Brand visibility detection method and system for generative artificial intelligence engine

Abstract

The application discloses a brand visibility detection method and system for generative artificial intelligence engines. A detection task containing a target brand logo and a prompt word set is obtained; the prompt word is submitted to multiple generative artificial intelligence engines, and answers are aggregated by engine-prompt word units; whether the target brand is mentioned and the reference source is analyzed; the reference source is merged by taking the domain name and the path end identifier as the merging key; the proportion visibility is calculated by taking the unit with the existing answer as the denominator and counting the non-mentioned into the denominator, and a global indicator G is obtained; |G-M1| and |G-M2| are compared for all-zero engines to distinguish between collection missing and true zero value, and when the collection is missing, the engine unit is excluded from the denominator; and a visibility detection result is output. The application realizes cross-engine comparable and auditable brand visibility measurement.

Description

Technical Field

[0001] This invention belongs to the field of network data analysis and artificial intelligence technology, specifically relating to a brand visibility detection method and system for generative artificial intelligence engines, as well as computer equipment and computer-readable storage media. Background Technology

[0002] Generative AI engines are becoming a crucial entry point for users to obtain information and support their decisions. Unlike traditional search engines that organize results in a stable list order, generative engines typically output answers as coherent natural language paragraphs or structured modules. Whether a brand is mentioned in the answer, in what context it is mentioned, and from which external content sources it is supported directly affect the brand's visibility in the AI-generated answer space. Therefore, the industry hopes to conduct reproducible and alignable brand visibility monitoring across multiple engine responses.

[0003] Existing approaches generally include: manually sampling engine-generated answers based on suggested keywords; performing keyword hit rate analysis on the full text of the answers; or superficially migrating ranking monitoring tools that are geared towards search results pages. When expanding the monitoring scope from webpage lists to generated answers, the above approaches face at least the following technical obstacles: (1) Lack of cross-engine alignable statistical units. The output structure, length and citation presentation of different engines are heterogeneous. If the mention ratio is only counted on the subset of "hit samples", the observation units that have not been mentioned but have successfully returned answers will be excluded from the denominator, resulting in the systematic overestimation of visibility and the failure of cross-engine horizontal comparison; (2) The source identifier has surface variations such as multilingual path prefixes, tracking query parameters, and text fragment anchors. If the contribution is accumulated based on the original Uniform Resource Locator string, the same content asset will be counted as multiple sources, the content contribution measurement will be distorted, and subsequent attribution and operational decisions will be polluted by noise; (3) When the visibility of a certain engine is zero across all relevant prompts, the existing process has difficulty distinguishing between "failed data collection channel, incomplete answer" and "the engine did not actually mention the brand". If the missing data collection is mistakenly treated as a true zero value and retained in the denominator, or if the engine is deleted in an unauditable manner, the global visibility metric will be distorted. (4) If there is no structured marking and completion mechanism for states such as timeout, empty response, fold not unfolded, and streaming interruption during the collection process, "not captured at the time" and "not explicitly mentioned by the engine" will be mixed together, making the visibility results unverifiable.

[0004] Therefore, there is an urgent need for a closed-loop solution consisting of executable data processing steps to obtain a brand visibility metric that can be compared across engines, merge references from the same source, and audit all-zero results, given that generative answers lack a unified ranking number, multiple engines have heterogeneous outputs, and there may be missing data collection. Summary of the Invention

[0005] [Technical Issue] How to obtain a brand visibility metric that can be compared across engines, merge references from the same source, and audit all-zero results under the conditions of lack of unified ranking for generative answers, heterogeneous output from multiple engines, and potential data collection gaps, while avoiding sample bias, reference variant noise, and incorrect denominator processing.

[0006] [Technical Means] This invention employs the following essential technical means to form the main closed loop: acquiring brand detection tasks; submitting prompt words to multiple generative artificial intelligence engines and forming an answer set with engine and prompt word combinations as record units; parsing whether the source is mentioned and cited; using the combination of domain name and path end identifier as the merging key to merge sources; using statistical units with existing answer content as the denominator and including unmentioned items in the denominator to obtain prompt word-level visibility and global visibility index G; determining whether the all-zero engine has missing data or a true zero value based on the comparison of |G−M1| and |G−M2|, and excluding the corresponding statistical unit from the denominator when it is determined to have missing data; outputting detection results that at least include the global visibility index. Search volume weighting, attribute expansion, content completion, bounded resampling, and time-series snapshots are preferred extensions, entering a more specific scope of protection without changing the essential feature set defined by the independent claims.

