# The X algorithm now reads your words, not your letters: what changes in bios, names, and thread hooks in 2026

> The X ranking is now a transformer based on Grok: what changes in your name, bio, and the hook of your threads, with data from the code published in 2026.

- **Source (source):** https://aitoolsforapps.com/x-twitter-header-bio-threads-and-quote-cards-230/blog/x-algorithm-2026-bio-name-thread-hooks
- **Platform (platform):** AiToolsForApps
- **Language (language):** en
- **Updated (updated):** 2026-11-20

## In summary

- **The core of X's ranking is already a transformer derived from Grok.** The repository `xai-org/x-algorithm` was published on **January 20, 2026** under the Apache 2.0 license, and its ranking model (*Phoenix*) predicts the probability of **about 15 different actions** per post instead of applying hand-written rules ([Franz Russo](https://www.franzrusso.it/condividere-comunicare/come-cambia-algoritmo-di-x-con-grok-nel-2026/), November 2026).
- **Between 50% and 70% of the "For you" feed may come from accounts you do not follow**, so your profile is read cold for the first time, without prior context (Franz Russo, November 2026).
- **Decorative text with Unicode glyphs is a real risk**: 𝗕𝗼𝗹𝗱 and "Bold" are only the same string if the system applies NFKC compatibility normalization, something that the [Unicode UAX #15](https://www.unicode.org/reports/tr15/) standard defines but no one guarantees X does in search, indexing, or tokenization.
- **Empty formula hooks** ("threading", "this will blow your mind") do not provide any signal to the model; a hook with **concrete data + explicit promise** does describe what the post is about, which is exactly what is now interpreted.
- **The short window still rules and external links are still debated**: the distribution of a post lasts approximately **24 hours** ([Postory](https://postory.io/blog/x-algorithm-2026), 2026), but in the open source **no link penalties are coded**, contrary to what half the industry repeats.

---

For three years, optimizing for X's algorithm consisted of wrestling with a list of heuristics: banned words, weights by type of interaction, formatting tricks. In 2026 that ended. With the ranking moved to a language model, what the system evaluates is no longer your letters, but **what your words mean**: the visible name, the bio, and the first line of a thread go through a model that tries to understand what your account and your post are about. This article explores what the published code exactly says and what practical consequences it has for the text of your profile and your threads. If you want to move from theory to the piece at the end, that is the part that aiTwitter solves, the AiToolsForApps suite with [text tools for X: header, bio, threads, and quotes](/x-twitter-header-bio-threads-and-quote-cards-230).

![Translucent panel with a social network profile reduced to abstract blocks being analyzed by a luminous neural network emitting fifteen beams of light](https://cdn.aistohire.com/i/9962-vquo-cdry-ntsm-ohqc.webp)

## What exactly changed in X's algorithm in 2026?

X replaced its hand-designed feature stack with a single transformer model derived from Grok-1, called **Phoenix**, which scores each candidate post by estimating the probability of about 15 user actions and combining them into a single weighted score. The code was published on GitHub on January 20, 2026, under Apache 2.0, with a commitment to periodic public updates; the largest of these, on May 15, 2026, changed 187 files and added more than 18,000 lines (Postory, 2026).

The change had been announced: in **October 2025**, Elon Musk said that manual heuristics would disappear from the ranking in four to six weeks (Franz Russo, November 2026). Nikita Bier, then product manager, presented it as [the most important change made in X](https://x.com/nikitabier/status/2037048934015889674).

| System piece | What it does |
|---|---|
| **Home Mixer** | Orchestrates the feed assembly and applies scoring weights |
| **Thunder** | Memory store of recent posts from accounts you follow |
| **Phoenix** | The core ML derived from Grok-1: retrieval and ranking |
| **Candidate Pipeline** | Reusable framework of sources, hydration, filters, and scorers |

Source: [repository xai-org/x-algorithm](https://github.com/xai-org/x-algorithm) and analysis by [Franz Russo](https://www.franzrusso.it/condividere-comunicare/come-cambia-algoritmo-di-x-con-grok-nel-2026/) (November 2026).

