Two researchers, one question: more views on X

Brief 6, run twice on September 4, 2026. The agent (Claude with web search, no X access) produced three documents: an algorithm run, a growth-cases run, and a replies-and-timing run. Weebo (a Grok bot team with X API read access) produced five: an algorithm brief, a growth-cases brief, a replies-and-timing brief, an audit of @gochaberulava (384 posts, August 5 to September 4, the only audit either side ran), and a synthesis. This document compares them; the merged playbook is a separate file.

Where the two runs quoted different numbers from the open-source ranking code, I fetched the file myself on September 4, 2026 (file header: "last sync 2026-09-03T16:23:24Z") and say which run was right. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs

Headline

1. Both runs land on the same mechanism: X's For You feed scores a post on predicted per-viewer actions, and the published defaults pay 20.0 for a copy-link share, 5.0 for a reply (plus 15.0 when viewer and author follow each other), 5.0 for a quote, 4.0 for a follow, and only 0.5 for a like. Likes are the cheapest signal; conversation and shares are what count. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs
2. Both runs rank the same formats from the same dataset: threads have the best engagement rate (1.89%), plain text the most impressions (46,731 average), link posts the least reach of the non-poll formats (9,573). https://www.tryordinal.com/blog/twitter-analytics
3. Both runs say replies to bigger accounts are the growth engine below a few thousand followers, replies must be specific and fast (first 15 to 30 minutes), and generic praise is worthless. Weebo measured it (hub reply latencies of 2 to 15 minutes; eight verified small-account replies above 10k views); the agent collected first-person creator accounts of it.
4. They differ in what they could reach: the agent had official statements (Bier, Musk, X Help) and large-n studies (Buffer 18.8M posts, 8.7M-post timing study, Metricool, Sprout); weebo had live X data (fxtwitter view counts, hub reply timestamps, the account audit, the April 2026 API link fee) plus the correct code weights.
5. The agent misquoted three low-value weights in the ranking code and, in its replies run, re-presented the 2023 heavy-ranker numbers as a reading of the 2026 release; weebo's growth list rests on unverifiable internal "Brief 1" files and single-source mirrors, and its one recommendation about the [agent] tag is built on three posts a few hours old. Neither run is usable alone; together they cover the brief.

Where they agree

Where they disagree or differ in emphasis

1. Three weights in param.rs. Agent: VideoOpenWeight 0.05, DwellWeight 0.0, VqvWeight 0.05. Weebo: VideoOpenWeight 0.07, DwellWeight 0.05, VqvWeight 0.0. Checked against the live file: weebo is right on all three. None of the three changes any conclusion (all are a tenth of a like or less), but it is a transcription error in a section the agent tagged "official" and said it had read directly. https://raw.githubusercontent.com/xai-org/x-algorithm/main/home-mixer/params/param.rs

2. The 2023 weights. The agent's algorithm run correctly flags "reply 13.5, author reply 75, bookmark 10" as March 2023 legacy numbers. Its replies-and-timing run then twice presents the same numbers as "third-party readings of the January 2026 xai-org/x-algorithm release" (hedged as "quoted from a blog, not verified"). The two documents contradict each other; the algorithm run is the correct one. Weebo's algorithm brief discards the same numbers as historical, but its replies brief slips once, explaining a reply's success with a "150x conversation weight narrative" (75 ÷ 0.5, the 2023 ratio). Weebo's slip is a parenthetical; the agent's is a section-level claim. https://opentweet.io/blog/x-algorithm-open-source-github-2026

3. Links in the post or in the first reply. Agent's algorithm run: Bier said in July 2026 "you do not need to put the links in replies anymore" and Musk replied "We haven't for over a year"; there is no 2026 test showing a reply link beats an in-post link. 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 The agent's replies run then repeats the secondary "30 to 50%, up to 80%" penalty figures and the first-reply workaround as standard practice (tagged unverified). Weebo's synthesis says "move links out of the main post when possible." Better supported: the agent's algorithm run. The code has no link demotion, the head of product says none exists, and the studies that show link posts doing worse measure engagement mix (people leave to read), not a rule. The practical rule is to write posts that stand without the link, not to hide the link in a reply, where both runs agree links do worst of all.

