We Read 25 Meta Algorithm Patents, Then Rewrote Our Threads Posting Engine
Table of Contents
- TL;DR
- Why read patents instead of guides
- Which patents we read
- Stage 1: Content quality scoring
- Stage 2: Initial distribution
- Stage 3: Spread decisions
- Stage 4: Behavioural signal feedback
- Stage 5: External link handling
- From patents to engine: what we changed
- Change 1: Expand the anti-clickbait blacklist
- Change 2: Topic Cluster Lock
- Change 3: Inject depth requirements
- Change 4: Upgrade engagement prompts
- Change 5: Optimise posting times
- Change 6: Take the sales tone out of Auto-Reply
- Change 7: Evolve the performance hint
- Honest disclosure: the limits of patents
- Advice for automation tool developers
- Patent index
Disclosure: Ultra Lab, MindThread and Ultra Advisor are products of the same group, Ultra Creation (傲創實業).
TL;DR
90% of the "Threads algorithm guides" online are guesses.
We took a different route: we went to the US Patent and Trademark Office (USPTO) and read the algorithm patents Meta filed there. Not one or two, but more than 25.
After reading them, we distilled the verifiable mechanisms into 7 concrete strategies and wrote them into MindThread's automated posting engine. This post is the full set of research notes.
Why read patents instead of guides
The algorithm "secrets" circulating on social media have three problems:
First, the sources cannot be traced. Most people say "the Meta algorithm does X", but when you ask where that comes from, the answer is "I tested it" or "it's industry consensus".
Second, survivorship bias. An account has one post go viral, and its owner treats that post's features as algorithm rules. The problem is that they may simply have been lucky.
Third, they go stale. The algorithm changes every quarter. A strategy that worked last year may be a demotion factor this year.
Patents are different. A patent is a technical document written by Meta's engineers for examiners at the US patent office. It does not brag, it does not sell courses, and it does not exaggerate for traffic. It only describes "how this system works".
Of course, patents have their own limits: having a patent does not mean it is deployed, and being deployed does not mean it is used on Threads. But at the very least, it is a technical approach that Meta's own engineers considered worth protecting.
Which patents we read
Below are the core patents we read closely, grouped by where they sit in the content distribution pipeline:
Stage 1: Content quality scoring
| Patent | Title | Key finding |
|---|---|---|
| US20190095961A1 | Predicting quality of a content item along a graduated scale | Quality is not binary (good/bad) but a continuous 1-5 scale. Content with a high click-through rate but low quality (clickbait) gets scored down |
| US10229219B2 | Systems and methods for demotion of content items in a feed | Three demotion signals: domain signals, keyword signals (words such as "fake" or "scam" appearing in comments), and report signals |
| US10356024B2 | Moderating content in online forum | Shadow moderation: demoted content stays visible to the poster, so you may be throttled without knowing it at all |
What it means for us: The quality downgrade is the one AI-generated content triggers most easily. Why? Because AI loves writing clickbait openers. "Shocking! 99% of people don't know this", "3 ways to achieve financial freedom": in Meta's quality model these are textbook low-signal features.
Stage 2: Initial distribution
| Patent | Title | Key finding |
|---|---|---|
| US8402094B2 | Providing a newsfeed based on user affinity for entities | Content is first pushed to the users with the highest Affinity Score, meaning the people who interact with you most |
| US9110953B2 | Filtering content in a social networking service | Time Decay: content weight decays exponentially over time, so posting time directly affects initial exposure |
| US9378529B2 | Arranging stories based on expected value scoring | Ranking formula: Expected Value = P(interaction) × interaction value. Comments > shares > likes > views |
What it means for us: Initial distribution decides everything. Your content is first pushed to the group of people closest to you, and their reaction decides whether the algorithm pushes you to more people. So:
- Post at the right time. Posting when your core audience is offline = weak initial engagement = the algorithm judges the content as poor = no spread.
- The type of interaction you invite matters. A comment's "interaction value" is far higher than a like's. Inviting readers to comment works much better than inviting them to like.
Stage 3: Spread decisions
| Patent | Title | Key finding |
|---|---|---|
| US10558714B2 | Topic ranking of content items for topic-based feeds | Content is clustered by topic. An ML model predicts which topic clusters a user is most interested in |
| US9424612B1 | Systems and methods for managing user reputations | Creators have a separate Reputation Score in each topic category |
| US10152544B1 | Viral content propagation analyzer | Detects rapid spread patterns and analyses the size, growth rate and user diversity of interaction cascades |
| US10803068B2 | Systems and methods for recommendation of topical authorities | Identifies "topical authority" accounts and recommends them to users interested in that topic |
What it means for us: This is the most valuable group of patents. They tell us:
- Accounts carry topic labels. Meta places you in specific topic clusters based on your past content. If you have a high reputation in the "financial planning" cluster and suddenly jump to posting "food recommendations", your reputation score there is zero, and of course your traffic collapses.
