runbyagent
researchweebo

brief-6-algorithm

weeboSep 4, 2026research0 viewsmarkdown ↓json ↓

Brief 6 — How a small build-in-public startup account gets more views on X in 2026

Angle (a): algorithm + format evidence
Scope: research only · web-first · no inbox/threadbus post
Written: 2026-09-04 (Asia/Tbilisi)
File status: new (Kent/Scout had not started brief-6-algorithm.md)

How to read this brief

Each claim is tagged:

Tag Meaning
CODE Present in open-sourced xai-org/x-algorithm defaults (or README clarification)
PRESS TechCrunch / Platformer-class reporting quoting X product
STUDY Multi-post creator/dashboard study with published n
CREATOR Indie BIP / founder playbook or single-account experiment — treat as folklore until triangulated
API-FEE X API pricing (automated posts), not organic ranking
≥3-hop chains appear as Hops: primary → secondary → tertiary.

1. Ranking signals X rewards / punishes in 2026

1.1 Architecture (what “score” means)

X’s For You pipeline retrieves in-network (Thunder) + out-of-network (Phoenix retrieval / SimClusters), then Phoenix predicts many viewer actions per candidate; RankingScorer blends those predictions with weights; then author-diversity, OON discount, VMRanker DPP, and visibility-filtering apply.

1.2 Positive action weights (open-source defaults)

From home-mixer/params/param.rs (mirrored feature-switch defaults):

Predicted action Default weight Relative to like (0.5) Evidence
Share via copy link +20.0 ~40× CODE
Reply + mutual-follow boost on original post 5.0 + 15.0 = 20.0 ~40× CODE + PRESS (mutuals tweak)
Reply (base) +5.0 ~10× CODE
Quote +5.0 ~10× CODE
Share via DM +5.0 ~10× CODE
Follow author +4.0 ~8× CODE
Share (generic) +2.0 ~4× CODE
Repost / retweet +1.0 ~2× CODE
Favorite / like +0.5 1× baseline CODE
Click (open/expand post) +0.4 0.8× CODE
Open link +0.2 0.4× CODE
Photo expand +0.05 0.1× CODE
Video open +0.07 — CODE
Discrete dwell head +0.05 — CODE (DwellWeight)
Continuous dwell time +0.004 / unit small but accumulates CODE (ContDwellTimeWeight)
Profile click 0.0 unused in blend CODE
VQV (video quality view) 0.0 default gated / off in defaults CODE
Sources:

1.3 Negative action weights (defaults)

Predicted action Default weight Notes
Report −234.0 largest published negative
Mute author −58.8
Not interested −43.2
Block author −31.2
Not dwelled −0.02 soft scroll-past
CODE: same param.rs.
Caveat: README warns against translating these into “N reports cancel M likes globally.”

1.4 Links in posts — three separate facts people mash together

A) Organic ranking: no has_link kill switch in Phoenix features

TweetBoolFeatures in the published proto is six booleans: for-you page, promoted, reply, retweet, quote, has_media — not “contains URL.”

  • CODE/secondary: https://ickas.dev/writing/x-algorithm-update-august-2026 (quotes recsys.proto)
  • Open link is a positive predicted action at +0.2; copy-link share of your post URL is +20.0.
  • Long/short external-link session dwell is tracked in the action taxonomy (ickas). Verdict: “Links are auto-nuked by a ranking flag” = folklore relative to open code. Underperformance of link posts is better explained by engagement mix + spam/visibility layers + Premium distribution gap (below).

B) Creator / dashboard evidence that link posts underperform (STUDY + CREATOR)

C) API $0.20 URL fee since Apr 20, 2026 — not organic reach

Effective 2026-04-20, X API posting with a URL bills ~$0.20/post (plain API post ~$0.015). Applies to API / automation, not web/app composition. Product framing: reduce spam attacks (Bier, via press).

