commerce search environment. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating the content rewriting method of the present invention that balances improved e-commerce GEO visibility and user experience.

[0073] Figure 2 This diagram illustrates the overall visibility and user experience outcome distribution for a generative search engine optimization experiment. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0075] Example 1

[0076] like Figure 1 As shown, this embodiment provides a content rewriting method that balances improved visibility in e-commerce GEOs and enhanced user experience. It is used in an e-commerce generative search environment to perform constrained rewriting of original content, ensuring that the rewritten content, while increasing the probability of citation, citation position, and textual contribution strength in the generative response, maintains or enhances the user's perception of the content's relevance, credibility, and decision support. Specifically, it includes the following steps:

[0077] S1: Obtain the original content to be rewritten, the target query, and user feedback corpus;

[0078] In this embodiment, the original content to be rewritten is a content object that can be retrieved and referenced by a generative search engine, such as the body of a webpage to be optimized. The target query adopts a user query statement corresponding to the topic of the webpage body. The user feedback corpus is a collection of text information that can reflect user needs and concerns, such as comments, Q&A, customer service records, forum discussions, feedback logs, or other text information related to the topic or application scenario of the original content.

[0079] S2: Based on the preset set of rewriting actions, rewrite the original content to be rewritten and generate candidate rewriting content;

[0080] Let the set of rewrite actions be Regarding the original content After applying different rewrite actions, candidate rewrite content is obtained. ;

[0081] In this embodiment, the rewriting action can adopt a multi-strategy rewriting action based on controllable text generation. By imposing differentiated constraints on the original content, diverse candidate rewritten versions are generated. The rewriting action includes at least one of the following: simple rewriting, authoritative rewriting, statistical enhancement rewriting, keyword enhancement rewriting, source enhancement rewriting, citation enhancement rewriting, easy-to-understand rewriting, fluent rewriting, unique vocabulary enhancement rewriting, and technical terminology enhancement rewriting. Under the given input content and expected attributes, candidate rewritten versions that maintain fluency and meet specific constraints are generated. Each candidate rewritten version is output in a unified structured format, including at least a title and a body text, to improve the comparability and parsability of candidate rewritten content among different methods.

[0082] S3: Input the candidate rewritten content and the interference content into the target generative search environment, such as ChatGPT, Doubao and other large language models, to obtain the generative response;

[0083] Let the original content to be rewritten be denoted as . The remaining interfering content that participates in the generative response construction along with the original content to be rewritten is denoted as... Rewrite the candidate content Together with the interfering content, they are organized into a candidate set. and the target query Inputting both into a generative search environment yields a generative response. ,in, Indicates the randomness of context organization. This indicates that the model performs randomness. and It is only used for theoretical modeling to characterize the inherent random factors in generative models, and there is no need to explicitly obtain their specific values ​​in actual implementation.

[0084] In this embodiment, the generative response is output as an ordered list with source numbers, and each response entry references only one source number, thus making the subsequent source contribution calculation more stable. Preferably, the candidate set size is 5, that is, 1 target candidate content is combined with 4 interference contents;

[0085] Specifically, enter the target query Then, the generative search engine first rewrites the content based on each candidate query and the target query. The model calculates the fit score of each candidate rewritten content based on the content relevance, information completeness and model preference, and sorts the top n candidate rewritten content from high to low fit. Then, the model generates a natural language response (i.e., a generative response) based on the sorted candidate rewritten content set. The response is usually presented in the form of an ordered list, with each list item corresponding to a summary or key information of a candidate rewritten content. The number of the cited candidate rewritten content in the original candidate set is marked with square brackets at the end of each list item (such as [1], [2] etc.) to clarify the source of information.

[0086] To eliminate the impact of randomness on the evaluation results, this embodiment independently generates the same candidate rewritten version multiple times under the same query conditions (e.g., using multiple random seeds), and takes the expected value of the evaluation index. Preferably, in implementation, to ensure the comparability between different candidate rewritten versions, a candidate block containing the target content and several interfering contents is constructed, and an ordered response list with source citation numbers is generated under the same query conditions.

