Google has announced an update to voice search that uses AI to make it faster and more accurate. This Speech-to-Retrieval (S2R) model is a major change in how spoken queries are processed. I wanted to unpick whether it has a knock-on effect for SEO too.
Traditionally, when users spoke into their devices, Google’s system first converted speech into text using Automatic Speech Recognition (ASR), and then ran a text-based search query. The S2R model removes that middle step. Instead of transcribing what you say, Google’s AI directly interprets the meaning of your voice input and maps it to the most relevant results.
This shift from ‘speech-to-text-to-search’ to ‘speech-to-meaning-to-search’ has been designed to improve both speed and accuracy. By skipping transcription, S2R reduces errors and better captures intent, even when users speak casually, mispronounce words, or use vague phrasing. According to Google, it marks the beginning of a new phase of semantic voice retrieval – one focused on understanding what users mean, not just what they say.
For marketers, SEOs, and content creators, this shift could be interesting (I’ll explain my caveat later on!) as it may alter how voice-based queries are interpreted, how search rankings are calculated, and ultimately, how content should be structured to remain discoverable.
1. From keywords to intent
S2R transforms queries into vectors (numbers) that represent the semantic meaning of what the person is asking for. It does the same for the potential results and tries to match them up. This all means that content that focuses solely on exact keywords may not be as successful in voice results as it was in traditional SERPS. Instead, Google’s AI will look for conceptual alignment: how well your content captures the intent or topic behind the query.
For example, a spoken query like “what’s that painting with the screaming guy?” no longer depends on keyword overlap with “The Scream by Edvard Munch.” The system recognises the meaning and retrieves relevant results even if the exact words differ.
So what does this mean for optimisation?
Comprehensive content written in natural language is likely to perform better than short, keyword-stuffed pieces. That was the case anyway but now there is an additional need to make sure content is written for real people and not search engines.
2. Conversational content
As people are intrinsically time poor and/or lazy, we tend to type as few words as possible for text search – “ISA rules UK”, or “car tax 2005 audi” whereas for a voice search we’re much more likely to say, “What are the ISA rules in the UK?” or “How much is car tax on a 2005 audi?”
S2R however thrives on understanding these idiosyncrasies of natural language. As users become more comfortable speaking to devices conversationally, the model will interpret context and intent, not just direct requests. So if a user is particularly talkative, it will still be able to pick out their need, even in longer chattier sentences.
So what does this mean for optimisation?
It may be helpful to