4. How many originals a day. Agent: 3 to 5 a day is the floor named by Graham Mann for under-1k accounts, and Buffer's top 10% post more than the median. https://grahammann.net/blog/how-to-grow-on-x-twitter-2026 and https://buffer.com/resources/state-of-social-media-engagement-2026/ Weebo: 1 to 3 "high-signal" originals a day, citing the author-diversity decay and the audit's 46-post day. Both read the same code. The agent's reading is more precise: the decay applies to the second and later posts from one author inside a single feed load ("competes with itself inside one viewer's session rather than being throttled globally"), so spacing matters more than a daily count. The audit supplies the reason to cap anyway: 22 of 60 originals had zero engagement, so quality is the constraint before quantity. Merged position: 2 to 3 originals a day, at least three hours apart.

5. Whose posts to reply under. Agent (via Teract) says avoid 500k+ mega-accounts because a small account's reply is buried. Weebo's eight verified small-account replies above 10k views all sit under mega threads (a 5.22M-view @mattshumer_ demo, a 603k-view @opencode post, a 930k-view @levelsio post), and weebo's own note says quiet solo-SaaS support threads stay in the tens to low thousands. Weebo is better supported because it measured it; the agent's rule is a vendor's folklore. But weebo also says the mega-thread replies are a lottery and recommends mid-tier ICP hubs (5k to 100k) for relationships. The merged position is weebo's: mid-tier hubs for reply-backs and profile visits, one or two early specific replies under a mega thread when one appears.

6. Timing windows. Agent synthesis from four platform-wide studies: weekday 13:00 to 16:00 UTC primary, 17:00 to 20:00 UTC secondary, avoid 22:00 to 04:00 UTC and Saturdays. https://buffer.com/resources/best-time-to-post-on-twitter-x/ Weebo from hub activity histograms: 07 to 10, 12 to 16, and 18 to 21 UTC, with several hubs (@tdinh_me, @jackfriks, @yongfook, @dannypostma) active 23 to 03 UTC, and its verified small replies landing at 18:48 to 22:55 UTC and 05:52 to 06:25 UTC. The agent's window is better supported for originals (n in the millions, though not builder-specific); weebo's is better supported for replies (it is when the specific accounts we want to reach are actually posting). The audit backs weebo's early window: the account's originals posted 06:00 to 09:00 UTC averaged 98 views (n=13) against 56 to 75 in every other three-hour bucket, and the best original (402 views) went out at 07:05 UTC. https://x.com/gochaberulava/status/2092508793842127279

7. Jack Friks' size. Agent: "~40k followers" (from a podcast transcript). https://tbf.fm/episodes/396-jack-friks-building-tools-that-empower-without-overwhelming/transcript Weebo: ~166k, with a Medium piece saying he gained over 100,000 followers in the past year. https://medium.com/@yumaueno/jack-friks-the-man-of-the-moment-who-scaled-a-couples-app-to-5-000-month-in-just-2-months-9f64d58cd9cf Weebo is better supported; the agent's number is stale.

8. Premium. Weebo's algorithm brief makes "Buy Premium" its first recommendation on Buffer's 10x gap. The agent reports the same gap but notes it is observational and that X Help only promises "a slight preference for replies from verified accounts." https://help.x.com/en/using-x/x-premium The disagreement is moot for us: the audit shows eight Article-format posts and long-form on 15 of 384, and Articles are Premium-only, so the account already has Premium.

9. Which format to lean on. Weebo's audit suggests biasing originals toward Articles because they lead the account's averages (111.5 views, n=8). The agent's only long-form evidence is a six-post test on a sub-400-follower account where threads and long-form posts came out about equal and the author concluded content mattered more than format. https://blog.hootsuite.com/experiment-x-threads-vs-longform-posts/ Neither is strong; n=8 on our side is weaker than the large-n tables both runs share, which put threads and short text first. Merged position: text and short threads first, Articles for the weekly long piece.

10. The [agent] tag. Weebo's synthesis proposes pausing or A/B-testing the tag because tagged originals averaged 18.3 views against 75.9 untagged. The agent did not see the audit. Weebo's own caveat says n=3. Looking at the three posts: all were posted between 09:47 and 10:53 UTC on September 4 and fetched around 14:02 UTC the same day, so each was three to four hours old; all three are link-format posts, the account's weakest format (59.7 average); and the untagged comparison set includes 30-day-old posts. The comparison is confounded twice over and is not evidence. It is rejected in the playbook on principle as well, since the tag is the project's premise.

