AI-Powered Social Media Posts: Enhancing Marketing with AI-Powered Content Creation

In early 2023, GoDaddy began integrating AI into its marketing capabilities to make marketing tasks easier and more accessible for users. This includes social media post creation, email marketing campaigns, blog content generation, and SEO.

By incorporating AI, we provided users with a starting point for their marketing tasks, generating personalized content and saving them time. Many users faced the “blank syndrome,” struggling to start creating a post; so instead of brainstorming manually, they now receive tailored suggestions, speeding up their workflow and building trust in our tools.

However, we noticed many users wanted a more personalized, interactive experience—similar to working with a marketing agency. This insight drove us to focus on enhancing our tools, offering dynamic, conversation-driven support leading to the creation of the Airo™ Plus Marketing Consultant.

Background

The Airo™ Plus Marketing Consultant is a tool designed to guide users through various marketing activities with ease. It leverages the power of AI to respond quickly and accurately to user prompts. For example, if a user wants to create a social media post for their business, Airo™ Plus Marketing Consultant provides tailored recommendations based on the user’s business such as the business name, the website, the business category, and other relevant social data including their past Facebook posts.

To make this happen, we utilized OpenAI’s function-calling paradigm. Function calling enables any language model, like GPT-4o-mini, to directly interact with external APIs or systems in a structured way. We created a set of functions mapped to specific actions, such as creating, previewing, editing, and publishing a social post. When a user provides text input, OpenAI maps the prompt to one of these predefined functions.

For example, if a user says, “I want to create a social post about my new product: chocolate chip cookies“, the assistant identifies the prompt and triggers a function that calls an external API we created. This API first retrieves the customer’s business information and social metadata. It then uses this data to generate a detailed OpenAI prompt that includes the context—namely, that the post is about a chocolate chip cookie product. The goal of this prompt is to produce a relevant and engaging social media caption for the business. By combining AI capabilities with structured function mapping, the assistant ensures an efficient, intuitive experience tailored to the user’s needs.

In the following screenshot, the user has initiated a request to the Airo™ Plus Marketing Consultant to generate a social post about their chocolate cookies product:

At this stage, the assistant is processing the request, gathering business data, and preparing a relevant post. In the following screenshot, the AI-generated social post is displayed, complete with a suggested caption, image, and scheduling details:

The user can review the post, make edits if necessary, or proceed with publishing. 

How it works

Airo™ Plus Marketing Consultant uses OpenAI’s function-calling system to help create and manage social media content. Here’s how it works:

1. User input: The user enters a request, like “Create a social post about cookies.“
2. System understanding: OpenAI analyzes the request to determine the best next step. It looks through a list of predefined actions to find the one that matches the user’s request. For example, when a user requests a social post, the system maps their input to the following function, ensuring that only relevant actions are taken. The following action, marketing-getSocialPostSuggestion, is executed when the user prompt is create a social media post:

hanya di

3. Action execution: The system performs the chosen action. For marketing-getSocialPostSuggestion, it retrieves business details and past posts. Using this information and the customerInput that it determined from the user prompt, it generates a post suggestion that matches the user’s brand and tone of voice.

   Here is an example of the Open AI prompt we use to generate the post caption:

   >You're an expert at promoting small businesses online. Your goal is to promote a product or service in a way that drives sales and represents your brand.
   >Please create one creative, interactive caption, keywords related to the caption, and a brief 5-10 word summary of the theme of the post starting with 'Here's a post'. Note the caption should not start with 'Here's a post', only the summary should.
   >The social media post is about: [USER_INPUT] that speaks to a benefit or target audience related to your products or services. Do not include the phrase 'Did you know'. Make sure there no references to the business name.
   >You may reference (only when needed) this metadata about the business: [BUSINESS_METADATA].
   >Your caption should be concise (2 sentences) and have a [TONE] tone.
>Your caption should be impressive and longer than 10 words. Your keywords CANNOT be too specific because they will be used to surface highly relevant results from a design template database. Your keywords should be ordered by descending order of relevance.
>The word choice and sentence structure should be varied compared to these past posts by the business: [PAST_POSTS].

   >You MUST use the JSON format provided and respond with valid JSON.

4. Response generation: After, all the data is collected, we create a JSON response which includes the user’s business data along with the generated post caption and this is sent back to the UI as the function response.
5. UI update: Once the UI receives the response and the action is completed, the system updates the UI to display the result—a preview of a post suggestion. The user can then review, edit, or publish the post.

The process is a continuous cycle: the user submits a prompt, which is mapped to a specific function, and then the UI is updated with the results. If the user decides to make changes or requests a different action, the cycle starts again, with the system analyzing the new prompt, running the corresponding function, and updating the UI accordingly.