[0007] [Technical Effects] It can provide proportional visibility that can be compared horizontally under cross-engine conditions; suppress content contribution noise caused by URL variations; improve the auditability of denominator processing in all-zero result scenarios; and enable the monitoring link to have a data foundation that can be labeled, completed, and verified when the collection is incomplete.

[0008] In view of this, embodiments of the present invention provide a brand visibility detection method for generative artificial intelligence engines. The present invention also relates to a brand visibility detection system, a computer device, and a computer-readable storage medium, to address the aforementioned technical deficiencies in the prior art.

[0009] According to a first aspect of the present invention, a method for brand visibility detection for generative artificial intelligence engines is provided, comprising the following steps: Step S1: Obtain a brand detection task, which includes at least the target brand identifier and a set of prompt words for initiating a probe to the generative artificial intelligence engine.

[0010] Step S2: Submit the prompt words in the prompt word set to multiple generative artificial intelligence engines to obtain the corresponding answer content, forming an answer set with the combination of engine and prompt words as the recording unit.

[0011] Step S3: parse the answer content in the answer set, identify whether the target brand is mentioned, and obtain the source identifier when it is mentioned.

[0012] Step S4: Merge the reference source identifiers, where the combination of the domain name and the path end identifier is used as the merge key to merge multiple reference source identifiers pointing to the same content into the same content source.

[0013] Step S5: Using the combined answer of the engine and the prompt words as the statistical unit, count the number of statistical units that mention the target brand as the numerator, and count the number of statistical units that have answer content as the denominator. Statistical units that have answer content but do not mention the target brand are included in the denominator. The prompt word-level visibility is obtained, and the global visibility index G is obtained accordingly.

[0014] Step S6: When there is an engine with zero visibility across all relevant prompts, calculate the first mean M1 including the engine and the second mean M2 after removing the engine. Based on the comparison result of |G−M1| and |G−M2|, determine whether the engine belongs to missing data or true zero value. If it is determined to be missing data, exclude the corresponding statistical unit of the engine from the denominator in the global visibility calculation.

[0015] Step S7: Output the brand visibility detection result, which includes at least the global visibility index.

[0016] Optionally, in step S1: the target brand identifier includes at least one of the brand name and the official website domain name; the prompt words in the prompt word set are associated with search volume weights and marked as enabled or disabled, and only prompt words in the enabled state proceed to step S2; the prompt words are generated by expanding business domain information and can be marked with intent type tags.

[0017] Optionally, in step S2: a collection status flag is written for each recording unit; when the answer content does not meet the preset integrity conditions, at least one of the following content completion operations is performed on the corresponding session or interface: wait, expand, scroll, follow the question, or re-request before writing it into the answer set.

[0018] Optionally, in step S3: entity recognition is performed based on the target brand alias set to determine whether it is mentioned; when it is determined to be mentioned, at least one attribute feature is extracted from sentiment polarity, mention position and co-occurring competitor entity, wherein the mention position includes at least one of the first appearance position, paragraph position or list sorting position.

[0019] Optionally, in step S4: the domain name and path of the source identifier are resolved, the path end identifier is extracted, and the multilingual path prefix, query parameters and fragment anchors that do not affect the content are removed before generating the merge key; the number of times the content source appears after merging is counted to obtain the content contribution.

[0020] Optionally, in step S5: the global visibility index G is obtained by weighting and summarizing the visibility of each prompt word according to the search volume weight of each prompt word; optionally, the global visibility result with the second caliber is obtained by assigning higher weights to the engines that are mentioned.

[0021] Optionally, in step S6: the first mean M1 and the second mean M2 are calculated using weights that are consistent with or configurable to the search volume weights of each prompt word; when |G-M2| < |G-M1|, it is determined to be a missing data collection; when |G-M1| < |G-M2|, it is determined to be a true zero value and retained; when |G-M1| = |G-M2|, it is determined to be a true zero value by default and retained, or it is marked as pending review and not immediately excluded from the statistical unit of the engine; after determining that the data collection is missing, bounded re-collection and rewriting can be performed on the relevant engine and prompt word tasks, and the visibility can be recalculated.

[0022] Optionally, steps S2 to S6 are repeated at a set period, the visibility index of each period is stored as a time-series snapshot, and the contribution mapping between the merged content source and the visibility of the prompt word is output.

[0023] Optionally, the multiple generative AI engines mentioned in step S2 include at least two engine instances that are distinguishable in the call chain or output structure; the data collection can be performed using parallel scheduling or serial scheduling; when parallel scheduling is used, independent rate limiting and timeout parameters are maintained for each engine to avoid congestion of a single engine causing the entire task grid to fail.

