Key takeaways<\/h2>\n\n\n- GoDaddy’s Airo Plus Advertising Advisor makes use of OpenAI’s function-calling system to create customized social media content material for companies.<\/li>\n
- The system overcomes challenges like non-deterministic prompts and response format variability by cautious immediate engineering and error dealing with.<\/li>\n
- Person interface design and clear steerage are essential for efficient AI-powered advertising instruments.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
In early 2023, GoDaddy started integrating AI into its advertising capabilities to make advertising duties simpler and extra accessible for customers. This consists of social media publish creation, electronic mail advertising campaigns, weblog content material era, and search engine optimisation.<\/p>\n
By incorporating AI, we offered customers with a place to begin for his or her advertising duties, producing customized content material and saving them time. Many customers confronted the “clean syndrome,” struggling to start out making a publish; so as a substitute of brainstorming manually, they now obtain tailor-made solutions, dashing up their workflow and constructing belief in our instruments.<\/p>\n
Nevertheless, we observed many customers needed a extra customized, interactive expertise\u2014much like working with a advertising company. This perception drove us to deal with enhancing our instruments, providing dynamic, conversation-driven assist resulting in the creation of the Airo\u2122 Plus Advertising Advisor.<\/p>\n
Background<\/h2>\nThe Airo\u2122 Plus Advertising Advisor is a instrument designed to information customers by numerous advertising actions with ease. It leverages the facility of AI to reply rapidly and precisely to person prompts. For instance, if a person desires to create a social media publish for his or her enterprise, Airo\u2122 Plus Advertising Advisor supplies tailor-made suggestions based mostly on the person’s enterprise such because the enterprise title, the web site, the enterprise class, and different related social information together with their previous Fb posts.<\/p>\n
To make this occur, we utilized OpenAI\u2019s\u00a0function-calling paradigm<\/a>. Perform calling permits any language mannequin, like GPT-4o-mini, to immediately work together with exterior APIs or methods in a structured manner. We created a set of capabilities mapped to particular actions, comparable to creating, previewing, modifying, and publishing a social publish. When a person supplies textual content enter, OpenAI maps the immediate to considered one of these predefined capabilities.<\/p>\nFor instance, if a person says, “I need to create a social publish about my new product: chocolate chip cookies<\/em>“, the assistant identifies the immediate and triggers a operate that calls an exterior API we created. This API first retrieves the client\u2019s enterprise info and social metadata. It then makes use of this information to generate an in depth OpenAI immediate that features the context\u2014particularly, that the publish is a couple of chocolate chip cookie product. The aim of this immediate is to provide a related and interesting social media caption for the enterprise. By combining AI capabilities with structured operate mapping, the assistant ensures an environment friendly, intuitive expertise tailor-made to the person’s wants.<\/p>\nWithin the following screenshot, the person has initiated a request to the Airo\u2122 Plus Advertising Advisor to generate a social publish about their chocolate cookies product:<\/p>\n
\n<\/figure>\n<\/div>\nAt this stage, the assistant is processing the request, gathering enterprise information, and making ready a related publish.\u00a0Within the following screenshot, the AI-generated social publish is displayed, full with a instructed caption, picture, and scheduling particulars: <\/p>\n\n![\"screenshot](\"https:\/\/www.godaddy.com\/resources\/wp-content\/uploads\/2025\/03\/marketing_assistant.png?size=3840x0\"\/)
<\/figure>\n<\/div>\nThe person can overview the publish, make edits if essential, or proceed with publishing.\u00a0<\/p>\n
The way it works<\/h2>\nAiro\u2122 Plus Advertising Advisor makes use of OpenAI\u2019s function-calling system to assist create and handle social media content material. Right here\u2019s the way it works:<\/p>\n
\n- Person enter:<\/strong> The person enters a request, like “Create a social publish about cookies.<\/em>“<\/li>\n
