genAI: Updating Search Engine Metadata for Pennies

SEO Title & Meta Description: Automated with genAI

Building on the success of our two-stage content processing workflow, this article demonstrates how generative AI can optimize SEO metadata at scale in ≤ 1 hour.

Search engine optimization (SEO) is constantly moving target. While AI based search recommendations have come a long way since 2024, when Google was recommending bizarre things… such as smoking and eating rocks:

Nevertheless, the AI recommendations remain ineffective. Some businesses suggest the AI responses provide hallucinated information.

With that being said, it is important to prepare for the ever changing landscape, especially when it comes to refined SEO. For example, Google’s recent updates targets ranking based on “helpfulness” rather than sheer ranking. Google is attempting to also identify and reduce AI-generated spam. So websites must maintain webpages that are helpful and concise in what they are providing. In other words, quality over quantity. So SEO must be refined with clarity and helpfulness in mind.

The Challenge

With metadata in-hand for a client’s products, I identified a critical optimization opportunity to enhance SEO titles and meta descriptions. The product pages had multiple issues which likely impacted 1) ranking and 2) click-through rate (CTR):

1. Outdating formatting styles
2. Too short or too long character titles/descriptions
3. Outdate shipping information.

Manual creation would require significant time investment, while poor SEO metadata directly impacts search visibility, CTR and helpfulness. For example, some product pages contained text in the title “Shop | Features | Reviews”, even though the pages didn’t always contain reviews or often lack information about features.

SEO Generation Goals:

Character Count Optimization: There are character limits to what is visible on mobile/desktop in search results.

• Titles: 40-60 characters with brand suffix “| NAME.com”
• Meta descriptions: 140-160 characters

For example, the below title has 59 characters and is visible in a mock search provided by Yoast SEO:

However, when we add 15-20 more characters, the latter portion of the title becomes cut-off. So if my intent was for that text to be visible during the initial impression, it would not be.

Content Quality: The quality of the information should be 1) informative 2) helpful to set you apart from competitors in the search and 3) provide some immediate value. In other words, each page should:

• Highlight key product benefits and features
• Incorporate relevant keywords from the product description/product manual
• Include value propositions (e.g., free shipping on $###+ orders)

Brand Consistency: The information should be relatively consistent for the business  across all products or product categories.

Implementation: Single-Stage SEO Generation

For this solution, I used the product title and HTML content from our previous workflow to prompt a genAI model to optimize the SEO metadata. The goal: include helpful and succinct information.

Model Choice: GPT-4o-mini (cost-effective for converting long content >> short, structured content)

seo_gpt_result = modify_seo_with_gpt(
    prompt=prompt_user, 
    sys_prompt=prompt_system, 
    model="gpt-4o-mini", 
    max_tokens=4000, 
    temp=.4
)

Results:

The genAI model was successfully deployed for summarizing each product’s content into titles and descriptions for 1175 products at approximately $0.47 cost in API. The processing time was less than 1 hour, which would have taken over 100 manual labor hours.

Quality Metrics

I evaluated the character count distributions for the titles and descriptions:

• SEO Titles: Old mean 72.9 chars → New mean 61.8 chars
• Meta Descriptions: Old mean 140 chars → New mean 160 chars

Importantly, the distribution of the character lengths were narrower, or less variable, between product descriptions. While the title was slightly above the target window (green, left pnael), the characters that exceeded it were the suffix including ” | <website.com> “. So loss is minimized. Importantly, the distribution was 1) within the target window for the description (green, right panel) and 2) included updated and consistent high value information.

Coverage Improvements: When I started, there was also missing information for titles and meta descriptions. The current solution reduces these missing values to 0%.

Implementation Timeline

Phase 1 (Day 1): Setup and Testing

• Prompt engineering for consistent JSON output
• 50-product test batch for baseline evaluation
• Character count optimization fine-tuning

Phase 2 (Day 2): Full Processing

• Automated processing of all products
• Quality validation and error handling
• Data export preparation for client implementation

Business Impact

Long-term Options

In this case, a universal optimization approach was used. In future iterations, it is possible to perform A/B testing across optimization themes. Across the clients’ multiple businesses and thousands of products, they can consider unique themes across product categories and/or adding different points targeting value at the start or end of the description.

With this, one can evaluate the performance over time across KPIs, such as:

• Changes in CTR
• Performance across mobile / desktop
• Increase in organic traffic
• Increase in sales and/or conversions.

Key Results

What Worked:

• Structured JSON output ensured consistent formatting and easy implementation
• Lower temperature (.4) provided reliable, focused results without creativity issues (avoid risk of making up “spam“)
• Scalable to future product additions and A/B testing across products, blogs, pages, etc.

Challenges:

• Prompt engineering required careful thought about “value” items.
• Character count targeting needed fine-tuning to stay close to target window.

Next Steps:

1. Monitor organic search performance over 3-6 months
2. A/B testing of AI-generated vs manually crafted SEO metadata
3. Integration with ongoing content update workflows
4. Extension to category and collection pages using similar methodology

Conclusion

This genAI powered SEO solution delivered exactly what the client requested. It minimized costs and reduced a slow, drawn-out process which is common with in-house, manual solutions. This single stage and the previous multi-stage solution illustrates the value of genAI in business. Focused implementations can deliver outsized returns.