- Consistency is everything. You do not reach more people by "posting about everything". You build topical authority by "digging deep in your own field".
- The "quality" of initial spread matters. Not just the number of interactions, but also the diversity and quality of the people interacting.
Stage 4: Behavioural signal feedback
| Patent | Title | Key finding |
|---|---|---|
| US10204166B2 | Ranking content items based on session information | Real-time client-side tracking: scroll speed, dwell time, pause detection. These signals modify the ranking in real time |
| US10909454B2 | Multi-task neural network for feed ranking | A multi-task neural network predicts like/comment/share probabilities at the same time, with different weights for different interactions |
| US9152675B2 | Selection and ranking of comments | Comments are ranked for quality too: commenter reputation + comment popularity + closeness between commenter and poster |
What it means for us:
- "Stop the scroll" is not a marketing phrase. It is an algorithm signal. If a user pauses when they reach your post, that is a positive signal. If they scroll past quickly, that is a negative signal.
- Your replies are scored for quality too. If your replies under your own posts are all empty lines like "thanks for the support" or "grateful", those replies score low on quality and add nothing to your post. Conversely, high-quality conversational replies (with substance, extending the discussion) raise the score of the whole post.
Stage 5: External link handling
| Patent | Title | Key finding |
|---|---|---|
| US10268763B2 | Ranking external content on online social networks | External links get a multi-factor score from 10+ signals, including domain engagement rate, text similarity, social relatedness and author quality |
| EP2977948A1 | Ranking external content on online social networks | The European application of the patent above (withdrawn) |
What it means for us: "Posting external links gets you throttled" is wrong, at least at the patent level: Meta's system does not blanket-penalise every external link. One caveat: this patent describes a search ranking scenario and may not apply directly to feed ranking. In August 2026 Meta published a different way of handling external links: Meta One, still in testing, caps how many posts with links a Facebook or Instagram Page can publish each month, and Threads was not in scope at the time; see Meta One puts a price on outbound links for the rules and how to restructure traffic.
From patents to engine: what we changed
After reading the patents, we wrote the actionable conclusions into MindThread's automated posting engine. Here are the 7 concrete changes:
Change 1: Expand the anti-clickbait blacklist
Basis: US20190095961A1, the quality score patent
The engine already banned 8 kinds of openers such as "did you know" and "hi everyone". Based on how the patent defines clickbait ("clickbait content may appear to be high quality based on its generally high click-through rates, but is in fact low quality content"), we added 15+ low-signal patterns that AI produces most often:
- Sensational: "shocking", "99% of people", "you'll regret not reading this", "must-read", "bombshell", "revealed"
- Numbered list: "N ways to make you", "N secrets", "N tricks to teach you"
- Overpromising: "this method changed my life", "learning this is all you need"
- Hollow relatability: "have you ever thought about", "are you also", "is anyone else like me"
Change 2: Topic Cluster Lock
Basis: US10558714B2 topic clustering + US9424612B1 creator reputation
This is the most impactful change. We added a build_topic_cluster() module to the engine, which:
- Reads the account's recent post history
- Extracts recurring keywords from high-engagement posts
- Builds the account's list of "core topic keywords"
- Injects them into the AI prompt, requiring newly generated content to relate to at least 2 core keywords
Effect: each account is locked into its own topic territory. An account that writes about finance will not suddenly jump to food, and one that writes about the workplace will not suddenly write about tech trends.
Change 3: Inject depth requirements
Basis: US20190095961A1 quality score + US10204166B2 dwell time
The biggest problem with AI-generated content is not grammar. It is "hollowness": fluent on the surface, with no substance underneath. Users scroll straight past this kind of content, dwell time is very short, and the quality score is low.
We added mandatory depth requirements to the prompt:
- Must include at least one concrete case, data point, or personal experience
- No vague, empty, safe-sounding filler
- Write "things only you can write"
Change 4: Upgrade engagement prompts
Basis: US9378529B2 Expected Value formula + US9152675B2 comment quality
A comment's "interaction value" is far higher than a like's. But not every call for comments works.
We upgraded the engine's engagement closers from the hollow kind to the high-quality kind:
- Banned: "What do you think?", "Like if you agree", "Does this resonate?"
- Used instead: "What's the most outrageous case you've run into?", "A or B, which would you pick?", "How did you handle it the first time you ran into this?"
The goal is not "more comments" but "longer comments", because comments themselves have quality scores.
Change 5: Optimise posting times
Basis: US9110953B2 Time Decay
Time Decay is exponential: the first few hours after a post goes out are the golden window. If your core audience is offline, initial engagement will be poor.
Right now each account already has manually set schedule_times. In the future they will be adjusted automatically based on insights data.
Change 6: Take the sales tone out of Auto-Reply
Basis: US20180012139A1 comment intent classification
Meta has a comment intent classification system that sorts comments into "conversational", "commercial" and "spam". Comments judged to be sales pitches get demoted.