- System understanding:<\/strong>\u00a0OpenAI analyzes the request to find out the very best subsequent step. It seems to be by a listing of predefined actions to seek out the one which matches the person’s request. For instance, when a person requests a social publish, the system maps their enter to the next operate, guaranteeing that solely related actions are taken. The next motion,
marketing-getSocialPostSuggestion<\/code>, is executed when the person immediate is create a social media publish:<\/li>\n<\/ol>\n\n{<\/span><\/span>\n title:<\/span> \"marketing-getSocialPostSuggestion\"<\/span>,<\/span><\/span>\n description:<\/span> \"Decide what the person desires to create a social publish about. Don't generate a publish caption. Use this operate if the immediate incorporates something a couple of publish.\"<\/span>,<\/span><\/span>\n parameters:<\/span> {<\/span><\/span>\n sort:<\/span> \"object\"<\/span>,<\/span><\/span>\n required:<\/span> [<\/span>\"customerInput\"<\/span>]<\/span>,<\/span><\/span>\n properties:<\/span> {<\/span><\/span>\n customerInput:<\/span> {<\/span><\/span>\n sort:<\/span> \"string\"<\/span>,<\/span><\/span>\n description:<\/span> \"What the person desires to create a publish about\"<\/span>,<\/span><\/span>\n }<\/span>,<\/span><\/span>\n }<\/span>,<\/span><\/span>\n }<\/span>,<\/span><\/span>\n }<\/span><\/span><\/pre>\n<\/div>\n\n- Motion execution:<\/strong> The system performs the chosen motion. For
marketing-getSocialPostSuggestion<\/code>, it retrieves enterprise particulars and previous posts. Utilizing this info and the customerInput<\/code> that it decided from the person immediate, it generates a publish suggestion that matches the person’s model and tone of voice.\nRight here is an instance of the Open AI immediate we use to generate the publish caption:<\/p>\n
>You are an professional at selling small companies on-line. Your aim is to advertise a services or products in a manner that drives gross sales and represents your model. <\/code>
>Please create one artistic, interactive caption, key phrases associated to the caption, and a quick 5-10 phrase abstract of the theme of the publish beginning with 'Here is a publish'. Be aware the caption shouldn't begin with 'Here is a publish', solely the abstract ought to. <\/code>
>The social media publish is about: [USER_INPUT] that speaks to a profit or audience associated to your services or products. Don't embody the phrase 'Do you know'. Be sure there no references to the enterprise title. <\/code>
>You could reference (solely when wanted) this metadata concerning the enterprise: [BUSINESS_METADATA]. <\/code>
>Your caption must be concise (2 sentences) and have a [TONE] tone. <\/code>
>Your caption must be spectacular and longer than 10 phrases. Your key phrases CANNOT be too particular as a result of they are going to be used to floor extremely related outcomes from a design template database. Your key phrases must be ordered by descending order of relevance.<\/code>
>The phrase alternative and sentence construction must be diverse in comparison with these previous posts by the enterprise: [PAST_POSTS].<\/code><\/p>\n>You MUST use the JSON format offered and reply with legitimate JSON.<\/code><\/p>\n<\/li>\n- Response era:<\/strong>\u00a0After, all the info is collected, we create a JSON response which incorporates the person\u2019s enterprise information together with the generated publish caption and that is despatched again to the UI because the operate response.<\/li>\n
- UI replace:<\/strong>\u00a0As soon as the UI receives the response and the motion is accomplished, the system updates the UI to show the consequence\u2014a preview of a publish suggestion. The person can then overview, edit, or publish the publish.<\/li>\n<\/ol>\n
The method is a steady cycle: the person submits a immediate, which is mapped to a particular operate, after which the UI is up to date with the outcomes. If the person decides to make modifications or requests a special motion, the cycle begins once more, with the system analyzing the brand new immediate, working the corresponding operate, and updating the UI accordingly.<\/p>\n
The next flowchart illustrates how the Airo\u2122 Plus Advertising Advisor processes person requests, maps them to the suitable AI operate, and generates a social publish based mostly on enterprise information and person enter:<\/p>\n
\n![\"diagram\"](\"https:\/\/www.godaddy.com\/resources\/wp-content\/uploads\/2025\/03\/function_calling_flow_chart.png?size=3840x0\"\/)
<\/figure>\n<\/div>\nEngineering challenges<\/h2>\nWhen constructing the assistant, we bumped into some challenges. These included issues like non-deterministic prompts, response format variability, chat interface limitations, and decreasing cognitive load.<\/p>\n