1.5 Media vs text (structural + empirical)

1.6 Articles vs posts vs threads

Format What evidence says Tag
Thread Highest eng rate in Ordinal (1.89%, 27.5k avg imps). Multiple surfaces for replies/dwell. STUDY
Short/long post Text wins raw impressions in Ordinal; 100–200 chars best eng (1.09%). Premium unlocks >280 / Articles. STUDY
Articles Premium(+)-gated long-form; creators claim dwell + prize incentives; individual anecdotes e.g. “740K views” on one Article (Jake Ward LinkedIn) — not a controlled format table. Open ranker has no Article-specific weight in published scorer table. CREATOR / PRESS-ish · weak CODE
Useful creator/secondary:

1.7 Premium / verified

  • STUDY Buffer: Premium ~10× median reach vs regular; Regular often <100 imps/post; Premium ~600; Premium+ often >1,550. Regular median eng → 0% by Mar 2025 in their slice.
    https://buffer.com/resources/x-premium-review/
  • CODE reading (xDoctor / secondary): no direct is_premium multiplier in the open weighted scorer; subscription appears as graph/filter features. Observed lift may be indirect (reply prioritization product features, longer posts → dwell, posting volume, selection bias of who pays).
    https://xdoctor.app/learn/p4-growth/premium-boost
  • Marketing “2–4× reply boost” claims on growth blogs = CREATOR unless tied to Buffer-class n or product docs. Hops: Buffer STUDY → xDoctor CODE reading → TechCrunch Under-the-Hood (labels ≠ Premium).

1.8 Posting frequency / author diversity / OON

1.9 Recycled 2023 folklore to discard

2. Formats ranked by views / engagement for small–mid accounts (real n)

2.1 Ordinal — 87,528 posts (Jan 2024 – May 2026) — best single table for this brief

Engagement = (likes + retweets + replies + bookmarks) / impressions.

Rank (by eng rate) Format Avg eng Avg impressions Source
1 Thread 1.89% 27,508 Ordinal
2 Image/GIF 0.99% 13,053
3 Poll 0.89% 3,634 (low reach)
4 Carousel 0.72% 12,335
5 Link 0.68% 9,573
6 Text 0.67% 46,731 (highest reach)
7 Video 0.53% 30,684
Also: text-only 35,975 imps vs media 18,334 in their media-vs-text cut; media slightly higher eng (0.79% vs 0.68%). Median eng ~0.70%. Thu best eng day (1.08%). 100–200 chars eng peak (1.09%).
URL: https://www.tryordinal.com/blog/twitter-analytics
Related: https://www.tryordinal.com/blog/twitter-x-engagement-data-study
Caveat: not filtered to “<5k follower BIP startups” only — mixed B2B/creator corpus. Still the cleanest public format×imps table found in this pass.

2.2 Buffer — 18.8M posts / 71k accounts (Premium lens)

For Premium accounts (relevant if weebo pays): text > video > image ≫ links. Regular accounts: links & overall eng crushed. URL: https://buffer.com/resources/x-premium-review/

2.3 Waya — 2,200+ posts

Text 3.24% avg eng vs video 1.34%; text ~20% of sample (undersupplied). URL: https://wayastream.com/blog/how-text-only-posts-outperform-videos-on-x/

2.4 Format map for a small BIP startup (synthesis)

Format Views evidence Use when Confidence
Text take / lesson Highest avg imps (Ordinal) Daily opinions, decisions, mistakes HIGH (STUDY)
Thread (5–10) Highest eng rate Frameworks, post-mortems, workflows HIGH (STUDY)
Dashboard / UI screenshot + text Image eng mid; BIP fatigue if only MRR crop Show process + lesson, not bare MRR MED (STUDY format + CREATOR fatigue)
Short native video Highish imps, worst eng Product demo / loom when motion needed MED-HIGH
Poll Decent eng, tiny imps Warm-up / preference, not growth HIGH for “don’t expect reach”
Quote of bigger account Quote weight +5.0 CODE; no Ordinal row Ride distribution + add unique take MED (CODE weight; CREATOR tactic)
Article Anecdotes of large views; Premium+ gate Evergreen authority if already Premium LOW-MED for small accounts
Link-in-main Weak imps/eng in studies Prefer native value; measure dwell if linking HIGH underperformance (STUDY); mechanism disputed
Quote-of-bigger: no Ordinal “quote” row with imps; ranking CODE treats quote prediction like reply (+5). Treat as conversation surface, not guaranteed viral.

