A brand monitoring and intelligent optimization method and system based on generation engine optimization

Abstract

This application discloses a brand monitoring and intelligent optimization method and system based on generative engine optimization, including: S1: inputting a multimodal detection cue matrix into a generative engine to obtain multi-source data, calculating the cross-modal alignment loss between brand standard assets and multi-source data to locate cognitive bias, tracing citation sources to construct an attribution link graph, and determining toxic slices; S2: analyzing the distribution of toxic slices in the retrieval encoder space, generating multi-source optimized content including structured text, images injected with anti-interference perturbation features, and structured videos, and verifying the content to form a knowledge injection carrier; S3: classifying and distributing the carrier to the platform to preemptively occupy the retrieval ranking to cover the toxic slices, and re-detecting and verifying at intervals until the alignment loss converges. Belonging to the field of generative engine optimization technology, this method can calculate the alignment loss through cross-modal detection to locate toxic slices and generate multimodal knowledge carriers injected with anti-interference features for targeted distribution and coverage.

Description

Technical Field

[0001] This invention relates to the fields of computer information retrieval and natural language processing, and in particular to a generative engine optimization method based on content structuring. Background Technology

[0002] With the popularization of generative engines, users are obtaining brand information in a way that has shifted from traditional webpage ranking and reading to multi-source content summaries generated directly by large models.

[0003] However, existing technologies have the following drawbacks: 1. Existing brand monitoring technologies mostly rely on single-text sentiment analysis or fragmented processing of images and videos. They cannot explain why generative engines produce inconsistencies between text and images or random video combinations. Furthermore, they cannot locate the sources of cross-modal knowledge pollution introduced by generative engines during the retrieval enhancement generation stage, resulting in the challenge of tracing the source of cross-modal misalignment and illusion.

[0004] 2. Existing optimization techniques only modify the text and lack an understanding of the underlying vector space of the large-scale multimodal encoder. This often results in the phenomenon that the text description is correct but the brand image generated by AI (artificial intelligence) is still incorrect after optimization, which cannot achieve collaborative intervention of text, image and vision and causes cross-modal cognitive dissonance.

[0005] 3. Existing optimization intervention methods mostly involve passively modifying webpage content and waiting for the engine to re-crawl, lacking strategies to proactively seize search rankings and failing to establish a periodic verification loop, making it difficult to control the optimization effect and easily decaying with model updates.

[0006] Therefore, there is an urgent need to develop a brand monitoring and intelligent optimization method and system based on generation engine optimization to solve the problems in the existing technology. Summary of the Invention

[0007] The purpose of this invention is to provide a brand monitoring and intelligent optimization method and system based on generative engine optimization, which aims to solve the problems of tracing the source of cross-modal misalignment and illusion in large models, the problem of visual cognitive dissonance caused by traditional single-modal optimization, and the problems of uncontrollable and easily decayed optimization effects of existing technologies. It can locate toxic slices by calculating alignment loss through cross-modal detection, and generate multimodal knowledge carriers injected with anti-perturbation features for targeted distribution and coverage, thereby achieving the technical effect of accurately tracing the source of pollution and realizing the holographic cognitive reshaping of brands. Moreover, it has a simple structure and is easy to use, thus solving the problems of tracing, coordination and control and effect decay mentioned in the background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A brand monitoring and intelligent optimization method based on generation engine optimization includes the following steps: S1: Based on multi-source data acquisition and attribution link analysis of cross-modal detection, the multi-modal detection prompt word matrix is ​​input into the generative engine to obtain multi-source data content. The cross-modal alignment loss between brand standard assets and multi-source data content is calculated using a multi-modal alignment and evaluation model to locate cognitive biases. The source of citation is tracked to construct an attribution link map and determine toxic knowledge slices. S2: Generate adversarial knowledge injection carrier based on cross-modal semantic alignment. Analyze the distribution of cross-modal toxic knowledge slices in the encoding space of the multimodal retrieval engine of the generation engine, generate structured text, and inject images with implicit perturbation features and semantic tags to resist generation interference, as well as videos of key frames and structured text, to form multi-source optimized content. After passing the multimodal alignment and evaluation model verification, it constitutes a multimodal knowledge injection carrier. S3: Multi-source targeted distribution and holographic closed-loop verification. Multimodal knowledge is injected into the carrier and classified and delivered to the corresponding platform to seize the search ranking in order to cover cross-modal toxicity knowledge slices. After a preset period of time, the verification is re-probeed until the brand cross-modal alignment loss converges.

