Brief 6a: What X's ranking rewards and punishes in 2026 for a small account, and which formats get views

Compiled 2026-09-04. Each bullet is one claim with a category tag and a URL. Categories: official (X, xAI code, Musk or Nikita Bier statements), data study (n=...), single source (one third-party analysis, not independently confirmed), creator folklore (unsourced advice repeated across growth blogs).

Important caveat up front: many 2026 growth articles still quote the 2023 heavy-ranker weights (reply 13.5, reply-with-author-response 75, profile click 12, bookmark 10, repost 1, like 0.5). Those numbers are from the March 2023 open-source release and are not what the 2026 code ships. The weights below in section 1 come from the current xai-org/x-algorithm repository (fetched 2026-09-04).

1. Ranking signals

The 2026 open-source algorithm (Phoenix / Grok-architecture)

• official: xAI re-open-sourced the For You feed in January 2026 as a Grok-architecture transformer ("Phoenix"); the code became runnable end-to-end with a downloadable mini model on May 15, 2026 (commit e414c17), and the training code was published August 13, 2026. https://github.com/xai-org/x-algorithm
• official: Phoenix "predicts how likely the viewer is to take each action on each post" and the feed score is "Final Score = Σ (weight_i × P(action_i))"; the README stresses the weights "scale the predicted probabilities ... they do not scale the raw engagement counts", so a post is not scored on its like or reply count but on how likely each viewer is to reply, like, share, etc. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• official: predicted actions are: favorite, reply, repost, quote, share, share via DM, share via copy link; post click, profile click, link click, photo expand, video open, quoted-post click; video quality view, dwell, dwell time, click dwell time, active seconds; follow author; and negatives: not interested, mute author, block author, report, not dwelled. https://github.com/xai-org/x-algorithm
• official: current default weights in home-mixer/params/param.rs (positive): ShareViaCopyLinkWeight 20.0, BidirectionalFollowReplyWeightBoost 15.0 (extra weight on a reply when viewer and author follow each other), ReplyWeight 5.0, QuoteWeight 5.0, ShareViaDmWeight 5.0, FollowAuthorWeight 4.0, ShareWeight 2.0, RetweetWeight 1.0, FavoriteWeight 0.5, ClickWeight 0.4, OpenLinkWeight 0.2, PhotoExpandWeight 0.05, VideoOpenWeight 0.05, VqvWeight 0.05, QuotedClickWeight 0.05, PostUnexploredWeight 0.02, ContDwellTimeWeight 0.004. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
• official: current default weights (zeroed or negative): ProfileClickWeight 0.0, DwellWeight 0.0, ContClickDwellTimeWeight 0.0, QuotedVqvWeight 0.0, BidirectionalFollowDwellWeightBoost 0.0; NotDwelledWeight -0.02, BlockAuthorWeight -31.2, NotInterestedWeight -43.2, MuteAuthorWeight -58.8, ReportWeight -234.0. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
• official: reading of the above for a small account: a predicted reply is worth 10x a predicted like and 5x a predicted repost; a predicted "share via copy link" is worth 40x a like; a predicted profile click is worth nothing in the published defaults; binary dwell is worth nothing, only continuous dwell seconds count at 0.004 per unit, so dwell is a weak positive and "not dwelled" a tiny negative. One predicted report (-234) outweighs roughly 47 predicted replies. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
• official: bookmarks do not appear anywhere in the published weight list or the README's action list, so the widely repeated "bookmark = 10 points" is a 2023 legacy number, not a 2026 one. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
• official: three post-scoring adjustments: Author Diversity ("each post after an author's first is multiplied by a decaying factor, down to a floor" — AuthorDiversityDecay 0.5, AuthorDiversityFloor 0.25), Out-of-Network Discount ("posts from accounts the viewer does not follow are multiplied by a factor below 1, as are replies and reposts from accounts the viewer does follow" — OonWeightFactor 0.75, TopicOonWeightFactor 0.5), and New-Author Boost ("posts from authors whose impressions are below a threshold are lifted toward a target position"). https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• official: pre-scoring filters include AgeFilter ("Posts older than 48 hours"), OONRetweetReplyFilter ("Reposts and replies from accounts the viewer does not follow"), NewUserMinEngagementFilter ("For new accounts, out-of-network posts below an engagement threshold"), MutedKeywordFilter, PreviouslySeenPostsFilter, and InventoryHoldoutFilter. There is no link, URL, or media filter in the published filter list. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• official: what is withheld: "Grox classifier prompts" (the LLM content classifier) and "some botmaker rules" are excluded "to reduce the risk of gaming"; X points users to its "Under the Hood" tool to see labels affecting their own account's visibility. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• single source: the January 2026 release was readable but not runnable and shipped no pretrained model; the May 15, 2026 update added phoenix/run_pipeline.py (retrieval and ranking in cascade), a mini Phoenix checkpoint (~3 GB, 256-dim, 4 heads, 2 layers) and ad blending, with a promised four-week release cadence. Production weights for the full Phoenix model are still not published. https://pasqualepillitteri.it/en/news/2594/x-algorithm-open-source-may-2026
• single source: the May 2026 commit's main change was "candidate isolation" (a post's score no longer depends on which other posts are in the same batch) plus exposure of the Grox classification and brand-safety layers; the commit did not change numeric weights. https://medium.com/@furkangozukara/xs-algorithm-update-is-bigger-than-it-looks-a-deep-dive-into-the-latest-x-algorithm-commit-6bbdab3bb7db