Non-deterministic prompts<\/h3>\nAI prompts don\u2019t all the time observe a predictable sample, which made it difficult to make sure they persistently mapped to the best capabilities. We needed to tweak and refine our operate descriptions and OpenAI prompts a number of instances to get them working as anticipated. To ensure the mannequin aligned with the person interface, we added particular guidelines and pointers for the AI comparable to ensuring that the responses the AI gave have been restricted to GoDaddy data. It was a balancing act between permitting flexibility in how customers work together with the instrument and ensuring the output match the wants of the system.<\/p>\n
Within the earlier instance, when a person sorts a request to create a social publish, the AI assistant is anticipated to map it to the next operate description:<\/p>\n
>Decide what the person desires to create a social publish about. Don't generate a publish caption. Use this operate if the immediate incorporates something a couple of publish.<\/code><\/p>\nNevertheless, we encountered instances the place the operate description was too imprecise, main the assistant to misread the immediate. As an alternative of appropriately mapping to the operate, the assistant would fall again to a default ChatGPT response, leading to an inconsistent expertise\u2014 the person may obtain a generic AI-generated reply reasonably than progressing by the anticipated workflow.<\/p>\n
To handle this, we refined and fine-tuned operate descriptions to make sure extra exact mapping. By bettering the readability and specificity of descriptions, we considerably decreased instances the place the assistant defaulted to unintended responses, guaranteeing a extra dependable and seamless interplay movement.<\/p>\n
Response format variability<\/h3>\nWhen working with OpenAI\u2019s system, one problem we confronted was response format variability. Even when the immediate explicitly requested responses in JSON format, the outputs weren’t all the time structured persistently. This inconsistency meant that the info returned by OpenAI generally didn’t observe the anticipated format, making it tough to course of and show it correctly inside our interface.<\/p>\n
This variability in response codecs created vital points, because it might trigger errors within the person expertise or disrupt the workflow. For instance, if the response was lacking important fields, was incomplete, or didn’t match the anticipated construction, it might break the movement, stopping customers from seeing the specified outcomes or interacting with the interface as meant. To unravel this downside, we carried out a collection of strong error-handling mechanisms to make sure the system remained dependable.<\/p>\n
First, we added retry performance, permitting the person to resubmit their request robotically if the response didn\u2019t meet the required format. This ensured that customers wouldn\u2019t be left ready for a problem to be resolved manually and will proceed their work easily.<\/p>\n
Moreover, we launched validation checks that fastidiously inspected every response earlier than it was processed additional. These checks verified that the response was full, correctly formatted, and contained all the required information to proceed the workflow. If any points have been detected, the system would both try and retrieve a corrected response or immediate the person to re-submit the request.<\/p>\n
Collectively, these options decreased the influence of response format points. They ensured that customers might nonetheless depend on the system to carry out their duties with out disruption, even within the face of occasional inconsistencies within the format of OpenAI\u2019s responses.<\/p>\n
Restricted display actual property<\/h3>\nDesigning a cellular first user-friendly chat-based interface got here with distinctive UX challenges. Options like modifying and previewing social posts needed to be built-in right into a compact, conversational format. We centered on creating an intuitive design the place customers might make modifications with out being overwhelmed. Interactive components comparable to inline previews, quick-edit buttons, and guided prompts ensured the interface was each practical and visually accessible.<\/p>\n
Person understanding and UI steerage<\/h3>\n
- \n
- GoDaddy’s Airo Plus Advertising Advisor makes use of OpenAI’s function-calling system to create customized social media content material for companies.<\/li>\n