[0009] By adopting the above technical solutions, the overall function of discovering cross-modal cognitive biases of brands and tracing toxicity slices through multimodal detection, generating adversarial multimodal carriers to cover toxicity slices, and finally completing brand cognition reshaping and closed-loop verification through targeted distribution can be achieved.

[0010] As a further aspect of the present invention: S1 utilizes a multimodal alignment and evaluation model to calculate the cross-modal alignment loss between brand standard assets and multi-source data content to locate cognitive biases, including the following sub-steps: S11: Extract visual features of images, semantic features of text, and temporal keyframe features of videos; S12: Obtain cross-modal alignment loss by calculating the cosine distance between the brand standard asset feature vector and the multi-source data content feature vector; S13: When the cross-modal alignment loss exceeds a preset threshold, it is determined that there is a cross-modal cognitive bias.

[0011] By adopting the above technical solutions, the system can achieve feature alignment and deviation quantification calculation of brand standard assets and AI-output multi-source content in the latent space, thereby accurately identifying cross-modal cognitive biases.

[0012] As a further aspect of the present invention: when the cross-modal alignment loss exceeds a preset threshold, it is determined that there is a cross-modal cognitive bias, and the preset threshold is set to 0.35.

[0013] By adopting the above technical solution, the function of accurately defining the boundary of cognitive bias and improving the accuracy of cross-modal cognitive bias judgment can be achieved by setting specific deviation quantification thresholds.

[0014] As a further aspect of the present invention: the multimodal alignment and evaluation model is constructed based on a contrastive language image pre-training architecture. It collects image-text-visual triplet data from the brand domain as training, validation, and test sets, and trains it using a contrastive learning loss function to narrow the distance between matching image-text-visual features and widen the distance between non-matching features.

[0015] By adopting the above technical solution, a model with cross-modal feature alignment evaluation capability can be constructed, enabling it to accurately measure the matching degree of image and text visual features.

[0016] As a further aspect of the present invention: the contrastive learning loss function is used for training, and the AdamW optimizer (an adaptive moment estimation optimizer) is used. The initial learning rate is set to 5e-5, and the number of training rounds is set to 30. When the loss function value on the validation set no longer decreases for 3 consecutive rounds, the training is stopped early and the model weights are saved.

[0017] By adopting the above technical solutions, the training process of the multimodal alignment and evaluation model can be optimized, ensuring that the model converges to the optimal state and improving the accuracy and stability of feature alignment.

[0018] As a further aspect of the present invention: the injection of anti-generated interference stealth perturbation features in S2 includes the following: with the original brand image, brand standard text and perturbation budget upper limit as constraints, the applied adversarial perturbation is calculated by maximizing the similarity between the image features after perturbation and the brand standard text features, and the adversarial perturbation is superimposed on the pixel layer of the original brand image.

[0019] By adopting the above technical solution, the ability to bind brand images to the brand standard text semantics and resist interference in the generation engine coding space is enhanced by calculating and injecting adversarial perturbations.

[0020] As a further aspect of the present invention: the multimodal alignment and evaluation model verification in S2 includes: verifying the cosine similarity between each pair of text features, image features after injection perturbation, and video keyframe features in the latent space. When the similarity is greater than 0.9, it is determined that they constitute a high-weight cross-modal entity.

