Nikola Todorovic, Director of Software Engineering at Google Search, appeared on an episode of Search Off the Record to discuss how AI evolved within Google Search.
Todorovic leads the Google SafeSearch engineering team and has worked in the search organization for 15 years. He said machine learning is difficult to apply widely to all of Search because complex models are more difficult to understand and modify than simple systems.
He was explaining why Google can’t use ML systems for all of Search at once. Todorovic said these models “can act like a black box” because engineers don’t always understand what’s going on underneath.
That makes debugging difficult if search programs change over time or if the model needs to be changed, he said.
SafeSearch as Site Proof
Todorovic said SafeSearch is one of the first places where Google can put AI models into Search because the team can separate those programs from the main ranking flow.
SafeSearch may run independent image and video classifiers that generate a signal, such as how obvious the result might be. If problems appear, developers can repeat the model without disrupting the rest of the search.
Convolutional neural networks began improving image understanding about 12 years ago, he said, making SafeSearch a natural early use case for machine learning within Search.
Overview of AI Built into Existing Search
Todorovic described the AI Overview as a feature that “puts a stamp on the top” of Google’s existing search and ranking systems. He said the recall and placement under AI Overviews is still what he calls “old style, old school.”
He said the program could include exit questions from fans. Google may identify additional queries related to the original input, run them in parallel, and return the returned results to a single answer.
The AI overview then aggregates and summarizes information from selected results, including source text, captions, headings, and other page context, he said.
The AI mode follows the same pattern but works with more autonomy, Todorovic said. He described it as still working on Sesho, while “it has a big platform of its own.”
Why This Matters
The quote “black box” gets the attention, but the full context is important. Todorovic was explaining why machine learning was difficult to implement widely across Search, noting that Google has no control over AI Overview or AI Mode.
His comments add useful context to Google’s AI Search documentation. Google has already said that AI Overview and AI Mode may use queries, generating more related searches across topics and data sources to improve answers.
The useful point is not that AI is a “black box.” His comments reinforce that traditional Search programs are still important in the AI Review, as Google narrows the layers and fan exits to the top.
That keeps the basics of traditional Search in line with AI features, as Google changes the way results are summarized and presented.
Looking Forward
The difference between AI Overviews and AI Mode is worth watching as Google expands AI Mode. Todorovic described the AI Overview as separate from the rest of Search, while the AI Mode has its own additional infrastructure.
That difference could matter in how Google defines visibility, measurement, and optimization guidelines as AI Mode evolves.