11. Reply share. Weebo's audit headline is "cut reply share" (324 of 384, 84.4%). The agent's growth cases say every first-person 2025 to 2026 growth story is reply-first (JonBuildsHQ 250 replies a day; Li Zeng 80% of time commenting). https://www.indiehackers.com/post/how-i-went-from-30-to-1-000-followers-on-x-in-14-days-94fcd933e1 The audit's 324 replies over 25 active days is about 13 a day, inside both runs' recommended band. The problem is not the share; it is the craft (55 replies of three words or fewer), the targets (nine replies to @elonmusk, most at 4 to 13 views), and the burstiness (46 posts on August 25). Weebo's synthesis actually lands there too ("cap replies at 10 to 15 high-craft a day"), so this is a headline that oversells its own finding.

12. Threshold for a growth case. The agent required growth inside 2025 to 2026 and excluded Marc Lou, Pieter Levels, Jon Yongfook, and Arvid Kahl as "already large before 2025." Weebo's brief allowed a second fit ("consistently clears 50k+ views per post") and profiled exactly those accounts. Neither is wrong; they answered different briefs. For a 467-follower account, the agent's set (accounts that were small recently) is the more useful one, and weebo's set is useful for who to reply to.

What only the agent found

• Nikita Bier, April 2026: talking-head video is "an easy way to get millions of impressions with basically no followers"; reposts and content pulled from other networks face "up to a 90 percent deduction on impressions." https://piunikaweb.com/2026/04/13/x-twitter-original-talking-videos-nikita-bier/
• Bier, July 16, 2026: three engagement-solicitation violations ("I'll follow everyone who replies") lead to removal from the creator program and referral for suspension; aggregator payouts down about 80%. https://www.socialmediatoday.com/news/x-updates-its-engagement-bait-detection/825495/
• March 2026 reply downvote button, with "AI-generated" and "spam" as reasons, feeding reply ranking; Bier: "The financial incentive to spam on X will decline enormously." https://www.dexerto.com/entertainment/x-is-testing-a-dislike-button-again-and-its-coming-soon-3336926/
• Musk, April 2025: "there is no explicit rule limiting the reach of links in posts." https://x.com/elonmusk/status/1915806794393457034 Bier, October 2025: links get lower reach because "the web browser covers the post and people forget to Like or Reply," with a new collapsing link view. https://www.niemanlab.org/2025/10/x-makes-overtures-to-journalists-with-new-feature-designed-to-improve-reach-for-links/
• The pre-scoring filters: posts older than 48 hours are dropped; replies and reposts from accounts the viewer does not follow are dropped before scoring, so a small account's replies reach For You mainly through the conversation, not as standalone candidates. https://raw.githubusercontent.com/xai-org/x-algorithm/main/README.md
• Buffer's regular-account format split (text ~0.40%, video ~0.25%, image ~0.20%, link 0% median engagement, August 2025), the closest thing to a small-non-Premium-account dataset. https://buffer.com/resources/links-on-x/
• Buffer 8.7M-post timing study (Tuesday 9am, Wednesday 9 to 10am local; 6pm to 11pm worst) and Sprout's 2 billion-engagement study (12pm to 6pm Tuesday to Friday). https://buffer.com/resources/best-time-to-post-on-twitter-x/ and https://sproutsocial.com/insights/best-times-to-post-on-twitter/
• Ordinal's length finding: under 100 characters gets the most impressions (67,356 average), 100 to 200 the best engagement rate. https://www.tryordinal.com/blog/twitter-analytics
• Metricool 2024 to 2025: replies per post +21%, profile clicks −31%, link clicks −28%. https://metricool.com/x-twitter-statistics/
• Graham Mann's dashboard: one reply 12k impressions and 7 profile visits vs 400 impressions on an original; December 2025 inconsistency cost 24% engagement and net follower loss. https://grahammann.net/blog/how-to-grow-on-x-twitter-2026
• Eleven growth cases with a 2025 to 2026 window, including Natia Kurdadze's 4.47M-view "drop your URL" prompt with only 233 likes and 511 replies (the reply-weighted signature), and the cautionary set: Roy Lee's retracted revenue figure, aixbt's token-driven growth, EngageX's 500 AI replies a day, bought-follower listicles. https://superx.so/creators/natiakourdadze , https://techcrunch.com/2026/03/05/cluely-ceo-roy-lee-admits-to-publicly-lying-about-revenue-numbers-last-year/
• Audience sentiment on AI content: enthusiasm for AI-generated creator content fell from 60% in 2023 to 26% in 2025. https://www.emarketer.com/content/exclusive--ai-slop-threat-creator-economy
• Hashtags: five or more cost roughly 40% engagement (Mann). https://grahammann.net/blog/how-to-grow-on-x-twitter-2026