- The system overcomes challenges like non-deterministic prompts and response format variability by cautious immediate engineering and error dealing with.<\/li>\n
- Person interface design and clear steerage are essential for efficient AI-powered advertising instruments.<\/li>\n<\/ul>\n<\/div>\n<\/div>\n
In early 2023, GoDaddy started integrating AI into its advertising capabilities to make advertising duties simpler and extra accessible for customers. This consists of social media publish creation, electronic mail advertising campaigns, weblog content material era, and search engine optimisation.<\/p>\n
By incorporating AI, we offered customers with a place to begin for his or her advertising duties, producing customized content material and saving them time. Many customers confronted the “clean syndrome,” struggling to start out making a publish; so as a substitute of brainstorming manually, they now obtain tailor-made solutions, dashing up their workflow and constructing belief in our instruments.<\/p>\n
Nevertheless, we observed many customers needed a extra customized, interactive expertise\u2014much like working with a advertising company. This perception drove us to deal with enhancing our instruments, providing dynamic, conversation-driven assist resulting in the creation of the Airo\u2122 Plus Advertising Advisor.<\/p>\n
Background<\/h2>\n
The Airo\u2122 Plus Advertising Advisor is a instrument designed to information customers by numerous advertising actions with ease. It leverages the facility of AI to reply rapidly and precisely to person prompts. For instance, if a person desires to create a social media publish for his or her enterprise, Airo\u2122 Plus Advertising Advisor supplies tailor-made suggestions based mostly on the person’s enterprise such because the enterprise title, the web site, the enterprise class, and different related social information together with their previous Fb posts.<\/p>\n
To make this occur, we utilized OpenAI\u2019s\u00a0function-calling paradigm<\/a>. Perform calling permits any language mannequin, like GPT-4o-mini, to immediately work together with exterior APIs or methods in a structured manner. We created a set of capabilities mapped to particular actions, comparable to creating, previewing, modifying, and publishing a social publish. When a person supplies textual content enter, OpenAI maps the immediate to considered one of these predefined capabilities.<\/p>\n
For instance, if a person says, “I need to create a social publish about my new product: chocolate chip cookies<\/em>“, the assistant identifies the immediate and triggers a operate that calls an exterior API we created. This API first retrieves the client\u2019s enterprise info and social metadata. It then makes use of this information to generate an in depth OpenAI immediate that features the context\u2014particularly, that the publish is a couple of chocolate chip cookie product. The aim of this immediate is to provide a related and interesting social media caption for the enterprise. By combining AI capabilities with structured operate mapping, the assistant ensures an environment friendly, intuitive expertise tailor-made to the person’s wants.<\/p>\n
Within the following screenshot, the person has initiated a request to the Airo\u2122 Plus Advertising Advisor to generate a social publish about their chocolate cookies product:<\/p>\n
\n<\/figure>\n<\/div>\n At this stage, the assistant is processing the request, gathering enterprise information, and making ready a related publish.\u00a0Within the following screenshot, the AI-generated social publish is displayed, full with a instructed caption, picture, and scheduling particulars: <\/p>\n
\n<\/figure>\n<\/div>\nThe person can overview the publish, make edits if essential, or proceed with publishing.\u00a0<\/p>\n
The way it works<\/h2>\n
Airo\u2122 Plus Advertising Advisor makes use of OpenAI\u2019s function-calling system to assist create and handle social media content material. Right here\u2019s the way it works:<\/p>\n
- \n
- Person enter:<\/strong> The person enters a request, like “Create a social publish about cookies.<\/em>“<\/li>\n
- System understanding:<\/strong>\u00a0OpenAI analyzes the request to find out the very best subsequent step. It seems to be by a listing of predefined actions to seek out the one which matches the person’s request. For instance, when a person requests a social publish, the system maps their enter to the next operate, guaranteeing that solely related actions are taken. The next motion,