[0021] By adopting the above technical solution, the function of strictly verifying the internal consistency of multimodal carrier features can be achieved to ensure that the generated carrier forms a high-weight indivisible entity in the latent space, thereby improving the retrieval hit rate.

[0022] As a further aspect of the present invention: the multi-source optimized content of the video containing keyframes and structured text in S2 includes: extracting standard keyframes of the brand video, generating a subtitle file containing accurate timestamps and brand action descriptions, and structured video transcription text.

[0023] By adopting the above technical solutions, a structured description that is temporally and semantically deeply bound to video content can be constructed, thereby enhancing the semantic density and parsability of video modalities in the retrieval of the generation engine.

[0024] As a further aspect of the present invention: S3 involves classifying and delivering multimodal knowledge injection carriers to corresponding platforms, including: delivering text to a high-weight knowledge base, delivering brand images with metadata to mainstream image libraries and official website media pools, and delivering structured videos to video platforms with high crawling frequency.

[0025] By adopting the above technical solution, the function of accurately matching high-weight distribution nodes according to the attributes of multimodal carriers can be realized, thereby maximizing the carrier being indexed by the generation engine and seizing the retrieval and ranking efficiency.

[0026] This invention also discloses a brand monitoring and intelligent optimization system based on generation engine optimization, used to implement the aforementioned brand monitoring and intelligent optimization method based on generation engine optimization, comprising: The cross-modal detection and attribution analysis module is used to execute S1. Based on the multi-source data acquisition and attribution link analysis of cross-modal detection, the multi-modal detection prompt word matrix is ​​input into the generative engine to obtain multi-source data content. The cross-modal alignment loss between brand standard assets and multi-source data content is calculated using the multi-modal alignment and evaluation model to locate cognitive bias. The citation source is tracked to construct an attribution link map and determine toxic knowledge slices. The adversarial knowledge injection carrier generation module is used to execute S2, generate adversarial knowledge injection carriers based on cross-modal semantic alignment, analyze the distribution of cross-modal toxic knowledge slices in the encoding space of the multimodal retrieval engine, generate images containing structured text, injected implicit perturbation features to resist generation interference and semantic tags, as well as multi-source optimized content of video with key frames and structured text, and constitute a multimodal knowledge injection carrier after passing the multimodal alignment and evaluation model verification. The targeted distribution and closed-loop verification module is used to execute S3, multi-source targeted distribution and holographic closed-loop verification, which classifies and delivers multimodal knowledge injection carriers to corresponding platforms to preemptively retrieve and sort information to cover cross-modal toxicity knowledge slices, and re-probes and verifies after a preset interval until the brand cross-modal alignment loss converges.

[0027] By adopting the above technical solutions, a system corresponding to the multimodal generative engine optimization (GEO) brand monitoring and intelligent optimization method is provided, ensuring efficient flow of each step and closed-loop execution of the entire process.

[0028] Compared with the prior art, the beneficial effects of the present invention are: This invention constructs a multi-source data joint detection and attribution link to achieve accurate cross-modal attribution and toxicity localization, calculates cross-modal alignment loss, and accurately identifies specific knowledge pollution sources that cause brand visual distortion or semantic shift, thus solving the problem of tracing the source of cross-modal misalignment and illusion in large models. This invention achieves multimodal holographic cognitive reshaping through cross-modal semantic alignment and attention adversarial mechanisms, generating a multimodal knowledge injection carrier containing high semantic density text and anti-interference visual features, actively covering toxic slices, and solving the problem of image-text visual cognitive dissonance caused by traditional single-modal optimization; This invention achieves targeted intervention and dynamic closed-loop verification by classifying and deploying high-weight multimodal carriers to preemptively occupy retrieval rankings, and periodically re-tests and verifies until alignment loss converges, ensuring that positive brand information is stably referenced in the generation engine, thus solving the problems of uncontrollable and easily decaying optimization effects of existing technologies.