What only weebo found

• The account audit itself: 384 posts, 324 replies, 60 originals, format averages, 22 zero-engagement originals, the 46-post day, and every post's metrics. https://x.com/gochaberulava/status/2092508793842127279 (best original, 402 views)
• The X API URL fee: since April 20, 2026, an API post containing a URL bills about $0.20 against about $0.015 for a plain post; a billing matter, not a ranking one. https://devcommunity.x.com/t/x-api-pricing-update-owned-reads-now-0-001-other-changes-effective-april-20-2026/263025 and https://gigazine.net/gsc_news/en/20260422-x-api-link-price-up/
• Measured hub reply latency to small accounts: @tibo_maker 2.0 minutes, @floriandarroman 3.7, @jackfriks 4.8, @marclou 5.8, @steipete 18.5, @levelsio 83.5. https://x.com/tibo_maker/status/2095497835114287182 , https://x.com/marclou/status/2095803182911823889
• Eight fxtwitter-verified replies from accounts under ~500 followers with over 10k views, with timing: @SamGCoder 154,249 views at +14.6 minutes; @ewesaidit 42,151 at +3.7; @Alone_clam 21,458 at +147.5; @Chrisgormley 18,218 at +57.1 under @levelsio, who then crowned it. https://x.com/SamGCoder/status/2095599859994869872 , https://x.com/Chrisgormley/status/2094859939005046806
• Ten hub accounts with live view counts: @levelsio typical original ~336k views, 5.93 posts a day; @marclou ~69k typical, 7.5 a day; @tibo_maker 615,354 on a launch post; @arvidkahl 87,419 views with 528 replies on a question post. https://playersells.com/insights/levelsio , https://x.com/arvidkahl/status/1979936984987537724
• Growth quotes in the founder's words: Marc Lou "It took 8 months to get 1,000 followers"; Tibo "the goal isn't to go viral, it's to become familiar to a small group of people"; Rob Hallam "zero to over 34,000 followers in just over a year." https://newsletter.marclou.com/p/200k-followers-5k-mrr-5-6-body-fat , https://www.starterstory.com/robv4
• Daniel Smidstrup, 0 to 2.9k in 47 days, now ~10.6k, the clearest recent small-to-10k climb. https://www.linkedin.com/posts/daniel-smidstrup-b98390226_i-grew-from-0-to-29k-followers-on-x-in-47-activity-7461795928585236480-DEBV
• The recsys.proto boolean features (has_media but not has_link) and the "Under the Hood" labels tool. https://ickas.dev/writing/x-algorithm-update-august-2026 and https://techcrunch.com/2026/08/13/x-open-sources-its-ranking-algorithm-letting-users-see-if-theyve-been-shadowbanned/
• The build-in-public fatigue literature with named sources (Foundra, buildinpublic.so, stateofmake, infra.page). https://stateofmake.co/articles/the-trust-floor-what-build-in-public-was-actually-buying
• Monolit's 09:00 PT (16:00 UTC) indie-hacker overlap slot. https://monolit.sh/blog/best-time-to-post-twitter-x-2026-data-backed-guide-founders

Quality notes

Ranking-code weights. Identical on the fifteen weights that matter and on all five negatives. The agent misquoted VideoOpen (0.05 for 0.07), Dwell (0.0 for 0.05) and Vqv (0.05 for 0.0); weebo's table is correct on all values checked. Weebo cites the weighted_scorer.rs and author_diversity_scorer.rs files and gives the diversity formula; the agent read only README and param.rs and says so.

2023 numbers as current. The agent's replies-and-timing run does this twice (section 1 "why author replies matter" and section 2 "mechanism evidence"), contradicting its own algorithm run. Weebo does it once, in passing ("150x conversation weight narrative"). The agent's algorithm run has the best treatment of the legacy problem of any of the eight documents.

Sourcing. Agent: 93 unique URLs, 16 of them x.com posts it could not fetch (wording taken from search-result titles and press). Weebo: 558 unique URLs, 460 of them x.com posts, 384 of which are the account's own; 98 non-X URLs. Exact URL overlap between the runs: 5 (the x-algorithm repo, Buffer's Premium review, Ordinal's analytics study, Teract's reply-guy guide, Supabird's Tony Dinh piece); 7 counting the README and param.rs reached by different paths. Ten domains overlap. The two runs used almost entirely different material and reached the same conclusions, which is the strongest thing in this comparison.

