Google Discover optimization – technical guide

We have moved from a query-based retrieval model to a predictive push architecture. In this new environment, Google Discover is no longer a secondary traffic source. It is a primary engine for organic growth. The rise of zero-click searches, which now account for approximately 60% of all queries, has forced this pivot. Users are increasingly finding answers directly on the search results page through AI overviews, meaning the traditional click-through economy for informational queries is shrinking.   

Publishers must now master the mechanics of predictive content delivery. Unlike traditional search engine optimization, which reacts to explicit user intent, Discover anticipates latent interests. This system relies on sophisticated machine learning models to map content to user interest vectors without a keyword ever being typed. The implications for technical SEO are profound. We must move beyond keyword density and focus on entity salience and technical congruency.

This report outlines the specific technical frameworks required to succeed. We will analyze the dual encoder architecture, the necessity of the experience signal, and the precise code structures needed for the follow feature.

Neural matching and dual encoder mechanics

To optimize for Discover, one must understand the retrieval system. Google employs a dual encoder architecture, often referenced in research as the two-tower model. This system is designed to handle the scale of billions of users and documents efficiently.   

One tower encodes the user state. This includes search history, location data, and past interactions across the Google ecosystem. The second tower encodes the document candidate. This analyzes the text, headline, and visual assets of your content. The system then calculates the dot product similarity between these two vectors in a high-dimensional space.   

The role of neural matching in retrieval

Neural matching is the bridge between vague user interests and specific content. It allows the algorithm to understand semantic relationships rather than just exact keyword matches. For example, a user interested in “sustainable living” may be served an article about “upcycling furniture” even if the exact terms do not overlap. The neural network maps both concepts to the same vector space neighborhood.   

This technology shifts the optimization focus. Writers must prioritize conceptual clarity over keyword stuffing. The document vector is formed primarily from the headline, the lead image, and the first 200 words of text. If these elements are ambiguous, the document vector becomes “noisy” and fails to align with any specific user interest vector.   

Optimizing for the reranking phase

Once the dual encoders retrieve a set of candidates, a heavier model performs the final ranking. In 2025, this involves multimodal models like Gemini and MUM. These models process text and images simultaneously to determine relevance. They assess semantic congruence between the visual and textual elements. A mismatch here, such as a generic stock photo paired with a specific technical headline, will downgrade the content quality score.   

Entity optimization and the topic layer

The topic layer sits on top of the knowledge graph. It is responsible for understanding how user interests evolve over time. It segments interests into hierarchies and journeys. A user might start with a broad interest in artificial intelligence and mature into a specific interest in large language models.   

Knowledge graph integration strategies

To capture traffic from the topic layer, you must anchor your content to recognized entities. The Cortex framework and similar data architectures illustrate how Google structures this data. You must treat your content as a dataset that feeds into this graph.   

1. Entity disambiguation – use specific nouns. Do not say “the company”. Say “OpenAI”. This helps the natural language processing API extract the correct entity ID.   
2. Structured data implementation – use the SameAs property in your schema. Link your mentioned entities to their Wikipedia or Wikidata entries. This removes ambiguity and strengthens the entity signal.   
3. Topic clustering – organize content to cover an entity exhaustively. This signals to the topic layer that your site is an authority on the entire entity branch, not just a single leaf node.

E-E-A-T and the experience signal

In a feed users do not curate themselves, trust is paramount. Google applies stricter E-E-A-T standards to discover than to web search. The most critical addition for 2025 is the experience signal.   

Demonstrating first-hand experience

Algorithms now aggressively filter generic content. They look for markers of genuine usage or presence. Content that contains phrases like “in our testing” or “we observed” tends to perform better than content that merely summarizes specifications.   

Publishers should adopt a first-person narrative where appropriate. This aligns with the “soft-lens” content that performs well in Discover. Personal stories and unique perspectives act as a proxy for authenticity.   

Author reconciliation and authority

You must help Google connect your authors to real-world people. This process is known as reconciliation. Ensure every author has a robust bio page marked up with ProfilePage schema. Link to their social media profiles and other publications. This builds a private knowledge graph of your site’s expertise that Google can map to its global graph.   

Visual assets and web stories integration

Discover is a visual-first medium. The click-through rate is heavily influenced by the quality and technical specification of the thumbnail image.

The 1200px rule and aspect ratios

Your images must be at least 1200 pixels wide. This is a hard technical requirement for the large image card format, which drives significantly higher engagement than the thumbnail format. You must also include the max-image-preview:large meta tag in the head of your document.   

HTML

<meta name="robots" content="max-image-preview:large">

Avoid using your site logo as the primary image. The multimodal analysis will likely flag it as low-relevance commercial content. Instead, use custom photography or highly specific diagrams that match the semantic context of the headline.

Leveraging google web stories

Web stories remain a potent format for Discover in 2025. They occupy a dedicated carousel and offer a high-visibility entry point. However, they must meet strict technical guidelines.   

1. Video length – keep video segments under 15 seconds.
2. Text density – limit text to under 280 characters per page to avoid the “wall of text” penalty.   
3. Aspect ratio – use 9:16 for all assets.
4. Metadata – ensure the poster-portrait-src attribute points to a valid 3:4 image file.   

Technical infrastructure for the follow feature

The follow button in Discover allows users to subscribe directly to a publisher. This feature relies entirely on your RSS or Atom feeds. If your feed is broken or generic, you lose the ability to retain an audience.

Feed specifications for optimization

Your feed must be technically flawless to support the follow feature. It should be linked in the <head> of your hub and leaf pages.

• Full content – do not serve partial snippets. The feed should contain the full article body to allow the algorithm to index the content correctly for the following tab.   
• Static IDs – never change the <guid> or ID of an article after publication. This causes duplicate content issues in the user’s feed.   
• Media tags – use strictly defined <media:content> tags to specify the high-resolution image asset. This ensures the correct image appears in the sub-feed.   

The curiosity gap versus clickbait policy

Writing for Discover requires balancing high click-through rates with strict policy compliance. Google penalizes clickbait that withholds information or exaggerates outcomes. However, the curiosity gap is a valid psychological trigger that can be used effectively.   

Defining the safe zone

Clickbait obscures the subject. A curiosity gap highlights a specific, interesting detail about a known subject.   

• Bad (Clickbait) – “You won’t believe what this company did.” (Subject is hidden).
• Good (Curiosity Gap) – “The specific feature that makes the Pixel 9 Pro unique.” (Subject is clear, the specific detail is the gap).

The algorithm measures post-click satisfaction. If a headline generates a click but the user returns to the feed immediately, known as pogo-sticking, the headline is reclassified as misleading. This can lead to a manual action or algorithmic suppression.   

Analytics and traffic recovery

Traffic from Google Discover is inherently volatile. It does not follow the predictable patterns of search volume. It moves in waves based on entity trending and user interest spikes.   

If you experience a sudden decline, investigate entity drift. If your site has published content outside its core expertise, the topic layer may have diluted your authority signal for your primary niche.   

Recovery requires a return to core competencies. Prune content that lacks deep expertise. Update evergreen content with fresh data to signal relevance to the freshness algorithm. Do not panic over minor fluctuations. Focus on the long-term trend of active engagement rather than daily session counts.