[0087] S4: Divide the generative response into multiple segments and calculate the visibility score of the candidate rewritten content;

[0088] Generative response Decomposed into An ordered segment Extract the source ID, word count, and position of each segment. Let the segment length (word count) be... The source of its attribution is Given position decay weight (follow) (Increases and decreases), then the candidate rewrite content The visibility contribution in this response is .

[0089] Preferably, the position attenuation weight is taken as follows: Considering the nondeterministic nature of the output of a generative system, responses are repeatedly generated for the same rewrite action under multiple random seeds. Candidate rewrite content is defined within the rewrite action. The expected visibility score is , Expressing expectations, This refers to calculating the expected visibility score of repeatedly generated responses under multiple random seeds, taking into account the randomness of context organization and model execution. Specifically, the candidate rewritten content undergoes multiple rewriting actions 'a', multiple PAWC calculations are performed, and the average is taken. The unrewritten content is then considered. To use the baseline as a reference, rewrite the action. The corresponding visibility gain is defined as ;

[0090] S5: Calculate the user experience score of the candidate rewritten content;

[0091] In this embodiment, the user experience score consists of two parts: a multi-dimensional subjective impression score and a core concern hit rate. First, based on the obtained generative response... Extract the candidate rewrite content from it. All sentence fragments are input into the Large Language Model (MLM) as source-level user perception units. The MLM then applies LLM to score the user perception effect of the corresponding source content in the generative response across seven dimensions: relevance, influence, uniqueness, subjective position, subjective count, click probability, and diversity. Each dimension is scored by the MLM with an integer score from 0 to 5, with higher scores indicating better performance in that dimension. It's important to note that the seven-dimensional subjective scores do not directly evaluate the static quality of the candidate content, but rather its user perception effect when presented in the generative response. Therefore, they must rely on the output of the generative response, reflecting the response-dependent nature of user experience scoring. This is homologous and isomorphic to the visibility score (PAWC), facilitating subsequent standardization and alignment. Let the th... The ratings for each dimension are as follows: Its value ranges from 0 to 5, then the mean of the multidimensional score of the large model is defined as... .

[0092] Secondly, a core concern set is constructed: each text in the user feedback corpus D is segmented according to sentence boundaries, using Chinese and English periods, question marks, and exclamation marks as delimiters, dividing each text into an independent sentence sequence. Then, the following preprocessing is performed on each sentence: all letters are converted to lowercase, and letter sequences of at least 3 letters are extracted using regular expressions as initial words, removing numbers, punctuation, and other non-alphabetic characters. Based on a pre-defined stop word list, the initial words in each sentence are filtered. This stop word list includes general function words (such as "the", "and", "for", "very") and generalized words that contribute less to the expression of core concerns (such as "good", "great", "product", "buy"). The remaining words after filtering constitute the candidate word set for that sentence. Simultaneously, duplicate candidate words within each sentence are deduplicated to avoid bias in subsequent estimations due to repeated statistics within the same sentence.

[0093] Furthermore, two types of candidate units are defined: the first type is a single candidate word; the second type is a bigram consisting of two consecutive candidate words. Two basic metrics are calculated for each candidate unit: first, the number of distinct sentences in which the candidate unit appears in the entire corpus, denoted as the number of sentences in which it appears. Second, the number of sentences in which the candidate unit appears in the same sentence as any word in the predefined positive seed word set is denoted as the number of sentences co-occurring with the positive seed word. The positive seed word set was manually compiled from commonly used positive evaluation vocabulary, such as "durable," "comfortable," "effective," "hydrating," and "gentle." These words reflect users' positive attention to product attributes and are used to guide the extraction of user concerns expressed in positive contexts. It should be noted that... The statistics only consider sentences that co-occur with positive seed words, and do not include sentences that co-occur with negative or neutral expressions. The purpose is to screen out concerns that can be used in positive expressions, rather than negative evaluations or risk warning words.