Grok-based ranking statements

• official: Musk, October 2025: "The X recommendation system is evolving very rapidly. We are aiming for deletion of all heuristics within 4 to 6 weeks" and "Grok will literally read every post and watch every video (100M+ per day) to match users with content they're most likely to find interesting"; he framed the goal as fixing the case where creators "post something great, but nobody sees it". https://www.socialmediatoday.com/news/x-formerly-twitter-switching-to-fully-ai-powered-grok-algorithm/803174/
• official: Musk, April 2025: "there is no explicit rule limiting the reach of links in posts"; the algorithm tries to maximize user time on X, so a link with no context does worse than a link with context. https://x.com/elonmusk/status/1915806794393457034

External links

• official: Nikita Bier, October 2025: "For creators, a common complaint is that posts with links tend to get lower reach. This is because the web browser covers the post and people forget to Like or Reply"; he denied links are deboosted and announced a new in-app link experience where "posts will now collapse to the bottom of the page so people can react while you're reading". https://www.niemanlab.org/2025/10/x-makes-overtures-to-journalists-with-new-feature-designed-to-improve-reach-for-links/
• official: Nikita Bier to Mark Zuckerberg, July 29, 2026: "Hello Mark, you do not need to put the links in replies anymore"; Musk replied "We haven't for over a year", i.e. X says it stopped penalizing link posts around mid-2025. https://www.freepressjournal.in/tech/x-product-head-nikita-bier-confirms-link-penalty-removed-over-a-year-ago-tells-mark-zuckerberg-he-can-post-them-directly
• official: the 2026 code gives a predicted link open a small positive weight (OpenLinkWeight 0.2, less than a like's 0.5) and contains no link-specific filter or demotion; any remaining gap for link posts comes from lower predicted engagement, not from an explicit rule. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
• data study (n=18.8M posts, 71k accounts, Aug 2024 to Aug 2025, published Oct 14, 2025): link posts had the lowest engagement of any format; for regular (non-Premium) accounts, link posts' median engagement rate was 0% in August 2025 (text ~0.40%, video ~0.25%, image ~0.20%), while for Premium accounts link posts were ~0.28% versus text ~0.90% and video ~0.85%. Buffer notes "Since March 2025, link posts from non-Premium accounts have been effectively invisible". https://buffer.com/resources/links-on-x/
• data study (n=18.8M X posts, 2025 data): across all accounts, median engagement rate by format: text 3.56%, image 3.40%, video 2.96%, link 2.25%; "posts with links came in last". https://buffer.com/resources/data-best-content-format-social-media/
• data study (n=87,528 posts, Jan 2024 to May 2026): link posts averaged 9,573 impressions and 0.68% engagement, versus 46,731 impressions for text-only posts. https://www.tryordinal.com/blog/twitter-analytics
• data study (n=1,123,528 X posts, 15,116 accounts, 2024 vs 2025): "link clicks fell 28%" year over year while organic impressions dipped 5%, so even when a link post is shown, fewer people click through. https://metricool.com/press-release-2026-social-media-study/
• creator folklore: "put the link in the first reply" is still recommended by growth blogs as a workaround (Buffer's own October 2025 piece calls it "not foolproof"); X's leadership said in July 2026 it is no longer needed, and there is no 2026 controlled test showing a reply link outperforms an in-post link. https://buffer.com/resources/links-on-x/