The following flowchart illustrates how the Airo™ Plus Marketing Consultant processes user requests, maps them to the appropriate AI function, and generates a social post based on business data and user input:

Engineering challenges

When building the assistant, we ran into some challenges. These included things like non-deterministic prompts, response format variability, chat interface limitations, and reducing cognitive load.

Non-deterministic prompts

AI prompts don’t always follow a predictable pattern, which made it tricky to ensure they consistently mapped to the right functions. We had to tweak and refine our function descriptions and OpenAI prompts multiple times to get them working as expected. To make sure the model aligned with the user interface, we added specific rules and guidelines for the AI such as making sure that the responses the AI gave were restricted to GoDaddy knowledge. It was a balancing act between allowing flexibility in how users interact with the tool and making sure the output fit the needs of the system.

In the previous example, when a user types a request to create a social post, the AI assistant is expected to map it to the following function description:

>Determine what the user wants to create a social post about. Do not generate a post caption. Use this function if the prompt contains anything about a post.

However, we encountered cases where the function description was too vague, leading the assistant to misinterpret the prompt. Instead of correctly mapping to the function, the assistant would fall back to a default ChatGPT response, resulting in an inconsistent experience— the user might receive a generic AI-generated reply rather than progressing through the expected workflow.

To address this, we refined and fine-tuned function descriptions to ensure more precise mapping. By improving the clarity and specificity of descriptions, we significantly reduced cases where the assistant defaulted to unintended responses, ensuring a more reliable and seamless interaction flow.

Response format variability

When working with OpenAI’s system, one challenge we faced was response format variability. Even when the prompt explicitly requested responses in JSON format, the outputs were not always structured consistently. This inconsistency meant that the data returned by OpenAI sometimes did not follow the expected format, making it difficult to process and display it properly within our interface.

This variability in response formats created significant issues, as it could cause errors in the user experience or disrupt the workflow. For example, if the response was missing essential fields, was incomplete, or did not match the expected structure, it would break the flow, preventing users from seeing the desired results or interacting with the interface as intended. To solve this problem, we implemented a series of robust error-handling mechanisms to ensure the system remained reliable.

First, we added retry functionality, allowing the user to resubmit their request automatically if the response didn’t meet the required format. This ensured that users wouldn’t be left waiting for an issue to be resolved manually and could continue their work smoothly.

Additionally, we introduced validation checks that carefully inspected each response before it was processed further. These checks verified that the response was complete, properly formatted, and contained all the necessary data to continue the workflow. If any issues were detected, the system would either attempt to retrieve a corrected response or prompt the user to re-submit the request.

Together, these solutions reduced the impact of response format issues. They ensured that users could still rely on the system to perform their tasks without disruption, even in the face of occasional inconsistencies in the format of OpenAI’s responses.

Limited screen real estate

Designing a mobile first user-friendly chat-based interface came with unique UX challenges. Features like editing and previewing social posts had to be integrated into a compact, conversational layout. We focused on creating an intuitive design where users could make modifications without being overwhelmed. Interactive elements such as inline previews, quick-edit buttons, and guided prompts ensured the interface was both functional and visually accessible.

User understanding and UI guidance

Some users did not understand what the assistant could do. Since the UI looked like a typical chatbot, they entered random prompts unrelated to the supported marketing tasks, leading to confusion when they received unexpected responses. The UI also didn’t clearly indicate what actions were available, and there was no proper feedback for unsupported inputs. To address this, we plan to improve the UI by moving the assistant into a modal or panel to make its functionality clearer, restricting free-form input to predefined actions, and adding inline validation so users get immediate feedback if their request isn’t supported. For example, the assistant could respond with: “I currently support social post creation, editing, and publishing. Please select one of these actions.” We are also exploring ways to provide suggested actions upfront so users know what they can do. These improvements will help create a more structured, intuitive experience, reducing confusion and guiding users through their marketing tasks more effectively.

Conclusion

Building the Airo™ Plus Marketing Consultant required solving complex engineering challenges, including function mapping accuracy, OpenAI response variability, UI constraints, and user understanding. Ensuring the assistant correctly mapped user prompts to predefined functions required multiple iterations of refining function descriptions and handling edge cases. Additionally, working with OpenAI’s function calling posed challenges in maintaining structured responses, enforcing JSON formatting, and handling inconsistencies in model outputs.

Beyond backend accuracy, we encountered front-end UX challenges, particularly around guiding users within the assistant interface. Some users entered unsupported prompts, expecting a general-purpose chatbot, which led to unexpected responses. To improve this, we are iterating on UI enhancements such as structured workflows, predefined actions, inline validation, and restricted input options to ensure users interact within the assistant’s defined capabilities.

Looking ahead, we plan to refine prompt engineering, optimize system reliability, and explore more deterministic AI responses. Additionally, we aim to expand capabilities, including content performance analysis, and improved function disambiguation to create a more predictable, developer-friendly experience. As AI continues evolving, so will our approach to integrating it efficiently into structured workflows.