marketing-getSocialPostSuggestion<\/code>, is executed when the person immediate is create a social media publish:<\/li>\n<\/ol>\n\n{<\/span><\/span>\n title:<\/span> \"marketing-getSocialPostSuggestion\"<\/span>,<\/span><\/span>\n description:<\/span> \"Decide what the person desires to create a social publish about. Don't generate a publish caption. Use this operate if the immediate incorporates something a couple of publish.\"<\/span>,<\/span><\/span>\n parameters:<\/span> {<\/span><\/span>\n sort:<\/span> \"object\"<\/span>,<\/span><\/span>\n required:<\/span> [<\/span>\"customerInput\"<\/span>]<\/span>,<\/span><\/span>\n properties:<\/span> {<\/span><\/span>\n customerInput:<\/span> {<\/span><\/span>\n sort:<\/span> \"string\"<\/span>,<\/span><\/span>\n description:<\/span> \"What the person desires to create a publish about\"<\/span>,<\/span><\/span>\n }<\/span>,<\/span><\/span>\n }<\/span>,<\/span><\/span>\n }<\/span>,<\/span><\/span>\n }<\/span><\/span><\/pre>\n<\/div>\n- \n
- Motion execution:<\/strong> The system performs the chosen motion. For
marketing-getSocialPostSuggestion<\/code>, it retrieves enterprise particulars and previous posts. Utilizing this info and thecustomerInput<\/code> that it decided from the person immediate, it generates a publish suggestion that matches the person’s model and tone of voice.\nRight here is an instance of the Open AI immediate we use to generate the publish caption:<\/p>\n
>
You are an professional at selling small companies on-line. Your aim is to advertise a services or products in a manner that drives gross sales and represents your model. <\/code>
>Please create one artistic, interactive caption, key phrases associated to the caption, and a quick 5-10 phrase abstract of the theme of the publish beginning with 'Here is a publish'. Be aware the caption shouldn't begin with 'Here is a publish', solely the abstract ought to. <\/code>
>The social media publish is about: [USER_INPUT] that speaks to a profit or audience associated to your services or products. Don't embody the phrase 'Do you know'. Be sure there no references to the enterprise title. <\/code>
>You could reference (solely when wanted) this metadata concerning the enterprise: [BUSINESS_METADATA]. <\/code>
>Your caption must be concise (2 sentences) and have a [TONE] tone. <\/code>>Your caption must be spectacular and longer than 10 phrases. Your key phrases CANNOT be too particular as a result of they are going to be used to floor extremely related outcomes from a design template database. Your key phrases must be ordered by descending order of relevance.<\/code>>The phrase alternative and sentence construction must be diverse in comparison with these previous posts by the enterprise: [PAST_POSTS].<\/code><\/p>\n>You MUST use the JSON format offered and reply with legitimate JSON.<\/code><\/p>\n<\/li>\n- Response era:<\/strong>\u00a0After, all the info is collected, we create a JSON response which incorporates the person\u2019s enterprise information together with the generated publish caption and that is despatched again to the UI because the operate response.<\/li>\n
- UI replace:<\/strong>\u00a0As soon as the UI receives the response and the motion is accomplished, the system updates the UI to show the consequence\u2014a preview of a publish suggestion. The person can then overview, edit, or publish the publish.<\/li>\n<\/ol>\n
The method is a steady cycle: the person submits a immediate, which is mapped to a particular operate, after which the UI is up to date with the outcomes. If the person decides to make modifications or requests a special motion, the cycle begins once more, with the system analyzing the brand new immediate, working the corresponding operate, and updating the UI accordingly.<\/p>\n
The next flowchart illustrates how the Airo\u2122 Plus Advertising Advisor processes person requests, maps them to the suitable AI operate, and generates a social publish based mostly on enterprise information and person enter:<\/p>\n
\n<\/figure>\n<\/div>\nEngineering challenges<\/h2>\n
When constructing the assistant, we bumped into some challenges. These included issues like non-deterministic prompts, response format variability, chat interface limitations, and decreasing cognitive load.<\/p>\n
Non-deterministic prompts<\/h3>\n
AI prompts don\u2019t all the time observe a predictable sample, which made it difficult to make sure they persistently mapped to the best capabilities. We needed to tweak and refine our operate descriptions and OpenAI prompts a number of instances to get them working as anticipated. To ensure the mannequin aligned with the person interface, we added particular guidelines and pointers for the AI comparable to ensuring that the responses the AI gave have been restricted to GoDaddy data. It was a balancing act between permitting flexibility in how customers work together with the instrument and ensuring the output match the wants of the system.<\/p>\n