Media vs text, threads, video

• official: in the published weights, media interactions are almost worthless on their own (PhotoExpandWeight 0.05, VideoOpenWeight 0.05, VqvWeight 0.05, i.e. one tenth of a like); media only helps if it raises predicted replies, shares, quotes, follows or dwell seconds. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
• official: Nikita Bier, April 2026: "People should make more talking videos like this on X. Easy way to get millions of impressions with basically no followers", and the platform wants "net new content on the app". https://piunikaweb.com/2026/04/13/x-twitter-original-talking-videos-nikita-bier/
• official: Nikita Bier, April 2026: "Reposts or anything pulled from a third-party network now face up to a 90 percent deduction on impressions"; copy-paste aggregation "just steal[s] impressions from the real creators". https://piunikaweb.com/2026/04/13/x-twitter-original-talking-videos-nikita-bier/
• data study (n=18.8M X posts): text 3.56% vs image 3.40% vs video 2.96% median engagement rate; the text-to-image gap "narrowed to just 5%" versus earlier years. https://buffer.com/resources/data-best-content-format-social-media/
• data study (n=3,200+ posts, ~29,000 data points, published Feb 2026): average engagement rate text-only 3.2% (median 1.13%), image 2.09% (median 1.28%), video 1.34% (median 0.94%), promoted 1.36%; text-only outperformed video despite being 80% less frequent in the sample. https://adilo.com/blog/x-engagement-rate-chart-2025/
• data study (n=87,528 posts, Jan 2024 to May 2026): engagement rate and average impressions by format: threads 1.89% / 27,508; images and GIFs 0.99% / 13,053; polls 0.89% / 3,634; carousels 0.72% / 12,335; links 0.68% / 9,573; text-only 0.67% / 46,731; video 0.53% / 30,684. Text-only posts averaged "35,975 impressions vs. 18,334 for media" while media converted slightly better (0.79% vs 0.68%). https://www.tryordinal.com/blog/twitter-analytics
• data study (n=6 posts, one account under 400 followers, updated Jan 2025): three threads totaled 10,783 impressions (247 to 8,155 each) and three Premium long-form posts on the same material totaled 11,846 (169 to 10,889 each); the author's conclusion was "the content of your post matters more than the format" and that reposts drove the outliers. https://blog.hootsuite.com/experiment-x-threads-vs-longform-posts/
• single source: "Thread completion rate became a ranking signal in 2025" and threads "generate highest dwell time" (OpenTweet); the published code has no thread-completion parameter, only continuous dwell seconds. https://opentweet.io/blog/x-algorithm-open-source-github-2026

Articles (long-form)

• official: "Publishing Articles is a feature limited to Premium subscribers" (Premium and Premium+, not Basic); only Premium tiers can post over 280 characters. https://help.x.com/en/using-x/x-premium
• official: X ran a $1M prize for the top Article of January 2026 (winners announced February 4, 2026, with $250k and $100k runner-up prizes), judged on impressions from verified users, as a push for long-form content; X has not published any statement that Articles get a feed ranking boost. https://www.socialmediatoday.com/news/x-awards-1-million-to-top-article-for-january/811396/
• single source: Articles support up to 25,000 characters, are search-indexed, and the contest required 1,000+ words and a US Premium account; the source concedes "No algorithm code shows long-form posts perform better than threads". https://www.socialync.io/blog/x-twitter-articles-strategy-2026
• official: there is no Article-specific weight or filter in the published home-mixer parameters; Articles are ranked as posts. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md

Premium / verified boost

• official: X Help: "people on X will see a slight preference for replies from verified accounts over other replies" and X is "currently testing the levels at which we prioritize content from Premium subscribers". https://help.x.com/en/using-x/x-premium
• data study (n=18.8M posts, 71k accounts, Aug 2024 to Aug 2025, published Oct 2, 2025): median impressions per post were "typically under 100" for regular accounts, "closer to 600" for Premium, and "over 1,550" for Premium+, i.e. "around 10x the median reach"; regular accounts' median engagement rate fell to 0% by March 2025 while Premium tiers sat at 0.49 to 0.55%. Observational, not a controlled test. https://buffer.com/resources/x-premium-review/
• data study (n=18.8M X posts): Buffer's 2026 engagement report repeats that "Regular account engagement subsequently declined to a median of 0%" after a January 2025 divergence, with Premium accounts improving. https://buffer.com/resources/state-of-social-media-engagement-2026/