[0024] Optionally, in step S5, before calculating the visibility of the prompt word, the unscoreable units are filtered according to the collection status mark, and only records that meet the integrity condition are kept in the denominator and numerator; for prompt words with insufficient data, a sample size field is attached to the output instead of forcibly giving misleading extreme proportions.

[0025] Optionally, in step S6, M1 and M2 calculations and distance comparisons are performed independently for each engine in the set of all zero engines, allowing some engines to be removed based on missing data and others to be retained based on true zero values, thereby avoiding the deletion of all zero-value engines at once.

[0026] It should be noted that the above optional implementation methods are used to provide a more specific technical path in the sense of subordinate protection. Those skilled in the art will understand that the main closed loop defined by the independent claims can still be implemented even without the above optional configuration.

[0027] According to a second aspect of the present invention, a brand visibility detection system for generative artificial intelligence engines is provided, comprising: The task configuration module is configured to acquire a brand detection task, wherein the brand detection task includes at least the target brand identifier and a set of prompt words for initiating a probe to the generative artificial intelligence engine; The multi-engine acquisition module is configured to submit the prompt words in the prompt word set to multiple generative artificial intelligence engines respectively, obtain the corresponding answer content, and form an answer set with the combination of engine and prompt word as the recording unit; The parsing module is configured to parse the answer content in the answer set, identify whether the target brand is mentioned, and obtain the citation source identifier when mentioned. The source merging module is configured to merge the reference source identifiers, wherein the combination of the domain name and the path end identifier is used as the merging key to merge multiple reference source identifiers pointing to the same content into the same content source. The visibility calculation module is configured to use the combined answer of the engine and prompt words as the statistical unit, count the number of statistical units that mention the target brand as the numerator, and count the number of statistical units that have answer content as the denominator. Statistical units that have answer content but do not mention the target brand are included in the denominator to obtain the prompt word-level visibility and obtain the global visibility index G based on it. The integrity determination module is configured to calculate the first mean M1 containing the engine and the second mean M2 after removing the engine when there is an engine with zero visibility across all relevant prompts. Based on the comparison result of |G-M1| and |G-M2|, the module determines whether the engine belongs to missing data or true zero value. If it is determined to be missing data, the corresponding statistical unit of the engine is excluded from the denominator in the global visibility calculation. The result output module is configured to output brand visibility detection results, and the detection results include at least the global visibility index.

[0028] The aforementioned modules support each other functionally in the data processing flow: the task configuration module provides alignable prompts and brand identifiers for data collection; the multi-engine data collection module provides a set of answers with status markers for parsing; mention parsing and source aggregation provide structured fields for visibility calculation; visibility calculation provides G and engine-level sequences for integrity determination; and the integrity determination writes back the denominator correction results before the results output module publishes them externally. Optionally, the system is also configured as an extension unit to implement any optional step in the method described in the first aspect.

[0029] According to a third aspect of the present invention, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the first aspects.

[0030] According to a fourth aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the first aspects.

[0031] Compared with the prior art, the present invention can achieve the following beneficial effects: (1) The combined answers of engine and prompt words are used as a unified statistical unit, and units with answers but no mention of the target brand are included in the denominator to avoid overestimating visibility due to only counting hit samples, so that cross-engine results can be compared horizontally. (2) Using the combination of domain name and path end identifier as the merge key, before merging, remove variants such as multilingual path prefixes, tracking parameters and fragment anchors to suppress the contribution noise caused by the repeated counting of the same content asset; (3) When a zero engine appears, the missing data or the true zero value is determined by comparing |G-M1| and |G-M2|, and the denominator is corrected when the data is missing, so that the global visibility processing process can be audited and verified. (4) By using optional mechanisms such as collection status marking and content completion, and bounded re-collection, “collection failure” and “not mentioned in reality” are separated at the data layer, thereby improving the feasibility and robustness of the monitoring link under unstable networks and heterogeneous interactive interfaces. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the brand visibility detection system provided in an embodiment of the present invention; Figure 2 This is the main flowchart of the brand visibility detection method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the citation source merging process provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the data acquisition integrity judgment process provided in the embodiments of the present invention. Detailed Implementation

[0033] 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. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. Numerical values, field names, concurrency parameters, and interface forms used in this document are illustrative and do not constitute a limitation on the scope of protection. This invention was completed by a natural person inventor; this embodiment does not involve the collection of facial information in public places for marketing purposes.

[0034] The terminology used in one or more embodiments of the present invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. The term “and / or” includes any or all combinations of one or more of the related listed items. The terms “first,” “second,” etc., are used to distinguish objects of the same kind and do not indicate a fixed constraint of order or importance.

[0035] The following is an explanation of some of the terms used in the embodiments of the present invention.

