FeedGen: Complete product feed revision in a few clicks

What is FeedGen?

FeedGen is an open-source tool that uses Google Cloud ‘s Large Language Models to optimize product names, generate more detailed descriptions, and fill in missing attributes. So if you’re missing some information in your product feeds or your product names aren’t attractive enough, FeedGen will fix it efficiently and save you a lot of time. Sophisticated and relevant titles, detailed descriptions and complete attributes not only increase the chances of products showing up for the right queries, but also improve click-through rates (CTR) and thus overall campaign performance.

1. Settings

First, make a copy of the configuration spreadsheet. Then you upload the product data you want to improve or add to the “Input Feed” tab.

2. Configuration

In the configuration spreadsheet you set:
✅ Whether you want to use Landing Page or Image Understanding.
✅ What language model will you use.
✅ In which language the data should be generated.
✅ Check the estimated processing cost.
✅ Enter your Google Cloud ID.

The key section is “Few-Shot Prompting Examples”, where you define examples of what the output data should look like. Make sure you pay attention to this part – it will significantly affect the quality of the results.

3. Starting up

Once your configuration is complete, go to the “Getting Started” section and click on “Initialise”. This will grant the necessary permissions. Then open the FeedGen sidebar in the top menu to start the process:

1️⃣ “Clean Generated Data” – clean previous outputs
2️⃣ “Generate” – start generating new data

4. Control and export

The generated data is displayed in the “Generated Content Validation” tab, where it is necessary to perform a manual check. Once approved, click “Approve Filtered” and then “Export to Output Feed”. This will send the products to the “Output Feed” tab for final export to your feed.

You can get data into the feed, for example, by using the Supplemental Feed directly in Google Merchant Center or by importing it into Mergada or another feed editor.

Remember that you should use structured_title and structured_description elements in the feed to import AI generated data.

Benefits ✅

✔ Filling in missing data from images and landing pages – the tool can analyse product images and landing page content to fill in missing information such as materials, colours or other specifications.

✔ Customisation to specific requirements – flexibility in adjusting the output to suit your needs.

✔ Easy integration with Google Spreadsheet – no complicated implementation, everything runs directly in spreadsheets.

✔ Increased relevance of product names and descriptions – better optimization for search engines and advertising.

Disadvantages ❌

✖ Manual checking of outputs is necessary – although the AI generates data automatically, it still needs to verify its accuracy.

✖ Higher costs for large feeds – with a large volume of products, the cost of generation can be quite high. The final price is also significantly influenced by the chosen language model. For example, generating a feed with 500 products costs only $1.33 using Gemini 1.5 Flash. When using Gemini 1.5 Pro, the price for one such generation goes up to $13.32. For a larger feed with 10,000 products, the cost can reach up to $266 per generation.

✖ Error rate – the model may not always understand the context correctly or produce the ideal output.

✖ Dependency on Google Cloud – proper functioning requires access to Google’s Large Language Models, which means you need a Google Cloud account and API keys.

Our tips

• The price is significantly influenced by the language model chosen. In our experience, Gemini 1.5 Flash offers the best price/performance ratio.
  
• Be careful with using landing pages to generate data. FeedGen can pull irrelevant information from the landing page, such as promotions or promotional texts, which it then inserts directly into the product name.
  
• Take extra care to fill in the “Few-Shot Prompting Examples” section. It is this part that fundamentally influences the quality of the results. Among other things, you can set not only the order of attributes, but also their formatting – for example, the font size.
  
• Try FeedGen on foreign accounts. The quality of output in Czech and Slovak is often lower than in English or German.
  
• Don’t forget to set a financial limit in your Google Cloud project. This ensures that you don’t spend more on FeedGen than you originally planned.

Conclusion

FeedGen will undoubtedly find its place and can greatly facilitate the optimization of product data, which will have a positive impact on the results. However, it has its limitations that make it ideal for smaller and medium-sized projects where we work with a small volume of data and there is no possibility to add relevant parameters to the XML feed directly via programmers. For larger projects, where it is possible to send all the necessary parameters directly to the feed, it is better to use specialized tools for XML feed management – for example Mergado.

Bottom line, we see huge potential in feedgen, it is just important to consider when its use makes real sense and when it is better to take a different approach. Artificial intelligence is unavoidable, so let’s at least make the most of it. And how do you deal with product feed optimization? Don’t you want to save some time with FeedGen too?