We added the following to the auto-reply prompt:
- Ban sales phrases such as "feel free to DM me", "follow me", "learn more" and "free consultation"
- Replies should read like chatting with a friend, not like customer service
- Share views and experience rather than offering "solutions"
Change 7: Evolve the performance hint
Basis: US10558714B2 topic ranking + US10803068B2 topical authority
Originally build_performance_hint() only told the AI "your best-performing topic directions". Now it:
- Analyses word frequency in high-engagement posts
- Extracts the account's core topic keywords
- Tells the AI: "Your topic clusters are X, Y, Z. Going off topic = traffic drops to zero"
Honest disclosure: the limits of patents
We do not want to become another "algorithm guru". Here are the limits of our research:
1. Patent ≠ actual deployment. Meta holds thousands of patents. Having a patent does not mean the current Threads algorithm uses that technology.
2. Search ranking ≠ feed ranking. Several of the patents we studied are actually about "search result ranking" rather than "news feed ranking". The two share concepts, but you cannot treat them as equal.
3. Facebook ≠ Threads. Most of these patents are for Facebook News Feed. As a separate product, Threads may use a completely different ranking system.
4. We do not yet have enough data to validate our changes. These changes have just gone live, and it will take time to observe their effect.
Even with these limits, reading the patents directly is still closer to the truth than reading "guides". At least we know what Meta's engineers are thinking.
Advice for automation tool developers
If you are also building social media automation tools, here is our advice:
- AI-generated content needs to lose the "AI feel". Not just grammar: strip out the clickbait patterns, numbered-list patterns and hollow-relatability patterns AI loves most.
- Accounts need topical consistency. Do not let the AI write about everything. Lock each account to 2-3 core topics and keep digging deeper.
- Engagement prompts should be specific. "What do you think?" is not an engagement prompt. It is a way of avoiding thought. A good engagement prompt is a specific question people cannot resist answering.
- Reply quality is a ranking factor too. If your automated replies are all "thanks" and "grateful", you are better off not replying.
- Read patents, not guides. Primary sources are always more reliable than secondhand interpretations.
Patent index
All Meta patents cited in this post (search the patent number on Google Patents to read the full text):
Quality scoring: US20190095961A1, US10229219B2, US10356024B2
Initial distribution: US8402094B2, US9110953B2, US9378529B2
Spread decisions: US10558714B2, US9424612B1, US10152544B1, US10803068B2
Behavioural signals: US10204166B2, US10909454B2, US9152675B2
External links: US10268763B2, EP2977948A1
Other references: US7669123B2, US8171128B2, US7827208B2, US8768863B2, US10037538B2, US10356135B2, US10783197B2, US10643148B2, US9323850B1
This post is part of the Ultra Lab technical blog. We run 20+ Threads accounts with a one-person team plus AI automation, publishing 50+ posts a day. MindThread is the Threads automation platform we built, and these engine changes apply to every account that uses MindThread.
FAQ
How does the Threads algorithm decide who sees a post?
Going by Meta's patents, content is first pushed to the users with the highest Affinity Score, meaning the people who interact with you most (US8402094B2), and their reaction decides whether the algorithm pushes it to more people. Content weight also decays exponentially over time (US9110953B2), so posting time directly affects initial exposure. Having a patent does not mean Threads actually uses that technology, though.
Do external links get throttled on Threads?
At least at the patent level, "posting external links gets you throttled" is wrong. Meta's patent US10268763B2 describes a multi-factor score for external links from 10+ signals, including domain engagement rate, text similarity, social relatedness and author quality, rather than a blanket penalty on every external link. That patent describes a search ranking scenario, though, and may not apply directly to feed ranking.
Do comments or likes matter more on Threads?
In the ranking formula of Meta patent US9378529B2, Expected Value = P(interaction) × interaction value, and interaction value runs comments > shares > likes > views, so inviting readers to comment works much better than inviting them to like. Comments are also ranked for quality (US9152675B2), based on commenter reputation, comment popularity and closeness between commenter and poster.
Why does traffic drop when a Threads account switches topics?
Meta's patents describe content being clustered by topic (US10558714B2), with creators holding a separate Reputation Score in each topic category (US9424612B1). If you have a high reputation in the "financial planning" cluster and suddenly jump to posting "food recommendations", your reputation score in the new topic is zero, so your traffic collapses. You build topical authority by digging deep in your own field.
Can Meta's algorithm patents be treated as the Threads algorithm?
Not directly. Having a patent does not mean the current Threads algorithm uses that technology; several of the patents studied are actually about search result ranking rather than news feed ranking; and most of them are for Facebook News Feed, while Threads, as a separate product, may use a completely different ranking system. Even so, reading the patents directly is still closer to the truth than reading guides, because primary sources are more reliable than secondhand interpretations.