[0036] Generative AI Engine: A service or interactive channel capable of receiving natural language prompts and generating answer text, including but not limited to application programming interface (API) call channels and human-computer dialogue page channels; this invention does not limit the commercial name of the engine.

[0037] Tip words: Natural language statements used to initiate a probe to the generative AI engine; tips words in the probe set can include search volume weights, enabled status, and intent type labels.

[0038] Statistical unit: A single answer record that uses the combination of engine identifier and prompt word identifier as the logical key and contains scoreable answer content; it is the basic alignment unit when calculating prompt word-level visibility and global visibility.

[0039] Merge key: A content identity identifier composed of a domain name and a path terminator, used to map multiple seemingly different source identifiers to the same content source.

[0040] Global visibility index G: The overall visibility measure is obtained by summing the visibility of each prompt word level according to equal weight or weighted rules; there may be a G before correction and a G_final after correction before and after integrity judgment.

[0041] First mean M1 and second mean M2: When an all-zero engine exists, these correspond to the global mean values ​​under the two statistical paths of retaining and eliminating the engine unit, respectively, and are used to compare the absolute distance with G.

[0042] In practical applications, brands or service providers want to observe whether multiple generative engines mention the target brand on a unified set of prompts. If only fragments indicating "brand name seen" are counted, a large number of units that have returned answers but do not mention the brand will be ignored; if the original URL is not denoised, similar content will proliferate; if an engine returns a batch of zeros without missing item checks, the denominator will be incorrectly contaminated. This invention addresses these obstacles by providing a deployable system and reproducible methods.

[0043] like Figure 1 As shown, in one system-side implementation, the brand visibility detection system includes a task configuration module, a multi-engine acquisition module, a mention parsing module, a source merging module, a visibility calculation module, an integrity discrimination module, and a result output module. Each module can be deployed on the same computing node or distributed, collaborating through in-process calls, remote procedure calls, message queues, or shared structured storage media. The internal naming of the modules can use engineering aliases such as prompt word configuration unit or session acquisition unit, without altering the set of technical means defined in the independent claims.

[0044] The task configuration module is used to acquire and solidify brand detection tasks. In one possible implementation, this module provides a task creation interface or configuration interface, receiving a target brand identifier and a set of prompt words. The target brand identifier includes at least one of the following: a brand name string and an official website domain name. Each prompt word in the prompt word set is a natural language probing statement that can be submitted to the engine. In a preferred implementation, each prompt word can also be associated with search volume weight, enabled or disabled status, and intent type tags such as recommendation, comparison, problem-solving, and selection decision; only prompt words in the enabled state are included in the collection and visibility statistics. The task is written to persistent storage and a unique task identifier is generated for alignment with collection logs, write-back, re-collection, and time-series snapshots.

[0045] In the example of interface fields, the task object may contain: task_id, brand_name, brand_domain, prompt_id, prompt_text, intent_type, weight, enabled, and created_at. The above field names are examples; the implementation can be mapped to different storage table structures, as long as the data relationship of "brand identifier - prompt word - enabled status - optional weight" is maintained.

[0046] The multi-engine data collection module is used to collect data from multiple generative AI engines on a unified set of prompt words, forming a reconcilable set of answers. "Multiple" refers to at least two engine instances that are distinguishable in their output structure or call chain, such as different vendor interfaces, different web session channels, or different model endpoints; commercial names are not a restriction. For each prompt word in the prompt word set, a single request or a set of requests initiated according to a strategy is sent to each engine, writing the answer text, a list of optional reference links, and a timestamp into a record unit. Each record unit uses "engine identifier × prompt word identifier" as a logical key to ensure cross-engine statistical alignment.

[0047] As one possible implementation, the acquisition module writes an acquisition status flag for each record unit. The flag enumeration can include: normal completion, timeout, interface error, authentication failure, empty response, content suspected of being truncated or collapsed, and manual termination. When the answer content does not meet the preset integrity conditions—for example, the number of characters in the body text is lower than the threshold, the collapse control is not expanded, or the streaming output is interrupted—the module performs at least one content completion operation on the corresponding session or interface, such as waiting, expanding, scrolling, following the question, or re-requesting, and then writes it to the answer set. Optionally, the acquisition adapter is loaded according to the engine type. The adapter exposes a unified submit(prompt_text)→{answer_text, references[], status, raw_meta} interface, so that the visibility calculation formula does not need to be modified when adding a new engine.