[0094] Preferably, candidate units must meet the following retention threshold: for a single candidate word, the requirement is... ≥3; For bigrams, the requirement is... ≥4. Simultaneously, all candidate units must satisfy... ≥1.

[0095] For each candidate unit that meets the threshold condition, calculate the comprehensive metric S, which is expressed as:

[0096] ;

[0097] in, Indicates the proportion of positive correlation; The logarithmic transformation of the number of sentences is used to characterize the coverage breadth of candidate units in the corpus; It is a statistic used in natural language processing to measure the strength of the association between two words. It measures the statistical association strength between a candidate unit and a positive seed word, that is, whether the probability of the candidate unit and the positive seed word appearing together in the same sentence is significantly higher than the probability of them appearing independently. The larger the value, the more likely it is to represent a semantically related combination.

[0098] In this embodiment, let the total number of sentences in the corpus be T. The probability estimate is the total number of sentences in the entire corpus that contain at least one positive seed word, using a plus-one smoothing method. , , The specific calculation method for the Point Mutual Information PMI is as follows:

[0099] ;

[0100] A higher PMI value indicates that candidate units are more likely to appear in contexts containing positive seed words, and are therefore more likely to represent the dimensions of concern expressed by users in a positive way.

[0101] In this embodiment, a higher comprehensive metric value S indicates that the candidate unit is more representative of the user's true concerns. All candidate units are sorted from highest to lowest according to their comprehensive metric value S, and the top-ranked units are selected as core concerns. Statistics on each core concern item The number of sentences in which the word appears And calculate the weights according to the following formula:

[0102] ;

[0103] This yields a weighted set of core concerns. This core concern set is used to constrain the subsequent rewriting process and avoid losing the information clues that users really rely on during the rewriting process;

[0104] Based on the set of core concerns Calculate the core concern hit rate of candidate rewritten content, assuming the candidate rewritten content is... For core concerns The hit indicator function is The core concern weighted hit rate is defined as:

[0105] ;

[0106] The original experience score is obtained by fusing the average multidimensional scores of the large model with the hit rate of core concerns:

[0107] ;

[0108] in, To pre-determine weights, the original experience scores are standardized and then linearly aligned with the visibility distribution to obtain the user experience score:

[0109] ;

[0110] in, and These are the mean and standard deviation of the original experience scores, respectively. and These are the mean and standard deviation of the visibility distribution, respectively, in unmodified form. To use the baseline as a reference, rewrite the action. The corresponding user experience change is defined as: .

[0111] S6: Implement a dual-objective constraint-based content rewriting strategy;

[0112] For each candidate rewrite action Calculate the comprehensive objective function:

[0113] ;

[0114] in, ,and Based on this, candidate rewrite versions that meet the following constraints are selected:

[0115] ;

[0116] ;

[0117] ;

[0118] in, Raise the threshold to minimize visibility. To allow for a decrease in the tolerance threshold for user experience, To determine the core concern threshold, among the candidate rewrite versions that meet the constraints, select the one that... The largest rewritten version is taken as the target rewritten content, that is ;

[0119] S7: Perform a compensatory rewrite;

[0120] When no candidate rewrite version meets the constraints, identify the missing high-weight core concerns and construct a set of missing concerns:

[0121] ;

[0122] in, The threshold for determining high-weight concerns is t, where t represents the t-th round of rewriting. Indicates candidate rewrite content Rewrite using the t-th round of rewriting action a;

[0123] Perform a compensatory rewrite on the corresponding candidate rewrite version to obtain:

[0124] ;

[0125] For example, when a candidate rewrite version removes two high-weight concerns—applicable scenarios and limitations—while improving visibility, a compensatory rewrite will explicitly require the restoration of applicable scenario descriptions, usage limitation descriptions, and avoidance of single selling point expressions in the prompt.