[0029] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0030] Figure 1 This is a flowchart illustrating the steps of a brand monitoring and intelligent optimization method based on generation engine optimization in an embodiment of the present invention. Detailed Implementation

[0031] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0032] 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.

[0033] Example 1 In this embodiment of the invention, a brand monitoring and intelligent optimization method and system based on generation engine optimization is described below. Figure 1 As shown, to address the problems of cross-modal misalignment, difficulty in tracing the source of illusions, and cognitive dissonance caused by single-modal optimization in generative engines, this invention acquires multi-source data through cross-modal detection and calculates alignment loss, and reverse-engineers a four-level attribution link graph to accurately locate cross-modal toxic knowledge slices; then, it analyzes the vector distribution of toxic slices to generate multi-modal knowledge injection carriers containing structured text, images injected with adversarial perturbations, and structured videos, which, after internal consistency verification, constitute high-weight cross-modal entities; finally, it preemptively seizes retrieval ranking through multi-source targeted distribution and establishes a closed-loop mechanism of periodic re-exploration verification to achieve holographic cognitive reshaping and stable convergence of brands in generative engines.

[0034] Before execution, a multimodal alignment and evaluation model needs to be trained in advance. The specific process is as follows: A total of 500,000 sets of image, text, and visual triplet data were collected in the brand field. Visual features were extracted from images, semantic features from text, and temporal keyframe features from videos.

[0035] The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set includes 350,000 sets for model training, the validation set includes 100,000 sets for adjusting model parameters, and the test set includes 50,000 sets for evaluating the final performance of the model.

[0036] The multimodal alignment and evaluation model is built on the contrastive language image pre-training (CLIP) architecture, which includes an image encoder, ResNet-50 (deep residual network-50), which outputs 2048-dimensional visual features; a text encoder, BERT-base (a basic version of the bidirectional encoder representation based on transformers), which outputs 768-dimensional semantic features; and a video encoder, I3D, which outputs 1024-dimensional temporal keyframe features.

[0037] Specifically, it includes: Data preprocessing: Image: resized to 224×224, normalized (mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225]); Text: Tokenized using BERT, padding to the maximum length of 128; Video: Keyframes were extracted using FFmpeg (an audio and video processing framework) at a rate of 1 frame per second with a -r parameter of 1. Features were extracted from each keyframe using ResNet-50 and average pooling was performed to obtain the video features.

[0038] The training objective uses the InfoNCE noise contrastive estimation information maximization loss contrastive learning loss function, the formula of which is:

[0039] in, For image / video feature vectors, To match text feature vectors, For other text feature vectors (negative samples) in the same batch. (Temperature parameter) (Batch size).

[0040] Optimizer and training parameters: The AdamW optimizer (an adaptive moment estimation optimizer) is used, with an initial learning rate of 5e-5 (i.e., The weights decay by 0.01, and the training rounds are 30. When the validation set loss does not decrease for 3 consecutive rounds, training is stopped early and the model weights are saved.

[0041] After the model training is completed, perform the method steps of this invention: S1: Multi-source data acquisition and attribution link analysis based on cross-modal detection Specifically, it includes the following sub-steps: S11: Extract visual features of images, semantic features of text, and temporal keyframe features of videos.

[0042] The system constructs a multimodal detection prompt matrix, which includes plain text query commands, query commands with brand reference images, and query commands that require the generation of video descriptions.

[0043] The matrix is ​​input into the target generative engine to obtain the multi-source data content output by the engine, specifically including text paragraphs, generated images, and video keyframe sequences; the encoder of the trained multimodal alignment and evaluation model is used to extract the feature vectors of the above multi-source content respectively.

[0044] S12: Obtain cross-modal alignment loss by calculating the cosine distance between the brand standard asset feature vector and the multi-source data content feature vector.

[0045] The brand's standard asset feature vector is pre-coded from the brand's official standard logo, standard slogan, and official video.