[0048] The mention parsing module analyzes the answer content. The basic output is a binary field "whether the target brand is mentioned"; if it is determined to be mentioned, at least a set of citation source identifiers is output. Citation source identifiers can be parsable URLs, annotation numbers mapped to URLs, or other locators pointing to external content carriers. One possible implementation for entity recognition is to load a set of aliases (full name, abbreviation, case and hyphen variations, multilingual spellings, misspelling whitelist), and combine this with dictionary matching, regular expression matching, or sequence labeling models to obtain the mention span; negative sentence filtering and domain whitelisting can be overlaid to reduce false triggers. Filter thresholds are standard engineering choices and are not used to limit the independent claims.

[0049] In the dependent extension implementation, when a mention is determined, the following can also be extracted: sentiment polarity (positive / neutral / negative); mention position (first character offset, paragraph number, k-th item in the list); and co-occurring competitor entities. The mention position can be used for interpretation or optional weighting, but the minimum sufficient statistics of the main loop of the independent claim do not depend on the position formula.

[0050] The source merging module is used to prevent the same content asset from being counted repeatedly due to different URL spellings. For example... Figure 3 As shown, a reproducible merging process is as follows: (1) Read in the original source identifier string and perform Unicode normalization and whitespace clipping; (2) Parse the protocol, hostname (domain name), path, query parameters and fragment anchors according to URI syntax; lowercase the hostname and remove the default port if necessary; (3) Remove the preset multilingual path prefixes, such as path segments composed of language or region codes, and verify whether the path still points to a parseable content path after removing the prefix. (4) Delete tracking query parameters that do not affect the content target, and retain business parameters that may distinguish the content; the list of tracking parameters is configurable; (5) Remove emphasized sections and pure anchor points from the browser; (6) Extract the path segment identifier: Take the last non-empty path segment and remove the trailing forward slash; if the last segment is a common entry file name, then fall back to use the previous path segment or the site's agreed content primary key; (7) Using the combination string of “domain name + path end identifier” as the merge key, multiple original references under the same merge key are recorded as the same content source, and the number of occurrences is accumulated as the basis for content contribution.

[0051] The path end identifier is an extractable abstraction of the content primary key, and its form can be a document alias, a numeric primary key, a semantic path end segment, or an object storage key end segment; the embodiment does not include the field name of a certain commercial content management system in the independent claim.

[0052] The visibility calculation module generates a proportional metric that can be compared across engines. A statistical unit is defined as a valid response record where the answer content exists on both the engine and the suggestion word combination. For a given suggestion word p, let the set of response units be U(p), and the set of units mentioning the target brand be U_hit(p), then the suggestion word-level visibility is: V(p) = |U_hit(p)| / |U(p)|.

[0053] The denominator |U(p)| counts units containing answer content; units containing answers but not mentioning the target brand must be included in the denominator. If all three engines return a complete answer but only one engine mentions the brand, then V(p) = 1 / 3, instead of being recorded as 100% for errors. The global visibility metric G is obtained by summing the visibility of each prompt word. The minimum implementation at the independent claim level can use the enabled prompt word arithmetic mean: G = (1 / |P|)·Σ_p V(p). In the dependent implementation, if the weight w(p) ≥ 0, then G = Σ_p w(p)·V(p) / Σ_p w(p). Example of the second caliber: The set of engines that appear at least once is weighted and then re-summarized for observation of sensitivity; the default report still uses the first caliber of the task configuration.

[0054] The integrity assessment module handles the auditable branch where "the visibility of a certain engine is zero across all relevant suggestion words," and the corresponding process... Figure 4 The relevant prompts refer to the subset of prompts enabled for this engine. If the engine visibility vectors in this subset are all zero, then it enters the all-zero candidate set Z.

[0055] Let G be the global metric calculated according to the default denominator rule (including the engine unit). Calculate M1 and M2: M1 is the global mean that is isomorphic to G or has configurable weights under the path of the all-zero engine unit; M2 is the recalculated global mean after removing the corresponding statistical unit of the engine. One implementation is to "calculate engine-level visibility first and then average" with / without the engine unit; another implementation is to directly perform two weighted averages with / without the engine unit on the global unit set. Both share the same idea: use the distance of G relative to M1 and M2 to determine whether all zeros are more like structural missing parts or a true unmentioned whole.

[0056] One possible implementation scheme for the discrimination rule: (1) If |G-M2| < |G-M1|, it is determined to be a tendency of missing data collection. The corresponding statistical unit of the engine is excluded from the global denominator and marked as missing data collection. (2) If |G-M1| < |G-M2|, it is determined to be a true zero value tendency, and the contribution of the engine unit in the denominator is retained (the numerator is zero), so as to avoid artificially increasing visibility by deleting engines; (3) If |G-M1|=|G-M2|, it is reserved as the true zero value by default, or marked as pending review and not excluded immediately.

