Google Feedgen
Google has released an open-source tool called FeedGen that uses Large Language Models (LLMs) to generate optimized Shopping Ads titles and descriptions, and fill missing attributes in product feeds. It helps merchants and advertisers surface and fix quality issues in their feeds using Generative AI in a simple and configurable way.
Feed Quality
Ensuring high feed quality is extremely important when running Google Shopping Ads. Google only recently shared how it weighs certain feed attributes in its algorithm. This could help you prioritize which things to optimize. One of the most important part of getting a more prominent spot is strong product titles.
Hallucination effect
Anyone who has worked regularly with ChatGPT and similar systems is likely aware of the hallucination effect. The hallucination effect is a phenomenon where the model generates text that is incorrect, nonsensical, or not real. This is something you would like to avoid with all means possible as this may give a false impression of the product to your (potential) customers. Google Feedgen tries to avoid this by providing a score for generated titles between -1 and 1 that acts as a quality indicator. Positive scores indicate varying degrees of good quality, while negative scores represent uncertainty over the generated content. If new words were added the Model will give the output a low score so you can use this score to prevent this output from being used.
My two cents
Google FeedGen has an amazing potential to quickly improve your product titles and descriptions in bulk. Some of the key benefits are imo:
– Great way to improve your product feed quality especially when you have to rely on third party input.
– Very fast, it goes through hundreds of products in minutes;
– It looks after the hallucination effect to ensure it does not add made up values to your titels and descriptions.
– There is an easy feature that allows you to select the new titles and descriptions that you like to include.
– When you selected the new titles and descriptions you like you can automatically convert it to a supplemental feed. This can then be added in your GMC.
What can be improved in my opinion:
– It’s not a plug a play solution. You need to create a Google Cloud Project and enable the Vertex API. I would recommend to watch Vertex AI for ML operations videos to get you started.
– The output quality highly depends on the quality of your prompts. For high quality output you must be skilled in setting the right prompting examples.
– You need to set prompting examples for each product category which can take up quite some time when first using the feature.
– To ensure high quality output you still have to go through all the products.
– Only English Language is supported for now. You can contact your Google rep if you like to apply it to another language. I have no experience with this as of yet.