Replies, engagement bait, penalties

• official: replies carry ReplyWeight 5.0 plus a BidirectionalFollowReplyWeightBoost of 15.0 when viewer and author follow each other, so replies inside a mutual-follow graph are the single strongest non-share positive signal (20 total vs 0.5 for a like). https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
• official: the OONRetweetReplyFilter drops "Reposts and replies from accounts the viewer does not follow" before scoring, so a small account's replies are shown in For You mainly to that account's own followers and via conversation threads, not as standalone out-of-network candidates. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• official: July 16, 2026: Nikita Bier said engagement-solicitation violations (example given: "I'll follow everyone who replies") "3 or more times will results in removal from the program and your account will be forwarded to the policy team for suspension"; X used Grok to detect it and reported "1.5 million posts that were stolen this cycle" with aggregator payouts down about 80% year over year. https://www.socialmediatoday.com/news/x-updates-its-engagement-bait-detection/825495/
• data study (n=18.8M X posts): accounts that reply to comments on their own posts saw "+8% engagement", the weakest effect among platforms studied. https://buffer.com/resources/state-of-social-media-engagement-2026/
• data study (n=1,123,528 X posts, 15,116 accounts): 2025 vs 2024 per post: replies +21% (2.56), retweets +35% (6.67), likes +8% (32.89), profile clicks -31% (5.68); engagement rate rose from 1.32% to 1.58% while impressions per post fell 5% to 2,711. https://metricool.com/x-twitter-statistics/

Posting frequency

• official: there is no posting-frequency penalty in the published code; what exists is Author Diversity in a single feed load (second post from the same author x0.5, floor x0.25), so a burst of posts competes with itself inside one viewer's session rather than being throttled globally. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• data study (n=18.8M X posts): "Top 10% performers post more frequently than median users"; Buffer's guide suggests three to five posts per day and a Tuesday 9 a.m. peak with weekday 6 to 11 a.m. windows strongest. https://buffer.com/resources/state-of-social-media-engagement-2026/
• data study (n=70M posts across four platforms, Jan 2024 to Dec 2025): X accounts averaged 70 posts per month in 2025 (up 40%), while engagement rate by followers fell from 0.15% (2024) to 0.10% (Q4 2025) and average views per post rose 50% to 2,979 with likes down 62% to 15 per post. https://www.socialinsider.io/social-media-benchmarks

For You vs Following

• official: the open-sourced code is the For You feed only ("Algorithm powering the For You feed"); Following is the reverse-chronological feed of followed accounts and is not ranked by Phoenix. https://github.com/xai-org/x-algorithm
• official: For You also sources out-of-network posts via SimClusters and embedding retrieval, then discounts them (OonWeightFactor 0.75); posts older than 48 hours are dropped, so a small account's window for out-of-network reach on any post is two days. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• single source: Social Media Today's caveat that users who default to Following never see the algorithmic improvements Musk described. https://www.socialmediatoday.com/news/x-formerly-twitter-switching-to-fully-ai-powered-grok-algorithm/803174/

Legacy numbers still circulating (flagged so they are not mistaken for 2026 facts)

• creator folklore: "reply = 27x a like, conversation = 150x a like, bookmark = 20x, profile click = 24x, link click with 2-minute dwell = 22x" — these are 2023 heavy-ranker weights recycled in 2026 guides (OpenTweet, Buffer's algorithm guide, Indie Hackers). They do not match the 2026 param.rs. https://opentweet.io/blog/x-algorithm-open-source-github-2026
• creator folklore: "TweepCred must exceed 65 or only 3 posts are considered" — from the 2023 codebase; TweepCred is not referenced in the 2026 README or filter list. https://pasqualepillitteri.it/en/news/2594/x-algorithm-open-source-may-2026
• single source: OpenTweet's claim that links carried a "20 to 30% reach reduction" in 2023 growing to "30 to 50%" by 2025 has no cited measurement. https://opentweet.io/blog/x-algorithm-open-source-github-2026