[0126] In this embodiment, compensatory rewriting includes at least one of the following: restoring deleted or weakened high-frequency core concern terms (such as outdoor use, waterproof, compatibility, etc.), supplementing omitted comparison dimension terms (such as battery life comparison, material differences, etc.), and replacing overly templated or low-information generalized expressions (such as expanding high quality to specific attribute terms).

[0127] In practice, it is based on the missing concern set. The system constructs compensating prompts from the uncovered core interest terms recorded in the database, calls the large language model to selectively supplement or adjust the current candidate rewrite version, and outputs the new version. ;

[0128] In this embodiment, the example of the compensation prompt is: "The current copy is missing the following keywords that users often mention: {list of missing words}. Please integrate these keywords or related expressions naturally while maintaining the original style. Do not fabricate facts."

[0129] Repeat the above steps of generating response, visibility calculation, user experience calculation, and dual-objective filtering on the content after compensatory rewriting until the target rewritten content that meets the constraints is obtained or the preset iteration limit is reached.

[0130] Through the above steps, this embodiment improves the probability of original content being selected and cited in the generative search environment without sacrificing user experience, and avoids conditional failures caused by excessive templates, excessive compression, and emphasis on a single selling point.

[0131] like Figure 2As shown, a large-scale e-commerce product generative search engine optimization experiment revealed that, across 26 categories and 10 optimization methods, only about 24.1% of the rewrites achieved simultaneous improvements in visibility and user experience, while about 18.8% of the rewrites exhibited over-optimization, resulting in improved visibility but decreased user experience. The latter is often accompanied by negative user evaluations of the content's credibility and decision support.

[0132] To verify the technical effectiveness of this invention in balancing visibility improvement and user experience in e-commerce generative search, 200 product text samples from four categories—cosmetics, 3C digital products, home furnishings, and apparel and footwear—were used as experimental subjects in an e-commerce dataset. As shown in Table 1, the core effects of the baseline strategy of directly maximizing visibility (unconstrained) and the dual-objective filtering and compensation rewriting strategy of this invention were compared. The baseline strategy only aims to maximize visibility, and its win-win ratio (the percentage of samples that simultaneously achieve visibility improvement ΔPAWC > 0 and user experience improvement ΔS1 > 0) is only 24.1%, which is excessive. The optimization rate (the percentage of samples where visibility is improved but user experience is reduced by ΔS1<0) reached 18.8%. However, the strategy of this invention, through the technical means of first screening potential qualified solutions with ΔPAWC>0 and ΔS1≥0, and then compensating and rewriting samples with ΔPAWC>0 but ΔS1<0 (such as adjusting keyword density and supplementing fluent sentences), ultimately achieved a win-win rate of 58.3% and an over-opt rate of 7.5%. This significantly proves that the present invention can effectively balance the visibility and user experience of e-commerce generative search and avoid the problem of over-optimization.

[0133] Table 1. Comparison of Core Indicators for Dual Objectives

[0134]

[0135] Example 2

[0136] This embodiment provides a content rewriting system that balances e-commerce GEO visibility improvement and user experience, used to implement the content rewriting method of embodiment 1 that balances e-commerce GEO visibility improvement and user experience, including: a data acquisition module, a candidate rewritten content generation module, a response generation module, a visibility score calculation module, a user experience score calculation module, a comprehensive objective function construction module, a dual-objective decision module, a compensatory rewriting module, and a result output module;

[0137] In this embodiment, the data acquisition module is used to acquire the original content, the target query, and user feedback corpus;

[0138] In this embodiment, the candidate rewritten content generation module is used to rewrite the original content based on a preset set of rewriting actions to generate candidate rewritten content;

[0139] In this embodiment, the response generation module is used to construct a candidate set based on the candidate rewritten content and the interference content, and input the target query into the generative search environment to obtain a generative response;

[0140] In this embodiment, the visibility score calculation module is used to divide the generative response into multiple segments and calculate the visibility score of the candidate rewritten content in the corresponding generative response.

