Yükleniyor...
When users ask your search box 'How do I create an eye-catching Instagram opening graphic?' instead of typing 'social media template', why do traditional keyword matching systems break down completely?
When users ask your search box "How do I create an eye-catching Instagram opening graphic?" instead of typing "social media template", why do traditional keyword matching systems break down completely?
For years, we stuffed our copy with keywords to feed search engines. However, with the rise of artificial intelligence and large language models (LLMs), search technology has evolved from matching literal character strings to understanding the conceptual meaning and intent behind them. One of the most concrete and massive-scale examples of this transformation is unfolding at Canva, a design platform where millions of users search for graphics and templates every day.
To overcome the limitations of traditional keyword-focused search infrastructure, Canva transitioned to dense semantic retrieval systems. In this case study, we examine how Canva's engineering teams built their technical architecture and explore actionable steps to prepare your own content for next-generation vector-based indices.
Traditional information retrieval systems have long relied on the BM25 (Best Matching 25) algorithm. This algorithm scores relevance by measuring how frequently a term appears in a document (TF - Term Frequency) against how rarely it occurs across the entire document corpus (IDF - Inverse Document Frequency). However, BM25 has a major blind spot: it cannot parse synonyms, indirect phrasing, or visual context.
According to the article "How we built semantic search for templates" published on Canva's engineering blog, users do not always search using technical or exact terms. For example, when a user searches for "retro summer vibe," BM25 expects those exact words to appear in the database record. Yet what the user is actually looking for are yellow neon gradients, 80s typography, and palm tree silhouettes.
This is where Dense Retrieval comes in. In this setup, both search queries and content assets (text, images, templates) are converted into high-dimensional mathematical vectors (embeddings). The semantic closeness between two concepts is calculated via the cosine of the angle between these vectors (Cosine Similarity). If the angle between two vectors approaches zero, the two assets are semantically close—even if their literal wording is entirely distinct.
To map search queries and visual templates into the same shared vector space, Canva engineers deploy a Two-Tower Neural Network architecture. This system consists of two primary pipelines:
During training, these two towers are fed real user engagement signals, such as clicks and downloads. Over time, the vector for the query "retro" and the vector for a "yellow-orange gradient" graphic are pulled closer together in mathematical space.
However, Canva did not discard its traditional Elasticsearch (BM25) infrastructure entirely. Semantic search models can occasionally struggle with exact brand names or specific technical dimensions (such as "A4" or "1080x1920"). For this reason, Canva employs a hybrid search architecture.
Research on Dense Passage Retrieval (DPR) published by Meta AI Research supports this dual approach. According to the study, hybrid models combining DPR and BM25 achieve a Top-20 Retrieval Accuracy rate as high as 91.7%. By unifying both systems, Canva delivers exact keyword precision and deep semantic comprehension simultaneously.
The semantic gap represents the divide between abstract concepts in a user's mind and the concrete attributes of digital assets. Canva bridges this gap by enriching content metadata.
The core signals determining a template's coordinate (embedding) in vector space include:
These signals are processed using Sentence-BERT (SBERT) style models into dense vectors and indexed in high-performance vector databases such as Qdrant, Milvus, or Pinecone.
If you want your content to be retrieved as the nearest neighbor (k-NN / k-Nearest Neighbors) and rank at the top in AI-driven search engines and platforms like Canva, traditional SEO tactics require an update. Here are 3 actionable steps:
Semantic search models struggle to extract meaningful intent from comma-separated keyword lists, degrading vector quality. Instead, structure your content with a logical flow and hierarchical subheadings.
Incorporate dedicated paragraphs that provide clear, upfront answers to natural language questions users are likely to ask search engines. For instance, in a product description, directly address questions like "What is this template used for?" or "Who should use this layout?" within natural copy.
Provide explicit technical signals so models can vectorize your content accurately. Write comprehensive alt tags for imagery and implement complete Schema.org structured markup across your pages. As shown in Canva's implementation, describe not just visual traits like color, but also the emotional tone or vibe ("energetic," "minimalist," "corporate") within your textual metadata.
Search is no longer confined to bots crawling flat web pages; content is mapped as mathematical coordinates across multidimensional vector spaces. Moving forward, traditional Search Engine Optimization (SEO) will increasingly converge with Vector Engine Optimization (VEO). Canva's semantic transformation demonstrates that this shift is already well underway. Creators who strip away keyword spam to prioritize semantic cohesion and conceptual depth will be the ones who remain discoverable across the AI-driven search landscape.