[0057] Preferably, after determining that the data collection is missing, bounded re-collection can be performed on the relevant engine and prompt words: each task can be retried a maximum of T times (T is a positive integer), using backoff waiting, changing the session, or switching to a backup data collection channel; if successful, the data is written back and recalculated from step S3; if it still fails, the denominator is excluded and the audit field is retained.

[0058] Multiple operating conditions are explained below. Condition A: If only one engine has all zeros and |G-M2| is significantly smaller, it is discarded as missing. Condition B: If the brand is not mentioned in any online suggestions, all engines are close to true zero, and |G-M1| is smaller or equivalent, the denominator is retained to avoid "deleting engines to artificially inflate values". Condition C: If both engines have all zeros, M1 / M2 comparison is performed independently for each engine in Z, allowing for partial rejection and partial retention. Condition D: If G itself has insufficient samples (too few effective suggestions), a data insufficiency flag can be output, and automatic rejection can be suppressed. The above conditions are used to illustrate the feasibility of the discriminator; the numerical thresholds are configurable.

[0059] The results output module outputs the detection results. The minimum output includes G or G_final after denominator correction; preferred outputs also include: a prompt word-level V(p) table, engine visibility, all-zero discrimination flags, a list of excluded engines, merged content contribution ranking, co-occurring competitor statistics, task timestamps, and calculation caliber version numbers. Output media can be interface responses, spreadsheets, database records, or page objects.

[0060] like Figure 2 As shown, the main process flow is executed sequentially from steps S1 to S7. More practical sub-steps are given below for implementation by those skilled in the art.

[0061] For step S1, one implementation includes the following sub-steps: Step S11: Create a detection task and write a unique task identifier; Step S12: Enter or import the target brand identifier, and verify that the name is not empty or the domain name can be resolved, at least one of the following conditions must be met; Step S13: Enter or import the prompt word set; optionally, expand candidate questions from seed words in the business domain and review and add them to the database; Step S14: Mark the enabled status; optionally mark the search volume weight and intent type; filter to obtain the probe set P.

[0062] For step S2, one implementation includes the following sub-steps: Step S21: Enumerate the engine set E and the probe set P to generate the data collection task grid E×P; Step S22: Schedule data collection according to concurrency C (C≥1); for each (e,p), call the adapter to submit the prompt text; Step S23: Receive the answer text and citations; if the integrity condition fails, perform the completion action subset and retrieve again; Step S24: Write to the recording unit and mark the acquisition status; timeouts and errors can enter the bounded retry queue.

[0063] For step S3, one implementation includes the following sub-steps: Step S31: Load the brand alias set and optional negation rules; Step S32: Standardize the text of the answer (full-width / half-width characters, case folding strategy according to language configuration); Step S33: Perform entity recognition to obtain a binary determination of whether it is mentioned and the alias form of the hit; Step S34: If the reference is true, extract the list of reference addresses and the optional sentiment, location, and co-occurring competitor fields and write them back to the cell.

[0064] For step S4, one implementation method is as follows: Figure 3 In addition to merge key generation, it also includes: Step S41: Generate a merge key for each original reference address; those that cannot be resolved are either placed in the unresolvable bucket or processed according to the discard policy. Step S42: Aggregate the counts within the task scope using the merge key to obtain the content contribution table; Step S43: Write back the merged content source identifier to the corresponding mention unit.

[0065] For step S5, one implementation includes the following sub-steps: Step S51: Select the units whose collection status is scoreable to form a statistical population; Step S52: For each prompt word, count the number of hit units and the number of answer units, and calculate V(p); if |U(p)|=0, then mark it as insufficient data and remove or place it according to the rules; Step S53: G is obtained by summing the results according to equal weight or search volume weight; Step S54: Optionally, calculate and cache the second caliber; the default reporting caliber is based on the task configuration.

[0066] For step S6, one implementation includes the following sub-steps: Step S61: Scan the visibility vector of each engine on the relevant prompt words and identify the set of all-zero engines Z; skip if Z is empty. Step S62: For each candidate engine z∈Z, calculate G, M1, and M2 based on the current unit set; Step S63: Compare |G-M1| with |G-M2| and write the discrimination label; Step S64: For the engine that collects missing labels, exclude its cells from the subsequent global denominator, and optionally trigger bounded resampling; Step S65: After correcting the denominator, the global visibility index G_final for output is recalculated.

[0067] For step S7, serialize G or G_final with the audit field and output the result, then push it to the subscriber or write it to the result database.

