# AI Overviews Leak: Does Google Know The Brands It Wants To Recommend?

3 min read

OK so VERY interesting discovery here from Stéphane Le Mentec. When doing some searching, he found what appears to be a leak in the internal thinking logic of AI Overviews. I'm going to paste the screenshots here:

![](https://firebasestorage.googleapis.com/v0/b/nectiv-website.firebasestorage.app/o/public%2Fmedia%2F1795010957773-1794888677487.jpeg?alt=media)![](https://firebasestorage.googleapis.com/v0/b/nectiv-website.firebasestorage.app/o/public%2Fmedia%2F1795010926861-1794888677303.jpeg?alt=media)![](https://firebasestorage.googleapis.com/v0/b/nectiv-website.firebasestorage.app/o/public%2Fmedia%2F1795010916121-1794888677448.jpeg?alt=media)

Now normally I'm skeptical of these things. However, this one seems legitimate. The reason being is that Dan Petrovic [wrote about AI Mode tags in September where he explored the underlying functions](https://dejan.ai/blog/ai-mode-tags/). Please go give it a read if you haven't already. Well the <FollowUp> tag in here EXACTLY matches the one he talked about. Which means it seems to be pulled from the systems. As well, the InlineEntity tags are strikingly similar to what he described.

So I'm going to jot down some of my quick takeaways as I think it's bigger than just a LinkedIn post can hold.

1.  The systems FIRST analyzes constraints for the response. So before even searching, it provides guardrails on what information can be surfaced
    
2.  It pulls a list of entities/brands before this process even begins. Suganthan Mohanadasan [wrote about how ChatGPT knows the brands it's going to include](https://suganthan.com/blog/chatgpt-decides-before-it-searches/) before it even responds. It appears AI Overviews is at least starting the process with something similar. While this screenshot doesn't include the search queries, if you analyze AI Mode/Gemini fan out queries, they often include direct brand searches.
    
3.  In the constraints, it even looks for Reddit specifically. It says "and Reddit if applicable" in this response. So that indicates there might be a direct function of the retrieval pipeline that sources Reddit for content.
    
4.  Finally in the constraints, it suggests a follow up question, via the <FollowUp> tag. This is obviously to keep us in Google's ecosystem
    
5.  The next step in synthesizing the data from search results. Using the retrieved pages, it then positions each entity (“focuses on privacy compliance without the heavy GRC layer”).
    
6.  During this process, it also seemed to introduce new brands into the pipeline that weren't already there. The initial entities started at a list of 8, this expanded to 13.
    
7.  It uses outside trust signals: It specifically looked up Gartner performance + consensus on Reddit for some of the brands.
    
8.  Finally it drafts the response. It cuts the entities/brands from the list of 13 to only 5. All of the 5 it landed on (in this response) were in the original entities considered.
    

So based on this small window, there are some definite considerations for marketers:

-   Similar to ChatGPT, AI Overviews might already arrive with pre-conceived shortlist about brands it wants to include. This makes sense why companies in listicles get cited but not mentioned. They are not in the shortlist that AI Overviews provides. These screenshots don't confirm that but this + analyzing a lot of fan-out query data makes me really think it's the case.
    
-   New brands CAN be introduced via search retrieval. We would need a lot more data to conclude about whether it is an uphill battle or not
    
-   Off-site matters. AI Overviews specifically looks for Reddit. It also pulled in Gartner. I suspect it might to the same with YouTube.
    

Obviously this is just one result. But I'm fairly certain that it provides us with a glimpse of how AI Overviews works