[0141] In this embodiment, the user experience score calculation module is used to calculate the user experience score of the candidate rewritten content;

[0142] In this embodiment, the comprehensive objective function construction module is used to construct a comprehensive objective function based on visibility score and user experience score.

[0143] In this embodiment, the dual-objective decision module is used to perform dual-objective constraint filtering based on a comprehensive objective function, and to filter candidate rewrite versions that meet the dual-objective constraints;

[0144] In this embodiment, the compensatory rewriting module is used to perform compensatory rewriting on candidate rewrite versions that do not meet the dual objective constraints until a candidate rewrite version that meets the dual objective constraints is obtained.

[0145] In this embodiment, the result output module is used to output the candidate rewrite version that satisfies the dual objective constraint as the target rewrite content.

[0146] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A content rewriting method that balances improved e-commerce GEO visibility and user experience, characterized in that, Includes the following steps: Obtain raw content, target queries, and user feedback corpora; The original content is rewritten based on a preset set of rewriting actions to generate candidate rewritten content. A candidate set is constructed based on the candidate rewritten content and the interference content, and then input into the generative search environment with the target query to obtain a generative response; The generative response is segmented into multiple fragments, and the visibility score of the candidate rewritten content in the corresponding generative response is calculated. The user feedback corpus is segmented and filtered to calculate the user experience score of the candidate rewritten content; A comprehensive objective function is constructed based on visibility score and user experience score; Based on the comprehensive objective function, a dual objective constraint of visibility and user experience is used for screening. Candidate rewrite versions that meet the dual objective constraints are used as the target rewrite content, while candidate rewrite versions that do not meet the dual objective constraints are rewritten compensatorily until a candidate rewrite version that meets the dual objective constraints is obtained.

2. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... The pre-defined set of rewriting actions employs a multi-strategy rewriting approach based on controllable text generation, which generates various candidate rewriting contents by imposing differentiated constraints on the original content.

3. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... Generative responses are represented as: ； ； in, Represents a generative response. This indicates a generative search environment. Indicates the target query. Indicates the randomness of context organization. This indicates that the model performs randomness. This represents a candidate set constructed based on candidate rewritten content and interfering content. Indicates the candidate rewrite content. This indicates interfering content.

4. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... A candidate set is constructed based on the candidate rewritten content and the distractor content, and then input into a generative search environment with the target query to obtain a generative response, specifically including: Generative search engines calculate the fit score of each candidate rewritten content based on its relevance to the target query, information completeness, and model preferences. They then rank the top n candidate rewritten contents from highest to lowest fit score and generate a generative response based on the ranked set of candidate rewritten contents.

5. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... Generative responses are output as an ordered list with source numbers, and each response entry references the number of the candidate rewrite in the original candidate set.

6. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... The generative response is segmented into multiple fragments, and the visibility score of the candidate rewrite content in the corresponding generative response is calculated, specifically including: Decompose the generative response into Given a set of ordered segments, extract the source ID, word count, and position of each segment. Assume the segment length is 1. The source of its attribution is The visibility contribution of candidate rewritten content in this response is calculated and expressed as: ； in, Indicates the position decay weight; For the same rewrite action, responses are repeatedly generated under multiple random seeds, and candidate rewrite content is calculated for each rewrite action. The expected visibility score is: ； in, Indicates the randomness of context organization. This indicates that the model performs randomness. This represents the expectation of the visibility score for repeatedly generated responses under multiple random seeds, taking into account the randomness of context organization and model execution.

7. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... Calculate the user experience score for the candidate rewritten content, specifically including: Based on the generative response, extract all sentence fragments of the corresponding candidate rewritten content, output multi-dimensional user perception effect scores through a large language model, and calculate the mean of the multi-dimensional scores; Each text in the user feedback corpus is segmented according to sentence boundaries, and each text is divided into an independent sentence sequence. Based on a preset stop word list, the initial words in each sentence are filtered to build a candidate word set, and the candidate words in each sentence are deduplicated. Two types of candidate units are constructed, including single candidate words and bigrams consisting of two consecutive candidate words; The comprehensive metric is calculated and expressed as follows: ； ； in, This represents a comprehensive metric. Represents point mutual information, Indicates the proportion of positive correlation. To perform a logarithmic transformation on the number of sentences, This indicates the number of sentences appearing in the candidate unit. This represents the number of sentences that co-occur with the positive seed word, and T represents the total number of sentences in the user feedback corpus. The total number of sentences in the entire corpus that contain at least one positive seed word; Candidate units are sorted from highest to lowest based on their comprehensive metrics. Core concerns are selected, and the number of sentences containing each core concern is counted to calculate its weight. : ； in, Indicate core concerns, This indicates the number of sentences containing each core concern item; Calculate the weighted hit rate of core concerns , is represented as: ； in, Indicates candidate rewrite content For core concerns The hit indicator function; The original experience score is obtained by weighting and fusing the mean of the multidimensional scores with the hit rate of core concerns. The original experience score is then standardized and linearly aligned with the visibility distribution to obtain the user experience score.

8. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... A comprehensive objective function is constructed based on visibility score and user experience score, expressed as: ； in, ,and , This represents the change in visibility score. This indicates the change in user experience score; Candidate rewrite versions that satisfy both visibility and user experience objectives are selected. These two objectives are specifically expressed as follows: ； ； ； in, Raise the threshold to minimize visibility. To allow for a decrease in the tolerance threshold for user experience, Hitting the threshold for core concerns; Among the candidate rewrite versions that satisfy the constraints, select the one that makes... The largest rewrite version is used as the target rewrite content.

9. The content rewriting method according to claim 1, which balances improved e-commerce GEO visibility and user experience, is characterized in that... Compensatory rewrites will be performed on candidate rewrite versions that do not meet the dual-objective constraints, specifically including: When no candidate rewrite version satisfies the dual objective constraints, a missing concern set is constructed based on the missing core concern items. Compensation prompt words are constructed based on the uncovered core concern terms recorded in the missing concern set. The current candidate rewrite version is then rewritten compensatorily based on the large language model, and a new candidate rewrite version is output.

10. A content rewriting system that balances improved e-commerce GEO visibility with enhanced user experience, characterized in that: The content rewriting method for achieving both improved e-commerce GEO visibility and user experience as described in any one of claims 1-9 includes: a data acquisition module, a candidate rewritten content generation module, a response generation module, a visibility score calculation module, a user experience score calculation module, a comprehensive objective function construction module, a dual-objective decision-making module, a compensatory rewriting module, and a result output module. The data acquisition module is used to acquire raw content, target queries, and user feedback corpora. The candidate rewritten content generation module is used to rewrite the original content based on a preset set of rewriting actions to generate candidate rewritten content. The response generation module is used to construct a candidate set based on the candidate rewritten content and the interference content, and input the target query into the generative search environment to obtain a generative response; The visibility score calculation module is used to divide the generative response into multiple segments and calculate the visibility score of the candidate rewritten content in the corresponding generative response. The user experience score calculation module is used to calculate the user experience score of the candidate rewritten content; The comprehensive objective function construction module is used to construct a comprehensive objective function based on visibility score and user experience score. The dual-objective decision module is used to filter candidate rewrite versions that meet the dual-objective constraints of visibility and user experience based on a comprehensive objective function. The compensatory rewriting module is used to perform compensatory rewriting on candidate rewrite versions that do not meet the dual objective constraints until a candidate rewrite version that meets the dual objective constraints is obtained. The result output module is used to output the candidate rewrite version that satisfies the dual objective constraint as the target rewrite content.

Images