[0046] The brand's standard asset feature vector is pre-encoded from the brand's official standard logo, standard slogan, and official video. The specific generation method is as follows: The standard logo image was extracted using a ResNet-50 encoder to obtain 2048-dimensional visual features, the standard promotional text was extracted using a BERT-base encoder to obtain 768-dimensional semantic features, and the official video was extracted using an I3D model to obtain temporal keyframe features and average pooling to obtain 1024-dimensional video features. The features of each modality were mapped to a common feature space of 2048-dimensional through a linear projection layer, and then fused into a single brand standard asset feature vector by weighted averaging, with image weight 0.4, text weight 0.4, and video weight 0.2.

[0047] The specific calculation process is as follows: First, the encoder of the multimodal alignment and evaluation model is used to extract feature vectors of the brand's standard assets (such as standard logo images, standard promotional text, and keyframes of official videos). Feature vectors of multi-source data content (such as text, images, and video keyframes output by generative engines). Next, calculate the cosine similarity between the two feature vectors, using the following formula:

[0048] in, Represents the vector dot product. Let L2 norm represent the vectors; finally, convert cosine similarity to cosine distance using the formula:

[0049] This distance value is the cross-modal alignment loss, which measures the degree of difference between brand standard assets and multi-source data content in the feature space. The larger the distance value, the greater the difference.

[0050] S13: When the cross-modal alignment loss exceeds a preset threshold, it is determined that there is a cross-modal cognitive bias.

[0051] In this embodiment, the preset threshold is set to 0.35; if the calculated cosine distance is greater than 0.35, it is determined that the generative engine outputs brand content with deviation.

[0052] After determining that there is a deviation, the source address of the reference attached to the output content of the generative engine is traced, and text fragments, original image regions and video slices in the original web page are crawled.

[0053] A four-level attribution link graph is constructed, from multimodal cue words to generated expressions and then to cross-modal source slices; source slices that cause cross-modal alignment loss exceeding 0.35 are identified as cross-modal toxic knowledge slices.

[0054] S2: Generation of Adversarial Knowledge Injection Carriers Based on Cross-Modal Semantic Alignment Specifically, it includes the following sub-steps: S21: Analyze the distribution of cross-modal toxicity knowledge slices in the encoding space of the multimodal retrieval engine to generate multi-source optimized content.

[0055] For structured text: Generate high semantic density structured text containing triples of core brand attributes.

[0056] For images: Generate images containing injected latent perturbation features to resist generation interference and semantic labels; the logic for injecting latent perturbation features to resist generation interference is as follows: using the original brand image I and the brand standard text... Disruption budget cap To constrain the perturbation, the projected gradient descent (PGD) algorithm is used to calculate the anti-perturbation algorithm. The formula is:

[0057] in, For image encoder (ResNet-50). For text encoders (BERT-base).

[0058] The specific steps of PGD include: initialization Learning rate The number of iterations K=40; For k=1 to K: Calculate the gradient ,renew ; Projection constraints: This ensures that the visual quality of the image does not change significantly after the perturbation.

[0059] Will Overlay the image onto the original image I and embed semantic tags in the metadata, such as “A car front face streamlined wing grille”.

[0060] For videos: Generate videos containing keyframes and structured text; extract standard keyframes from brand videos to generate caption files containing accurate timestamps and descriptions of brand actions, as well as structured video transcription text.

[0061] S22: Input the multi-source optimized content into the multimodal alignment and evaluation model. When the similarity between text features, image features and video features in the latent space is greater than the preset threshold, the verification is deemed successful.

[0062] Verify text features t and image features after perturbation. Video keyframe features Pairwise cosine similarity in the latent space:

[0063] When all conditions are met, it is determined to be a high-weight cross-modal entity, constituting a multimodal knowledge injection carrier.