Within the earlier instance, when a person sorts a request to create a social publish, the AI assistant is anticipated to map it to the next operate description:<\/p>\n
>Decide what the person desires to create a social publish about. Don't generate a publish caption. Use this operate if the immediate incorporates something a couple of publish.<\/code><\/p>\nNevertheless, we encountered instances the place the operate description was too imprecise, main the assistant to misread the immediate. As an alternative of appropriately mapping to the operate, the assistant would fall again to a default ChatGPT response, leading to an inconsistent expertise\u2014 the person may obtain a generic AI-generated reply reasonably than progressing by the anticipated workflow.<\/p>\n
To handle this, we refined and fine-tuned operate descriptions to make sure extra exact mapping. By bettering the readability and specificity of descriptions, we considerably decreased instances the place the assistant defaulted to unintended responses, guaranteeing a extra dependable and seamless interplay movement.<\/p>\n
Response format variability<\/h3>\n
When working with OpenAI\u2019s system, one problem we confronted was response format variability. Even when the immediate explicitly requested responses in JSON format, the outputs weren’t all the time structured persistently. This inconsistency meant that the info returned by OpenAI generally didn’t observe the anticipated format, making it tough to course of and show it correctly inside our interface.<\/p>\n
This variability in response codecs created vital points, because it might trigger errors within the person expertise or disrupt the workflow. For instance, if the response was lacking important fields, was incomplete, or didn’t match the anticipated construction, it might break the movement, stopping customers from seeing the specified outcomes or interacting with the interface as meant. To unravel this downside, we carried out a collection of strong error-handling mechanisms to make sure the system remained dependable.<\/p>\n
First, we added retry performance, permitting the person to resubmit their request robotically if the response didn\u2019t meet the required format. This ensured that customers wouldn\u2019t be left ready for a problem to be resolved manually and will proceed their work easily.<\/p>\n
Moreover, we launched validation checks that fastidiously inspected every response earlier than it was processed additional. These checks verified that the response was full, correctly formatted, and contained all the required information to proceed the workflow. If any points have been detected, the system would both try and retrieve a corrected response or immediate the person to re-submit the request.<\/p>\n
Collectively, these options decreased the influence of response format points. They ensured that customers might nonetheless depend on the system to carry out their duties with out disruption, even within the face of occasional inconsistencies within the format of OpenAI\u2019s responses.<\/p>\n
Restricted display actual property<\/h3>\n
Designing a cellular first user-friendly chat-based interface got here with distinctive UX challenges. Options like modifying and previewing social posts needed to be built-in right into a compact, conversational format. We centered on creating an intuitive design the place customers might make modifications with out being overwhelmed. Interactive components comparable to inline previews, quick-edit buttons, and guided prompts ensured the interface was each practical and visually accessible.<\/p>\n
Person understanding and UI steerage<\/h3>\n
- Response era:<\/strong>\u00a0After, all the info is collected, we create a JSON response which incorporates the person\u2019s enterprise information together with the generated publish caption and that is despatched again to the UI because the operate response.<\/li>\n
- System understanding:<\/strong>\u00a0OpenAI analyzes the request to find out the very best subsequent step. It seems to be by a listing of predefined actions to seek out the one which matches the person’s request. For instance, when a person requests a social publish, the system maps their enter to the next operate, guaranteeing that solely related actions are taken. The next motion,
- Person enter:<\/strong> The person enters a request, like “Create a social publish about cookies.<\/em>“<\/li>\n