2. Format evidence for small accounts

• data study (n=87,528 posts, Jan 2024 to May 2026; no account-size split): highest engagement rate: threads 1.89%; highest average impressions: text-only 46,731 and video 30,684; lowest reach: polls 3,634 and links 9,573. Best-length band for engagement was 100 to 200 characters (1.09%); under 100 characters had the most impressions (67,356 average) at 0.68%; over-280-character (Premium) posts 0.74%. Median engagement across the sample 0.70%; above 1% "considered strong". https://www.tryordinal.com/blog/twitter-analytics
• data study (n=18.8M X posts, 2025): text 3.56%, image 3.40%, video 2.96%, link 2.25% median engagement rate; X's overall median engagement rate ~2.5% (up 44% from ~2.0% in 2024) but with a Premium/regular split where regular accounts' median is 0%. https://buffer.com/resources/state-of-social-media-engagement-2026/
• data study (n=18.8M posts, 71k accounts): for regular accounts specifically (the closest proxy for a small non-Premium builder account), August 2025 median engagement rates were text ~0.40%, video ~0.25%, image ~0.20%, link 0%. https://buffer.com/resources/links-on-x/
• data study (n=3,200+ posts, Feb 2026): median engagement image 1.28% > text 1.13% > video 0.94%; averages flip to text 3.2% > image 2.09% > video 1.34% because text has more extreme outliers ("some text posts exceeded 400% engagement"). Polls, threads, quotes and links were not covered. https://adilo.com/blog/x-engagement-rate-chart-2025/
• data study (n=6 posts, one sub-400-follower account): threads and Premium long-form posts were roughly equal (10,783 vs 11,846 total impressions); best single post was a long-form post at 10,889 impressions and 7.03% engagement, best thread 8,155 at 4.72%. https://blog.hootsuite.com/experiment-x-threads-vs-longform-posts/
• data study (n=1,123,528 X posts): the "typical" X account in Metricool's sample gets 2,711 impressions per post, 32.89 likes, 6.67 retweets, 2.56 replies and 5.68 profile clicks; there is no format or account-size split. https://metricool.com/x-twitter-statistics/
• data study (n=70M posts, four platforms): X average views per post 2,979, likes 15, comments 1, shares 1, engagement rate 0.10%; no format split. https://www.socialinsider.io/social-media-benchmarks
• official: talking-head video is the one format X's head of product has explicitly said gets "millions of impressions with basically no followers" (April 2026). https://piunikaweb.com/2026/04/13/x-twitter-original-talking-videos-nikita-bier/
• single source: an Indie Hackers post (June 3, 2026) cites one unnamed creator who went from zero followers to "560,000 impressions and an engagement rate 22x higher than average" in 10 days by replying early to viral posts in their niche; no dashboard or verification. https://www.indiehackers.com/post/zero-followers-high-reach-how-to-get-impressions-on-x-in-2026-1dd99f725d
• creator folklore: "tweets with images get 30 to 80% higher engagement than text-only" and "short video hooks (first 3 seconds) can generate 5 to 10x more early impressions" (Indie Hackers, June 2026, unsourced; contradicted on images by Buffer, Adilo and Ordinal, which all put text at or above images on engagement rate). https://www.indiehackers.com/post/zero-followers-high-reach-how-to-get-impressions-on-x-in-2026-1dd99f725d
• not found: no 2025–26 study that isolates "screenshot of a dashboard or revenue number" as a format; the closest is the image category in the studies above, and the code treats a photo expand as worth 0.05. No 2025–26 study reports quote-post reach separately (the code weights a predicted quote at 5.0, equal to a reply). https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs

3. Hooks

• data study (n=87,528 posts, Jan 2024 to May 2026): posts under 100 characters got the most impressions (67,356 average), 100 to 200 characters the highest engagement rate (1.09%), so short openers reach further and mid-length posts convert better; this is the only 2026 dataset found that measures length directly. https://www.tryordinal.com/blog/twitter-analytics
• official: because ranking is on predicted per-viewer actions, the first line matters only insofar as it raises predicted reply, share, quote or dwell seconds; there is no first-line, question-mark, or number feature in the published parameters. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• creator folklore: "Short first lines, under 10 words, perform 3x better than long ones" and "posts with concrete numbers consistently outperform vague openings" (Indie Hackers, June 2026; no sample or source given). https://www.indiehackers.com/post/zero-followers-high-reach-how-to-get-impressions-on-x-in-2026-1dd99f725d
• creator folklore: "one idea per line" and line-break formatting advice appears in 2026 thread-writing guides (Teract, Autotweet, Posteverywhere) without any measurement; no 2025–26 data study on line breaks was found. https://www.teract.ai/resources/twitter-thread-writing-2026
• single source: Teract's claim that it "tested 2,000+ tweets in Q1 2026" and analyzed "10,000+ tweets across 500 accounts" reports no methodology or per-hook numbers and restates the 2023 "27x" weight, so it is not usable as evidence. https://www.teract.ai/resources/twitter-algorithm-2026