[0068]

Example 1: Minimal Operation of Main Closed Loop

[0069] [Example 2: Denominator Overestimation Prevention and Numerical Exercise] A certain prompt keyword has 37 statistical units with existing answer content, of which 30 mention the target brand, so V = 30 / 37 ≈ 81.08%. If only the hit units are counted incorrectly, 7 "answer-not-mentioned" units will be discarded, inflating the percentage to 100%. Assuming the aggregated G = 27.6%, a certain zero-based search engine calculates M1 = 17.9% and M2 = 29.9%. Since |G-M2| = 2.3% is less than |G-M1| = 9.7%, it is treated as missing data and recalculated. The above values ​​are exemplary data and do not constitute a limitation on the scope of protection.

[0070] [Example 3: Merge Key Denoising] When the same content page is referenced by multiple addresses with different language prefixes and tracking parameters, after stripping the language prefixes, tracking parameters, and emphasis anchors, if the domain names are the same and the path end identifiers are the same, then the merge key is the same, and the contribution is accumulated as for the same asset. If the path ends are different, then they are not merged.

[0071] [Example 4: Dependent Enhancement Implementation] Based on Example 1, this embodiment incorporates a weighted index of prompt word search volume; enables folding / expanding and scrolling completion at the data collection end; outputs sentiment and co-occurring competitors at the parsing end; writes weekly time-series snapshots and outputs a descriptive contribution mapping from the source to the visibility of prompt words; and configures the bounded number of resampling times as a positive integer. This embodiment corresponds to a more specific protection circle in the dependent claims.

[0072] [Example 5: Computer Equipment and Storage Media] Figure 2 The process is embedded in a computer program, stored on non-volatile media, and loaded and executed by a processor. The program reads the task configuration, calls the acquisition adapter, writes intermediate results to disk, and finally writes them back to the result table. Changing the engine type, prompt word size, weight source, field order, re-sampling limit, concurrency, etc., are all conventional substitutions that do not depart from the essence of this invention.

[0073] [Example 6: Failure and Boundaries] When all engines have no valid answers for all prompt words, output insufficient data and do not provide misleading extreme visibility readings. When only a single engine is available, cross-engine comparison conditions are insufficient; the main comparison task should be rejected or downgraded to single-engine diagnostic mode with explicit annotation. When there are very few valid units for a prompt word, provide a sample size hint to avoid over-extrapolation. This type of guardrail improves robustness without reducing the essential feature set of the independent claim.

[0074] [Example 7: Independent Determination of Simultaneous Zero Values ​​in Both Engines] Suppose that engines A and B are both zero in the relevant prompt words. Calculate M1 / M2 for A and B respectively and compare it with G: If only A satisfies |G-M2|<|G-M1|, then only the elements of A are excluded; if B satisfies the true zero value condition, then it is retained. This example illustrates that elements within the set Z of zero values ​​should be determined independently, rather than deleting all zero-value engines at once.

[0075] [Example 8: Data Acquisition Adaptation and Status Write-back] For interface-based engines, the adapter acquires text via synchronous or streaming requests; for human-computer dialogue-based engines, the adapter automatically submits prompts through the page and reads the answer container from the reachability tree. Both write back to the same record unit pattern: {engine_id, prompt_id, answer_text, references, status, fetched_at}. Visibility calculation relies only on standard fields in this pattern, thus achieving "replaceable acquisition channels and stable statistical caliber".

[0076] The present invention also provides system embodiments corresponding to the second aspect and computer device and storage medium embodiments corresponding to the third and fourth aspects: the memory stores a program, which, when executed by the processor, implements any step of the above method; the program on the storage medium, when executed, implements any step of the above method. Details not elaborated in the device embodiments can be found in the corresponding descriptions of the method embodiments.

[0077] The technical features of the embodiments described above can be combined arbitrarily. For the sake of brevity, not all possible combinations have been described; any combination that does not contradict itself should be considered as described in the specification. The above embodiments only illustrate several implementation methods and are described in a relatively specific and detailed manner, but should not be construed as limiting the scope of protection of the invention patent. Modifications and improvements made by those skilled in the art without departing from the concept of the invention are all within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A brand visibility detection method for generative artificial intelligence engines, characterized in that, Includes the following steps: Step S1: Obtain a brand detection task, wherein the brand detection task includes at least the target brand identifier and a set of prompt words for initiating a probe to the generative artificial intelligence engine; Step S2: Submit the prompt words in the prompt word set to multiple generative AI engines to obtain corresponding answer content, forming an answer set with the combination of engine and prompt words as the recording unit; Step S3: Analyze the answer content in the answer set, identify whether the target brand is mentioned, and obtain the citation source identifier when mentioned; Step S4: Merge the citation source identifiers, using the combination of domain name and path end identifier as the merge key, merging multiple citation source identifiers pointing to the same content into the same content source; Step S5: Use the combined answer of engine and prompt words as the statistical unit to count the mentions of the target. The number of statistical units for the brand is used as the numerator, and the number of statistical units with answer content is used as the denominator. Statistical units with answer content but without mentioning the target brand are included in the denominator to obtain the visibility at the prompt word level and thus obtain the global visibility index G. Step S6: When there is an engine with zero visibility across all relevant prompt words, calculate the first mean M1 including the engine and the second mean M2 after removing the engine. Based on the comparison result of |G-M1| and |G-M2|, determine whether the engine belongs to missing data or true zero value. If it is determined to be missing data, exclude the corresponding statistical unit of the engine from the denominator in the global visibility calculation. Step S7: Output the brand visibility detection result, which includes at least the global visibility index.

