You know that I love talking about the vector search. I had a good success for the vector search. This success often does not stop. I think I recognized why and if this thought is true, it can turn out that it may be more harmful than benefits if this thought applies that a great job could possibly do more damage when optimizing the vector search.
What is vector search?
Vector search uses mathematics to convert words and phrases in a way into numbers in which the numbers represent several aspects or dimensions of the word. If these numbers are embedded in a multi -dimensional space, concepts that are similar in this area appear close together.
Here is an excellent video too Explain the vector search even further.
https://www.youtube.com/watch?v=YVOBVU1L7GI
Rankbrain and Rankembed Bert -Use vector -embedding
Google’s AI systems such as Rankbrain and Rankembed Bert use vector codes and vector search. If a query is searched, the query (and probably probably extensions to the query) is embedded in a vectors and the content that is embedded nearby in this room is probably regarded as relevant. (Here is more of A 2016 Search engine country items written by Danny Sullivan It says that Google tells us that we can learn more about Rankbrain Convert words to vectors.))
In the Doj against Google TestWe have learned that Rangbrain can be used to reach the top 20-30 results again.
Rankbrain has probably developed enormously since Google told us about the conversion of words into vectors for the first time. However, there is no doubt that vectors and vectors are important. In fact, Google expressly tells us that Vector search does Google Search.
For this reason, we see so many advice from SEOS to write as it looks good for the vector search. This will help our content look more relevant to machines. At least initially …
Google says that writing for vector search does not help us run better
A few years ago I had a fascinating conversation with Google’s Gary Illyes about the vector search. He told me it was something interesting that you could learn about, but if I know about the search for vector, I would not give an advantage in the search.
I didn’t believe him. (Sorry, Gary! I had to test things for myself before I came to the realization that you are right.)
I developed a theory that I wrote about in my book. The concept contained the expansion of a query to write related micro-inside and then text with well-structured sentences that answered each of these microintents clearly and precisely. The idea was that this technique should lead to the content of the content in the vector room of Google in a place that is close to both the query and the extension of the query. We would therefore look more relevant in the eyes of the vector search.
I had good success times:
In some cases we have retained these ranking lists. In some cases we won them for a short time and then lost all of our profits.
As Google determines the relevance, continues to develop
I now understand why Gary encouraged me not to concentrate too much on the vector search. I think we can learn a lot by understanding Google’s different methods to predict what content a user is relevant. However, I think that it may be more harm than good damage if it can harm extensively to look good for Google’s vector search systems.
Before I explain why, it is worth noting that Google has systems that are far more sophisticated than my simple explanation for the vector search. For example, at the end of 2024 they wrote a paper in which a technology was described where They gave an LLM a content and asked them to create a relevant query On the content and another question that seem Relevant, but really not. Then use these couples to train a system to better understand the nuances of the relevance. I wrote a lot more about my thoughts about how Google determines relevance and helpfulness.
However, we come back why it is the wrong way to focus mainly on optimizing the search for vector in order to go down.
If we write for the vector search, we look good for machines. But what if users don’t agree that we are the most relevant and helpful choice?
Google predicts what the viewfinders click on and then refined these predictions
The DOJ VS Google test version has shared a wealth of information about how Google learns from the one that people deal with.
(Documentation from the doj vs Google test)
These functions improved even more with the development of the helpful content system and finally with the start of the core update in March 2024, in which Google learned to use a variety of signals to determine which content is really satisfied.
There are Many articles that have been written in Google’s Clickness models This attempt to predict what people click on and with whom he will be satisfied. Here is a Google paper on Include clicks, attention and satisfaction in a model that learns from the actions of seekers.
Google’s systems work to present users with results that they believe will be helpful. Then they learn from what people actually found helpful And machine learning systems use this information to improve your predictions.
Can you see where I go with it?
I think it is possible if you look consistently good for vector search systems and select Google as the most likely relevant result. However, user actions do not confirm that they are actually actually was The best result, then the systems will adapt to find out that their content is not as good as the system originally predicts.
We could do more harm than benefit.
Should we ignore the vector search?
I still think that it is a good idea to understand the questions that your audience has on a topic and to do a good job to answer them in a way that is easy to digest and present for machines.
However, I think that understanding should usually come from Really to understand your audience instead of using the use of tools. That is why experience is so important. I think you can be inspired by tools that people also tell the results, and even more inspiration from the question of an LLM, which questions a user probably has on a topic. Ultimately, your original, helpful value results from your unique insight and know what your audience is really helpful. If we write in this way, we look good for viewfinder that searches the original and really informative information and at the same time optimize the search for vector. If we only rely on the tools, only the machines that love our content love it.
Here are some from Google’s bullet points Create helpful content that I am relevant for this discussion. These are not specific things that try to measure Google’s search systems, but the types of things that tend to use users:
- Does the content offer? Original information, reporting, research or analysis?
- Does the content offer? Inexpensive analysis or Interesting information That means Beyond the obvious?
- When the content decreases to other sources Offer considerable additional value and originality?
- Does the content offer? Significant value compared to other pages In search results?
- Is this content written or checked by an expert or enthusiast who has been shown to know the topic well?
- Do you have an existing or intended audience for your company or your website, which feels the content useful if you come directly to you?
- Make your content clearly from first hand expertise and a depth of knowledge demonstrate (For example, specialist knowledge that arises from the use of a product or service or a location)?
- Will someone who reads their content will feel that they have a satisfactory experience?
- Is the content mainly led to attract visits to search engines?
- Do you mainly summarize what others have to say without adding much value?
Google’s latest advice on how to cut well in their AI search systems unique, valuable content for people. I think it’s okay to think a little bit of the vector search while we write this content. If we understand and write the needs of our audience to answer these needs in a precisely way, we will create content that is easier for machines to digest and quote AI mode and Aios. This can give us an initial advantage in the ranking. But in the end it is people who decide which pages the most deserved ranking lists are good.
I think it is interesting to play around with the understanding of the cosine similarity that is used to determine whether content is relevant for a query that is considered via the lens of vector search. However, if we concentrate too much on optimizing for machines instead of optimizing people, we can teach Google that our content is really not as good as its vector search systems predict. In this way we teach Google systems to adapt to the adaptation not Prefer our content.
Marie
PS a day after I originally published the preview in this article in the search bar, John Mueller wrote Google: “The optimization of websites for embedding is * literally * keywords.“
((This was originally made for my paid members of the Search bar Pro community. If you are serious to learn AI because you not only change the search, but also our world, we would like to make it to join.)