Modern AI Search and Semantic SEO increasingly rely on text vectorization. Instead of matching only exact keywords, Google and other platforms compute cosine similarity between embeddings to estimate content relevance, link relevance, topicality, and overall topical authority.
This study investigates which publicly available sentence embedding models can act as low-cost, high-fidelity proxies for Google LaBSE on English similarity tasks. The goal is to identify models that are “Google-like” enough for advanced Semantic SEO analysis—without the latency and cost of running LaBSE or OpenAI’s text-embedding-3-large at scale.
Google describes LaBSE as a language-agnostic BERT sentence embedding model used to measure semantic similarity across languages and content types. For SEO, that implies that much of the real matching—what pages are considered relevant, which links are treated as thematically related—likely happens in embedding space, not purely via keyword matching.
To apply this logic inside your own tools, such as QueryMatch for measuring content relevance and link relevance, you need a model that is close to LaBSE but cheaper and faster to run. This study compares several public models to identify the best proxy candidates for Semantic SEO and AI Search workflows.
Methodology (Short Summary)
We compare three public sentence embedding models against Google LaBSE:
– sentence-transformers/all-mpnet-base-v2
– sentence-transformers/all-MiniLM-L6-v2
– thenlper/gte-base
The models are evaluated on approximately 3,500 English word and sentence pairs from STS-B (sentence similarity), SimLex-999 (word similarity), and SemEval STS 2012–2016 (legacy sentence similarity tasks). For each model we compute cosine similarity for every pair and then:
– compute Spearman rank correlation between each model and LaBSE
– where human gold labels exist, measure correlation with human similarity judgements.
We additionally run a modern-only analysis on STS-B + SimLex-999 (1,999 pairs) to avoid noise from older STS datasets and better reflect current NLP capabilities.
Key Results (Summary Tables)
Table 1 shows the global Spearman correlation with Google LaBSE for all three models, both on the full dataset mix and on the modern-only subset (STS-B + SimLex-999).
| Model | Spearman ρ vs LaBSE (all datasets) | Spearman ρ vs LaBSE (modern only) | Comment |
| all-mpnet-base-v2 | 0.763 | 0.728 | Highest correlation – best overall LaBSE proxy |
| all-MiniLM-L6-v2 | 0.732 | 0.711 | Very close to mpnet – more resource-efficient |
| gte-base | 0.688 | 0.609 | More variable – strong on some tasks, weaker overall |
Table 1. Global Spearman correlation with Google LaBSE for three public models, over both the full dataset mix and the modern-only setting (STS-B + SimLex-999).
Table 2 summarises how the models correlate with human gold labels on STS-B (sentence similarity) and SimLex-999 (word similarity) in the modern-only analysis.
| Model | STS-B ρ vs gold | SimLex-999 ρ vs gold | Strength | Profile |
| LaBSE | 0.758 | 0.619 | Very good | Strong on word-level (lexical similarity) |
| all-mpnet-base-v2 | 0.867 | 0.536 | Excellent | Best at sentence-level – strong all-round proxy |
| all-MiniLM-L6-v2 | 0.838 | 0.446 | Very good | Good balance between performance and speed |
| gte-base | 0.875 | 0.373 | Mixed | Very strong on STS-B, clearly weaker on word-level tasks |
Table 2. Spearman correlation with human gold labels on modern datasets (STS-B and SimLex-999).
Illustrative Result (Figure)
Figure 1 is a simple bar chart showing which models perform best as LaBSE proxies on the modern-only dataset (1,999 text pairs). It makes it visually clear that all-mpnet-base-v2 is the strongest candidate, closely followed by all-MiniLM-L6-v2, with gte-base as a more specialised option.
Figure 1. Best LaBSE proxy models on modern datasets only (STS-B + SimLex-999).
Implications for Semantic SEO, QueryMatch and Link Building
The findings directly extend our work on QueryMatch, where we use Sentence-BERT-style models to analyse query–content similarity, internal links and link relevance in backlink profiles:
👉 QueryMatch – Sentence-BERT for topical authority and link relevance
As AI Search grows (AI answers in the SERP, ChatGPT-style search, AI Overviews), it becomes increasingly important to optimise for embeddings, not just keywords. This connects to our broader work on:
For anyone investing in link building services, this research also updates the requirements for a modern partner: a serious provider now needs to understand embedding models, cosine similarity and topical relevance, not just domain metrics and anchor text:
👉 Guide: Link Building Services in an AI and embeddings world
Further Reading and Resources
- Google’s original blog post on Language-agnostic BERT Sentence Embedding (LaBSE):
👉 https://research.google/blog/language-agnostic-bert-sentence-embedding/ - GitHub repository with code, data summaries and figures for this study:
👉 https://github.com/VesterlundCoder/Proxies-for-Google-LaBSE-Evaluating-Low-Cost-Sentence-Embedding-Models-for-Semantic-SEO - Author profile and related research articles (TrustRank, Semantic Vectorization SEO, QueryMatch and more):
👉 https://independent.academia.edu/DavidVesterlund