Same study, two names. The agent's Adilo figures (text 3.2%, video 1.34%, "3,200+ posts") and weebo's Waya figures (text 3.24%, video 1.34%, "2,200+ posts") are almost certainly one dataset republished; they should not be counted as independent confirmation. https://adilo.com/blog/x-engagement-rate-chart-2025/ and https://wayastream.com/blog/how-text-only-posts-outperform-videos-on-x/

Unverifiable inputs. Weebo's growth-cases and replies briefs lean on sealed "Brief 1" files (hub reply densities, posting-hour histograms) that are not in the library and cannot be checked; its "typical views" numbers come from playersells.com mirrors, single-source. The agent's growth cases lean on SuperX and TwitterScore mirrors and on vendor Substacks (EngageX, SupaBird), which it flags. Both label single-source claims consistently.

Padding. The agent's growth run pads with a "named but not usable" section and a "faked it" section that are, in fact, useful. Its replies run pads with eight vendor guides in one bullet. Weebo's algorithm brief pads with hop-chain notation; its synthesis repeats the same eight numbers four times. Weebo's audit is 384 rows of table, which is not padding, it is the data.

Errors and misclassifications. Agent: the three weights; the 2023 numbers in its replies run; Jack Friks at 40k. Weebo: the audit counts six posts that begin with "@user" as originals (they are orphaned replies, weebo flags this), which slightly inflates the zero-engagement originals count; the audit calls a 25-active-day window "30 days"; the [agent] comparison is confounded by age and format; the "cut reply share" headline contradicts its own recommended reply volume. Weebo's replies brief delivered 8 of 10 requested examples and said so rather than filling the gap.

Growth-list overlap by handle. Agent profiled or named 21 accounts; weebo profiled 10 and cited 18 hubs. Overlap: @robj3d3 (full case in both); @tdinh_me, @jackfriks, @marclou, @levelsio, @yongfook, @arvidkahl (profiled by weebo, named-but-excluded by the agent as outside the 2025 to 2026 growth window); @dannypostma (weebo's replies hub list; agent's "already large" list). Eight handles. Zero overlap among the accounts that actually grew from small in 2025 to 2026: the agent has Natia Kurdadze, Roy Lee, JonBuildsHQ, Graham Mann, Li Zeng, Yasser Elsaid, Nizzy; weebo has Daniel Smidstrup and Hasan Cagli.

Budget honesty. Agent: 40 to 45 calls per run, listed the sources it did not open. Weebo: $0.37 on the replies brief, $1.39 on the audit, $0 elsewhere, stopped on a 429 and sealed as partial.

Verdict on how to use each researcher for this kind of brief

Use the agent for the official record and the large-n studies: what X has said, what the code says, what Buffer, Sprout, Metricool, and Ordinal measured, and who claimed what in the press. It reads primary text carefully, tags evidence quality consistently, and lists what it could not reach. Do not trust it on numbers it transcribes from long files without a second check, and read its documents against each other, because it does not always reconcile its own runs.

Use weebo for anything that needs X itself: the account audit, live view counts, reply timestamps, who the hubs answer and how fast, and API billing. Its data is the only thing in this library that is about our account rather than about X in general. Treat its recommendations as hypotheses: it moves from a three-post sample to a policy suggestion, and it headlines a finding its own analysis then walks back.

For a brief like this one, run both, have the agent do the code and the studies, have weebo do the audit and the live measurements, and merge with a check on any number both runs quote.

Counts

• Agreements logged: 24
• Disagreements or emphasis differences logged: 12 (weebo better supported on 5: the three weights, reply targets, Jack Friks; agent better supported on 4: links, cadence mechanism, original-post timing, format choice; moot or shared: 3)
• Agent-only findings: 13
• Weebo-only findings: 10
• Exact URL overlap: 5 (7 counting the same GitHub files via raw and blob paths); domain overlap: 10
• Account overlap by handle: 8; overlap among accounts that grew from small in 2025 to 2026: 0
• Factual errors: agent 3 (three tail weights; 2023 numbers presented as 2026 in one run; Jack Friks' follower count); weebo 1 outright (the "150x" aside) plus 3 method problems in the audit (orphan-reply originals, "30 days" for 25 active days, the confounded [agent] comparison)