Ten patterns that quietly tank your account.
Each anti-pattern is ranked by directional cost to reach. Most are not visible until you have accumulated weeks of them — which is why they tank accounts that look identical to high-performing ones from the outside.
Outbound link in the main tweet body
An outbound URL in the parent tweet is the single most replicated deboost from the 2023 open-source release. The fix is mechanical: hook in the parent, link in the first reply. The deboost survives the 2026 rewrite because attention-leaving content lowers P(dwell) and P(reply) simultaneously.
The engagement throat-clear
"🧵 1/ Excited to share my new project — strap in." The first 9 words must do work, not perform enthusiasm. The transformer reads first-line legibility as the primary signal of whether dwell time will exceed 2 seconds. "Excited to share" is the literal canonical example of throat-clearing.
Hashtag stacking
"#crypto #web3 #defi #ai #buildinpublic" reads to the ranker like spam-pattern content and to the reader like a desperate signal. One project-specific tag is the maximum. The mechanic: more than one hashtag triggers a heuristic in the spam classifier inside Grox, which pushes P(report) and P(not_interested) up.
Tagging celebrities for reach
An unsolicited @-mention of a famous account does not push your tweet to their followers; the algorithm has been calibrated against this since 2018. What it does push is your tweet to your own followers as "a tweet directed at someone else" — i.e. into a thinner candidate pool. Tag only when the mentioned account would plausibly reply.
Screenshot of plain text
An image of a paragraph of text gets the image boost on dwell time and photo_expand, but is read by Grox as "text masquerading as image" — which triggers a mild deboost as of 2024–2026 (the policy was added to discourage screenshotted blocked tweets). The fix: if the content is text, post it as text. If the image is data, a chart, or a real screenshot of a third-party UI, the boost holds.
Controversy bait without payoff
Highly polarizing tweets earn a P(reply) spike that the ranker rewards in the short term. They earn an equally high P(not_interested), P(mute_author), and P(block_author) over the following 48 hours that the ranker punishes for far longer. Net effect: a single bait-tweet can knock your distribution down for two weeks. The pattern is the most common cause of inexplicable account-wide reach drops.
Follow-for-follow and engagement pods
Coordinated reciprocal engagement looks to Phoenix's two-tower retrieval like a tightly self-citing cluster — which the ranker reads as "low information value" because the cluster does not export embeddings to the broader graph. Engagement pods worked in 2018; they have been a net negative since 2022 and a stronger net negative under the 2026 transformer ranker.
Post and delete cycling
Deleting tweets within hours of posting registers as a spam-like behavior pattern. The 2023 spam classifier weighted this as a soft signal; the 2026 Grox classifier appears to weight it harder. Posters who delete every tweet that does not perform will, after several months, see baseline reach decline by ~20–30%.
Thread padding past T8
Each tweet in a thread is scored independently. Tweets 9+ in a thread routinely score below the in-network display threshold, which means the thread visibly truncates for most readers. Two crisp threads of 6 tweets each beat one 14-tweet thread in nearly every metric. The pattern: if your thread cannot end at T6–T8, it should be two threads.
Generic engagement-bait questions
"What's your favorite founder advice?" / "Drop your handle below." These questions get replies and therefore an initial ranker lift, but accumulate P(mute_author) from sophisticated readers (the ranker has been calibrated against generic-question patterns since 2023). Specific questions tied to your actual claim ("who is currently churning above 8%?") perform better and do not accumulate mute signal.