2. The method according to claim 1, characterized in that, In step S1: the target brand identifier includes at least one of the brand name and the official website domain name; the prompt words in the prompt word set are associated with search volume weights and marked as enabled or disabled, and only prompt words in the enabled state proceed to step S2; the prompt words are generated by expanding business domain information and can be marked with intent type tags.

3. The method according to claim 1, characterized in that, In step S2: a collection status flag is written for each recording unit; when the answer content does not meet the preset integrity conditions, at least one of the following content completion operations is performed on the corresponding session or interface: wait, expand, scroll, follow the question, or re-request before writing it into the answer set.

4. The method according to claim 1, characterized in that, In step S3: entity recognition is performed based on the target brand alias set to determine whether it is mentioned; when it is determined to be mentioned, at least one attribute feature is extracted from sentiment polarity, mention position and co-occurring competitor entity, wherein the mention position includes at least one of the first appearance position, paragraph position or list sorting position.

5. The method according to claim 1, characterized in that, In step S4: the domain name and path of the reference source identifier are resolved, the path end identifier is extracted, and the multilingual path prefix, query parameters and fragment anchors that do not affect the content are stripped before generating the merge key; The content contribution is calculated by counting the number of times each content source appears after aggregation.

6. The method according to claim 1 or 2, characterized in that, In step S5: the global visibility index G is obtained by weighting and summarizing the visibility of each prompt word according to the search volume weight of each prompt word; optionally, the global visibility result with the second caliber is obtained by assigning higher weight to the engines that are mentioned.

7. The method according to claim 1, characterized in that, In step S6: the first mean M1 and the second mean M2 are calculated using weights that are consistent with or configurable to the search volume weights of each prompt word; when |G-M2| < |G-M1|, it is determined to be a missing data; when |G-M1| < |G-M2|, it is determined to be a true zero value and retained; when |G-M1| = |G-M2|, it is determined to be a true zero value by default and retained, or it is marked as pending review and not immediately excluded from the engine statistical unit; After determining that the data collection is missing, bounded recollection can be performed on the relevant engine and prompt word tasks, and the visibility can be recalculated after writing back.

8. The method according to claim 1, characterized in that, Also includes: Repeat steps S2 to S6 according to the set period, store the visibility index of each period as a time-series snapshot, and output the contribution mapping between the merged content source and the visibility of the prompt words.

9. A brand visibility detection system for generative artificial intelligence engines, characterized in that, include: The task configuration module is configured to acquire a brand detection task, wherein the brand detection task includes at least the target brand identifier and a set of prompt words for initiating a probe to the generative artificial intelligence engine; The multi-engine acquisition module is configured to submit the prompt words in the prompt word set to multiple generative artificial intelligence engines respectively, obtain the corresponding answer content, and form an answer set with the combination of engine and prompt word as the recording unit; The document describes a system that includes a parsing module, a source merging module, and a source merging module. The source merging module is configured to merge the source quoting identifiers, using a combination of domain name and path terminator as the merging key to merge multiple source quoting identifiers pointing to the same content into a single source. The visibility calculation module uses the combination of engine and prompt words as the statistical unit, calculating the number of statistical units mentioning the target brand as the numerator and the number of statistical units containing answer content as the denominator. Statistical units containing answer content but not mentioning the target brand are included in the denominator, resulting in prompt word-level visibility and a global visibility index G. The integrity determination module is configured to calculate the first mean M1 containing the engine and the second mean M2 after removing the engine when there is an engine with zero visibility across all relevant prompts. Based on the comparison result of |G-M1| and |G-M2|, the module determines whether the engine belongs to missing data or true zero value. If it is determined to be missing data, the corresponding statistical unit of the engine is excluded from the denominator in the global visibility calculation. The result output module is configured to output brand visibility detection results, and the detection results include at least the global visibility index.

10. The system according to claim 9, characterized in that, The system is also configured to implement additional steps of the method according to any one of claims 2 to 8.

11. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 8.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 8.

Images