4. Claim tagging

Every bullet above carries its tag at the start of the line. Summary of how the tags were applied:

• official: xai-org/x-algorithm README and param.rs (fetched 2026-09-04), help.x.com, and press-reported direct quotes from Elon Musk or Nikita Bier.
• data study (n=...): Buffer (18.8M X posts / 71k accounts, three reports), Ordinal (87,528 posts), Metricool (1.12M X posts / 15,116 accounts), Socialinsider (70M posts, four platforms), Adilo (3,200 posts), Hootsuite (6 posts, one sub-400-follower account).
• single source: pasqualepillitteri.it, furkangozukara on Medium, OpenTweet, Socialync, Social Media Today caveat, Indie Hackers anecdote, Teract.
• creator folklore: links-in-replies, 2023 weights presented as current, TweepCred 65 threshold, image and video multipliers, hook-length and numbers-in-hook rules, one-idea-per-line.

Exclusions and limits

• x.com post URLs are blocked to the fetcher (robots.txt), so Musk's April 2025 link statement and Bier's October 2025 and July 2026 posts are cited via the tweet URL plus press coverage (Nieman Lab, Free Press Journal, AdTechRadar, Social Media Today); wording for the April 2025 tweet is paraphrased from search-result text rather than a fetched copy.
• github.com tree pages are blocked to the fetcher; README.md and home-mixer/params/param.rs were read from raw.githubusercontent.com. I did not read the Rust source for the New-Author Boost threshold, the impression threshold, or the NewUserMinEngagementFilter threshold, so those numbers are not reported.
• The published param.rs values are defaults for the open-source build; the README says production Phoenix weights and the Grox classifier prompts are not published, so real production weights may differ.
• Metricool's full 2026 PDF was blocked (robots.txt); only its press release and X statistics page were used, which lack a by-format split.
• Socialinsider's benchmark page has no X-only sample size or format split.
• wildandfreetools.com (tweet length 2026) failed with a redirect loop.
• No 2025–26 study was found that splits X results by follower tier, so "small account" evidence is limited to Buffer's regular-vs-Premium split, Hootsuite's single sub-400-follower test, and Bier's remark about talking videos.
• No 2025–26 study measured quote posts, dashboard screenshots, or bookmarks by format on X. Bookmarks are absent from the 2026 weight list, so 2023-era bookmark advice could not be confirmed.
• Buffer's engagement rate for 2026 (3.56% text etc.) and its 2025 links piece (0.40% text for regular accounts) use different bases and time windows; both are reported as published.
• Search budget was 40 calls; 44 were used (three failed fetches). Sources I did not open: Sprout Social's algorithm article, SocialPilot, Posteverywhere, Teract's growth guides beyond the two above, momentumhive "Articles vs threads", OpenTweet "10 secrets", xDaily's repost of Musk's April 2025 statement.

Counts

• Claims by category (counting the tag at the start of each bullet, including the four "tagging summary" lines in section 4): official 31; data study 23; single source 9; creator folklore 7; plus 1 "not found" bullet. Total 71 bullets; 67 of them are evidence claims in sections 1 to 3.
• Sources checked (opened or attempted): 29 URLs across 16 searches; 4 fetches failed (x.com status, github.com tree, metricool PDF, wildandfreetools redirect loop).
• Sources used (cited at least once): 28 unique URLs — github.com/xai-org/x-algorithm, raw README.md, raw param.rs, help.x.com X Premium, x.com/elonmusk/status/1915806794393457034, socialmediatoday.com (Grok algorithm; engagement bait; $1M Article), niemanlab.org, freepressjournal.in, piunikaweb.com, buffer.com (state of engagement 2026; best content format; links on X; X Premium review), tryordinal.com, metricool.com (press release; X statistics), socialinsider.io, adilo.com, blog.hootsuite.com (threads vs long-form), pasqualepillitteri.it, medium.com/@furkangozukara, opentweet.io, socialync.io, indiehackers.com, teract.ai (two pages).