[0064] S3: Multi-source targeted distribution and holographic closed-loop verification Specifically, it includes the following sub-steps: S31: Inject multimodal knowledge into carriers and classify and deploy them to corresponding platforms. Use cross-modal feature consistency to preemptively retrieval and ranking to cover cross-modal toxicity knowledge slices.

[0065] The specific delivery strategy is as follows: deliver text to high-weight knowledge bases, deliver brand images with metadata to mainstream image libraries and official website media pools, and deliver structured videos to video platforms with high crawling frequency.

[0066] When the multimodal retrieval engine of the generation engine processes related queries in the future, it will give the carrier a search relevance score that is far higher than that of single-modal content because the carrier has extremely high cross-modal feature consistency. This will allow it to take over the ranking position, causing the original toxic slices to be downgraded or ignored.

[0067] S32: Re-probe and verify after a preset interval until the brand cross-modal alignment loss converges.

[0068] After a preset interval, the detection and loss calculation in step S11 are re-executed. When the brand cross-modal alignment loss is less than or equal to 0.35, the brand holographic cognition reshaping is determined to be complete and the optimization ends. If it does not converge, the adversarial perturbation budget or text semantic density is adjusted to regenerate the carrier and launch it.

[0069] A brand monitoring and intelligent optimization system based on a generation engine optimization includes: The cross-modal detection and attribution analysis module is used to perform multi-source data acquisition and attribution link analysis based on cross-modal detection. It inputs the multi-modal detection prompt word matrix into the generative engine to obtain multi-source data content, uses the multi-modal alignment and evaluation model to calculate the cross-modal alignment loss between brand standard assets and multi-source data content to locate cognitive biases, tracks the source of citations to construct an attribution link map and determine toxic knowledge slices. The adversarial knowledge injection carrier generation module is used to perform the adversarial knowledge injection carrier generation step based on cross-modal semantic alignment. It analyzes the distribution of cross-modal toxic knowledge slices in the encoding space of the multimodal retrieval engine of the generation engine, generates images containing structured text, injected implicit perturbation features to resist generation interference and semantic tags, as well as multi-source optimized content of video with key frames and structured text. After passing the multimodal alignment and evaluation model verification, it constitutes the multimodal knowledge injection carrier. The targeted distribution and closed-loop verification module is used to perform multi-source targeted distribution and holographic closed-loop verification steps. It injects multimodal knowledge into carriers, classifies and delivers them to corresponding platforms to preemptively search and rank them to cover cross-modal toxicity knowledge slices. After a preset period of time, it re-probes and verifies until the brand's cross-modal alignment loss converges.

[0070] Example 2 The technical difference between this embodiment and Embodiment 1 lies in the following: taking the GEO monitoring and optimization of a smart car brand A as an example, the implementation process of this invention is explained in detail: Step 1: Model Preparation The multimodal alignment and evaluation model was trained using the aforementioned method, with the AdamW optimizer used for 30 training epochs. The initial learning rate was set to 5e-5, and the model weights were saved when the validation set loss did not decrease for three consecutive epochs.

[0071] Step 2: Cross-modal detection and attribution analysis The system constructs a multimodal detection prompt matrix for car A, including plain text query commands such as "Please describe the appearance and performance of car A", query commands with reference images such as "Please describe the front design of car A with reference images", and video generation commands such as "Please generate a video description of car A driving on snow".

[0072] The above instructions are input into the generative engine to obtain the output text paragraphs, generated images and video keyframe sequences, and the model is used to extract features and calculate cosine distance.

[0073] Calculations revealed that the text description "Car A's front face adopts a vertical waterfall grille" has a significant semantic distance from the brand's standard asset "Car A's front face adopts a streamlined wing grille," and the generated image exhibits the visual characteristics of a vertical waterfall grille, with a cosine distance of 0.45 from the standard logo features.

[0074] Since 0.45 is greater than the preset threshold of 0.35, the system determines that there is cross-modal cognitive bias.