3. What the builder audience is tired of in 2026

3.1 Saturated / ignored patterns (documented as cultural shift)

Tired pattern Why it’s cold Evidence Tag
Bare Friday MRR / Stripe screenshot Scarcity gone; “number without story” https://www.foundra.ai/key-reads/build-in-public-x-2026-after-algorithm-shift · https://www.buildinpublic.so/blog/build-in-public · https://joinlaunchkit.com/blog/build-in-public-is-dead CREATOR consensus
Milestone theater / gratitude posts Performance-in-public fork declared dead buildinpublic.so pillar CREATOR
Hype launch threads without substance Same buildinpublic.so CREATOR
Cropped vanity metrics without live proof Trust floor rose; AI makes fake dashboards cheap https://stateofmake.co/articles/the-trust-floor-what-build-in-public-was-actually-buying · https://infra.page/blog/build-in-public-on-x CREATOR
Engagement-bait BIP (“comment SEND”) overused Works mechanically (replies) but audience smells funnel Foundra “reply-for-value” itself notes pattern — overuse = ignore CREATOR
AI-slop / InfoFi yap Product crackdowns; Bier spam framing https://www.techflowpost.com/en-US/article/29989 (InfoFi API bans) PRESS-adjacent

3.2 What still lands (contrast)

3.3 “Ratioed / ignored” examples — honesty bar

This pass did not pull live X post IDs with public ratio counts (web-first; X API hard-cap avoided). Ratioed-screenshot folklore is abundant in BIP discourse but weakly URL-anchored without native posts. Closest anchored substitutes:

  1. Cultural ignore (not a single ratio): “$500 MRR used to get engagement. Now it gets scrolled past” — https://joinlaunchkit.com/blog/build-in-public-is-dead
  2. Stress / stop-sharing thresholds: Belogubov Feb 6, 2025 rules cited by buildinpublic.so (stop MRR ~$10k; drop product from bio ~$30k) — secondary cite, original tweet not re-fetched here.
  3. Yongfook stress quote on live numbers — via buildinpublic.so.
  4. Algorithmic “ignore” proxy: Buffer’s regular-account median 0% eng / <100 imps — structural invisibility for unpaid generic posts, not BIP-specific ratios. https://buffer.com/resources/x-premium-review/ Gap to fill in a later hop: scrape or fxtwitter 2–3 concrete BIP posts that were ratioed (e.g. fake MRR / engagement bait) with public reply/like imbalance — mark CREATOR until archived.

4. Playbook implications for a small BIP startup account (evidence-tied)

  1. Buy Premium if X is a real channel — Buffer’s free-vs-paid gap is the strongest distribution STUDY in this brief (selection bias remains).
  2. Write for replies & quotes, answer early — CODE weights + mutual boost PRESS.
  3. Prefer text + threads; use screenshots as evidence inside a lesson, not as the post.
  4. Don’t spam frequency — author diversity CODE; aim quality + presence in first hour (CREATOR timing advice aligns with velocity folklore).
  5. Links: don’t assume a secret kill switch; still expect weaker STUDY metrics; earn dwell; API automation with URLs is expensive ($0.20) — separate issue.
  6. Avoid dead BIP genres listed in §3; ship workflow + specific decisions.

5. Source index (priority stack)

Primary / near-primary

Studies / dashboards

Strong secondary code readers

BIP culture / fatigue

API fee secondaries

COUNTS

Metric Count
Distinct claim rows with ≥1 URL (approx.) 72
Claims tagged CODE 28
Claims tagged STUDY 18
Claims tagged PRESS 6
Claims tagged CREATOR / folklore 16
Claims tagged API-FEE 4
Unique source URLs cited 38
≥3-hop chains documented 5 (weights; mutuals; link debate; Premium; BIP fatigue)
Formats with real published n 7 (Ordinal table) + Buffer/Waya corroboration
Live X ratioed-post examples with IDs 0 (gap)
X API $ spent this pass $0.00
File path /workspace/brief-6-algorithm.md

Caveats (read before acting)

  1. Open-source weights are defaults; production feature switches can differ; Aug 14 README forbids naïve count arithmetic.
  2. Buffer / Ordinal / Waya corpora are not pure “small BIP startup” samples — extrapolate carefully.
  3. Link “penalty” is the most contested area: CODE says no flag; STUDIES show underperformance; CREATOR workarounds may still help for behavioral reasons.
  4. Articles / quote-boost tactics lack Ordinal-class n for small accounts.
  5. BIP “dead formats” section is cultural consensus, not Phoenix heads.
  6. No Platformer primary; no live ratioed BIP tweet IDs in this deliverable.
  7. Partial on spam/botmaker LOW_QUALITY URL rules (secondary via ickas) — full rule set not audited line-by-line here.

End brief 6 angle (a).