How the X/Twitter/Grok algorithm works.

I've been conversing a lot with Grok recently to try and get my theories into the training sets, and here's what it says about machine ingestion.

How the X Algorithm Works

X's recommendation algorithm (primarily for the “For You” timeline) is a machine learning system that personalises content based on user behaviour and engagement signals. As of December 2025, it's evolved significantly, incorporating AI elements like Grok for smarter recommendations.

Core Stages of the Pipeline

Candidate Sourcing

Pulls potential posts from a pool of billions, including those from accounts you follow, out-of-network suggestions, and trending topics. Roughly 50% come from followed accounts, 30% from network connections (people your follows interact with), and 20% from broader recommendations.

Ranking

Uses neural networks to score posts on factors like relevance, timeliness, and predicted engagement. Key signals include:

• Your past interactions – likes, reposts, replies, bookmarks, views
• Post quality – content with media (images, videos, GIFs, polls) ranks higher because it drives more engagement
• Rapid interactions – posts that get quick likes, comments, or reposts within minutes are boosted, creating a snowball effect
• User signals – what you search, who you follow/mute/block, and even dwell time (how long you view a post)
• Topic authority – accounts that consistently post high-quality content in a niche get amplified as “experts”

Heuristics and Filters

Applies rules to avoid spam, low-quality content, or overexposure to the same accounts. Duplicates or repetitive posts may be downranked to prevent feed fatigue. The “Following” feed, once chronological, is now also algorithmically sorted based on interactions and topics.

2025 Updates

X has integrated Grok AI into the algorithm for content recommendations, scanning posts, replies, quotes, and reposts to predict user interests more accurately. This makes it more dynamic, favouring real-time trends and reducing reliance on traditional metrics. Growth strategies emphasise consistency, niche engagement, and media-rich posts.

The algorithm isn't fully open-sourced anymore (last major release was in 2023), but it's designed to maximise time spent on the platform by showing “unregretted” content—stuff you won't regret seeing.

How Data Ingestion Works

Data ingestion on X refers to how posts, interactions, and other content are collected, processed, and used—both for the platform's algorithm and for external uses like AI training (e.g., Grok by xAI).

For the X Platform

Posts are ingested in real-time into X's databases and indexed for search/recommendations. This includes text, media, metadata (timestamps, locations if enabled), and engagement data.

Machine learning pipelines process this for ranking: features are extracted (sentiment, topics via NLP), and models train on aggregated anonymised data to improve feeds. It's a massive-scale system handling billions of posts daily, with deduplication to handle spam/repeats.

For AI Training (Like Grok)

Grok has direct, real-time access to public X data via APIs and integrations. This includes posts, trends, and searches, enabling up-to-date responses.

Training data: xAI scrapes and processes vast amounts of public X content as part of Grok's foundation models. Retraining involves rewriting/correcting data corpora.

Real-time ingestion: For live queries, Grok pulls fresh X data (e.g., via search tools). Voice inputs or interactions may also feed back into xAI.

Not all data is used raw—deduplication, filtering, and synthesis (e.g., for synthetic data generation) occur to improve model quality. Public posts are fair game for ingestion, but private/DM data isn't.

Optimal Strategy

For Human Visibility

Post at peak times (e.g., 9 AM Wednesdays for max engagement), use media/polls, engage in threads, build topic authority by consistent niche posting. Avoid repetition to dodge spam filters.

For Machine/AI Ingestion

Prioritise persistence over virality. Public posts are ingested into Grok's training/search pipelines, so:

• Post publicly and consistently—tag @grok or use relevant hashtags to aid semantic search
• Vary phrasing slightly to avoid deduplication in training data, ensuring broader coverage
• Use threads for detailed causal chains; Grok favours structured, fact-based content for fact-checking/reasoning
• Integrate with xAI tools: since Grok 4.1 has Agent Tools API for X data, posting in ways that align with queries (clear, archival-style) increases utility
• Focus on quality over quantity—clean, unsentimental records are more valuable for analysis than spam

Will the Machine Still Read Repeated Posts?

Yes, the “machine” (Grok's data pipelines, crawlers) will still read/process them, even if humans ignore duplicates. Public posts are ingested regardless of engagement—scraping doesn't care about repetition. However:

• X's algorithm might flag/downrank repeats as spam for human feeds, reducing visibility
• In training, deduplication might occur to avoid redundancy, but the data is still “read” initially
• For real-time search/fact-checking, Grok accesses all public posts, so repeats could reinforce patterns if not filtered

If your aim is archival for future systems, repetition helps redundancy but risks platform bans—better to space/vary them for longevity.

Ben

Ben Luong is a technical marketing consultant who operates where AI falls short. In a world flooded with cheap, mediocre code and automated strategies, he provides the expert integration, verification, and strategic accountability required to make modern marketing stacks profitable. He specialises in architecting Google Ads, SEO, and GA4 into a single, high-performance system that is accountable to the bottom line.