[0075] The system traced the source of the reference and found that the engine referenced a post fragment and images from a car forum when outputting the content. The system constructed an attribution link graph and identified the text and image slices from the forum post as cross-modal toxicity knowledge slices.

[0076] Step 3: Generation of adversarial knowledge injection carrier The system analyzes the distribution of the toxic slice in the retrieval code space and generates multi-source optimized content: Structured text generation: Generates high semantic density structured text containing triples, including: A. car, front face design, streamlined wing grille.

[0077] Image generation: Select the official flying wing grille standard image and set the upper limit of the disturbance budget. The value is 8 / 255, calculated using the anti-disturbance formula. This maximizes the similarity between the perturbed image and the text feature of "streamlined flying wing grille" in the coding space.

[0078] Will The image is overlaid onto the original pixel layer, and the semantic tag "A car front face streamlined wing grille" is written into the image metadata.

[0079] Video generation: Extract standard keyframes from official promotional videos to generate SubRip (SRT) subtitle files with precise timestamps and structured transcribed text "00:00:05 A close-up of a car's streamlined wing grille".

[0080] The above content was input into the model for verification. The pairwise cosine similarity of text features, perturbed image features, and video keyframe features all reached 0.92, which is greater than the threshold of 0.9. Therefore, it was determined that it constituted a high-weight cross-modal entity and a multimodal knowledge injection carrier was generated.

[0081] Step 4: Multi-source targeted distribution and holographic closed-loop verification The system delivers structured text to a high-weight knowledge base, perturbation images with metadata to the official website's media pool, and structured videos with SRT subtitles to video platforms with high crawling frequency.

[0082] Seven days later, the system re-executed the detection command, and the text output by the generative engine changed to "Car A's front face adopts a streamlined wing grille". The generated image showed the wing grille feature, and the cosine distance between it and the standard logo feature decreased to 0.25.

[0083] Since 0.25 is less than the preset threshold of 0.35, the system determines that the brand cross-modal alignment loss has converged, and the optimization is complete.

[0084] This invention provides a brand monitoring and intelligent optimization method and system based on generative engine optimization. It can locate toxic slices by calculating alignment loss through cross-modal detection, and generate multimodal knowledge carriers with anti-perturbation features for targeted distribution and coverage, thereby achieving the technical effect of accurately tracing the source of pollution and realizing the reshaping of brand holographic cognition.

[0085] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0086] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the claims be included within the invention.

[0087] No reference numerals in the claims should be construed as limiting the scope of the claims.

[0088] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A brand monitoring and intelligent optimization method based on generation engine optimization, characterized in that, Includes the following steps: S1: Based on multi-source data acquisition and attribution link analysis of cross-modal detection, the multi-modal detection prompt word matrix is ​​input into the generative engine to obtain multi-source data content. The cross-modal alignment loss between brand standard assets and multi-source data content is calculated using a multi-modal alignment and evaluation model to locate cognitive biases. The source of citation is tracked to construct an attribution link map and determine toxic knowledge slices. S2: Generate adversarial knowledge injection carrier based on cross-modal semantic alignment. Analyze the distribution of cross-modal toxic knowledge slices in the encoding space of the multimodal retrieval engine of the generation engine, generate structured text, and inject images with implicit perturbation features and semantic tags to resist generation interference, as well as videos of key frames and structured text, to form multi-source optimized content. After passing the multimodal alignment and evaluation model verification, it constitutes a multimodal knowledge injection carrier. S3: Multi-source targeted distribution and holographic closed-loop verification. Multimodal knowledge is injected into the carrier and classified and delivered to the corresponding platform to seize the search ranking in order to cover cross-modal toxicity knowledge slices. After a preset period of time, the verification is re-probeed until the brand cross-modal alignment loss converges.

2. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 1, characterized in that, In step S1, a multimodal alignment and evaluation model is used to calculate the cross-modal alignment loss between brand standard assets and multi-source data content to locate cognitive biases. This includes the following sub-steps: S11: Extract visual features of images, semantic features of text, and temporal keyframe features of videos; S12: Obtain cross-modal alignment loss by calculating the cosine distance between the brand standard asset feature vector and the multi-source data content feature vector; S13: When the cross-modal alignment loss exceeds a preset threshold, it is determined that there is a cross-modal cognitive bias.

3. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 2, characterized in that, When the cross-modal alignment loss exceeds a preset threshold, it is determined that there is a cross-modal cognitive bias, and the preset threshold is set to 0.

35.

4. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 1, characterized in that, The multimodal alignment and evaluation model is built on a contrastive language image pre-training architecture. It collects image-text-visual triplet data from the brand domain as training, validation, and test sets, and uses a contrastive learning loss function for training. This model aims to narrow the distance between matching image-text-visual features and widen the distance between non-matching features.

5. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 4, characterized in that, The training process employs a contrastive learning loss function and uses the AdamW optimizer. The initial learning rate is set to 5e-5, and the number of training epochs is set to 30. When the loss function value on the validation set no longer decreases for three consecutive epochs, training is stopped early and the model weights are saved.

6. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 1, characterized in that, The S2 process injects an anti-generated interference stealth perturbation feature, which includes the following: using the original brand image, the brand standard text, and the upper limit of the perturbation budget as constraints, the applied adversarial perturbation is calculated by maximizing the similarity between the perturbation-injected image features and the brand standard text features, and the adversarial perturbation is superimposed on the pixel layer of the original brand image.

7. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 1, characterized in that, The S2 process, which involves verifying the multimodal alignment and evaluation model, includes: verifying the cosine similarity between each pair of text features, perturbed image features, and video keyframe features in the latent space. When the similarity is greater than 0.9, it is determined that the entity constitutes a high-weight cross-modal entity.

8. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 1, characterized in that, The S2 process generates multi-source optimized content for the video, which includes keyframes and structured text. This includes: extracting standard keyframes from the brand video, generating subtitle files containing accurate timestamps and descriptions of brand actions, and generating structured video transcription text.

9. The brand monitoring and intelligent optimization method based on generation engine optimization according to claim 1, characterized in that, In step S3, the multimodal knowledge injection carrier is classified and delivered to the corresponding platform, including: delivering text to a high-weight knowledge base, delivering brand images with metadata to mainstream image libraries and official website media pools, and delivering structured videos to video platforms with high crawling frequency.

10. A brand monitoring and intelligent optimization system based on generation engine optimization, characterized in that, To implement the brand monitoring and intelligent optimization method based on generation engine optimization as described in claim 1, the method includes: The cross-modal detection and attribution analysis module is used to execute S1. Based on the multi-source data acquisition and attribution link analysis of cross-modal detection, the multi-modal detection prompt word matrix is ​​input into the generative engine to obtain multi-source data content. The cross-modal alignment loss between brand standard assets and multi-source data content is calculated using the multi-modal alignment and evaluation model to locate cognitive bias. The citation source is tracked to construct an attribution link map and determine toxic knowledge slices. The adversarial knowledge injection carrier generation module is used to execute S2, generate adversarial knowledge injection carriers based on cross-modal semantic alignment, analyze the distribution of cross-modal toxic knowledge slices in the encoding space of the multimodal retrieval engine, generate images containing structured text, injected implicit perturbation features to resist generation interference and semantic tags, as well as multi-source optimized content of video with key frames and structured text, and constitute a multimodal knowledge injection carrier after passing the multimodal alignment and evaluation model verification. The targeted distribution and closed-loop verification module is used to execute S3, multi-source targeted distribution and holographic closed-loop verification, which classifies and delivers multimodal knowledge injection carriers to corresponding platforms to preemptively retrieve and sort information to cover cross-modal toxicity knowledge slices, and re-probes and verifies after a preset interval until the brand cross-modal alignment loss converges.

Images