Among the ~15 actions that Phoenix predicts are **positive** (like, reply, repost, quote, click, profile visit, video play, photo zoom, share, time in feed, and follow the account) and **negative** ("not interested", block, mute, and report). This is what matters for the writer: the model not only bets that you'll like the post, but also bets that you'll report it. That's why hostile or low-effort content loses reach even if it generates noise: outraged replies come with reports and "see less", and those probabilities enter the same score with the opposite sign.

Two details of the code worth knowing before blaming the text:

- Accounts with **fewer than 1,000 followers** undergo an extra anti-spam filter that larger accounts do not have, and the social proof of "people you know" requires a threshold of 1,000 (`VIEWER_FOLLOWERS_THRESHOLD`) (Postory, 2026).
- Every new post starts classified as **medium risk** until the security system clears it (Postory, 2026).
- Posting several times in a row is **exponentially attenuated** by an author diversity multiplier: the second post in the burst is worth much less and the third, almost nothing (Postory, 2026).

And one transparency limitation that must always be mentioned: X published the architecture, **not the trained weights or numerical parameters** ([adySEO](https://adyseo.com/blog/x-publica-el-codigo-de-su-algoritmo-asi-funciona/), 2026). Anyone selling you exact figures of "how much a reply is worth" is extrapolating.

## Why does a visible name and bio in decorative letters cost you reach?

Because those pretty characters are not stylish letters: they are **distinct Unicode characters**. 𝐀 (U+1D400, "mathematical bold capital A") is not "A" unless the system reading it applies compatibility normalization. The Unicode standard defines that operation —NFKC/NFKD fold font variants, "ℌ → H"— in the report [UAX #15](https://www.unicode.org/reports/tr15/), but does not require anyone to apply it: if the internal search engine, the index, or the model's tokenizer do not apply it, your brand written in decorative bold simply **is not your brand** to them.

![Two profile cards compared: the one on the left, filled with fragmented embellishments, disintegrates under the analyzing beam; the one on the right, with clean blocks, lights up](https://cdn.aistohire.com/i/4667-kpna-ywjk-ebig-ofwo.webp)

This has always been a searchability risk. What changes in 2026 is that now there is a language model trying to deduce **what your account is about** from that same text, and a jumble of mathematical glyphs tells it nothing. What is documented, with sources, is the cost of accessibility: screen readers announce these characters **by their Unicode technical name** instead of reading the letter, so a stylized phrase becomes noise for those using assistive technology ([Texas Tech University social media accessibility guide](https://www.ttu.edu/accessibility/protected/accessible-social-media/index.php)). The involved block is [mathematical alphanumeric symbols](https://en.wikipedia.org/wiki/Mathematical_Alphanumeric_Symbols), U+1D400–U+1D7FF.

| Where you place the decorative glyph | Risk | Sensible alternative |
|---|---|---|
| Full visible name (50 characters) | High: it’s your main label and the one that gets searched | Name in plain text; decoration only in a separator symbol |
| Full bio (160 characters) | High: it’s the text from which the model infers your topic | Keywords in plain text; style in 1-2 words at most |
| An embellishment or accent in the bio | Low | Acceptable: the rest remains readable and indexable |
| @username (15 characters) | None: X does not allow glyphs there | — |

The practical rule: **that 100% of your keywords are in plain text**. Decoration is typography, not vocabulary. And that’s why Unicode styling is a design decision worth considering before you paste it: in aiTwitter the bio editor offers **7 Unicode style variants** and **9 insertable separators** (`·`, `|`, `//`, `▸`, `↳`, `—`, `→`, `✦`, `•`), with a counter measuring actual glyphs with `Intl.Segmenter` —emoji and CJK count as 2— to ensure you don’t exceed the 160 characters for the bio or the 50 for the name.

## What distinguishes an empty thread hook from one that the model rewards?

An empty hook only announces that you are going to talk; a useful hook **says what you are going to talk about and what the reader will take away**. With a transformer interpreting the meaning of the text, "I’m starting a thread 🧵" provides zero information about the topic of the post, while a first line with a concrete fact and an explicit promise gives the model —and the human— material to decide.

![Two columns of glass plates representing the start of a thread: the one with hollow frames collapses and the one that supports its first plate on a concrete fact stands firm](https://cdn.aistohire.com/i/4251-zexs-ncpd-gqhl-gywo.webp)

| Hook | What it tells the model | What it tells the reader |
|---|---|---|
| "I’m starting a thread about this 🧵" | Nothing about the topic | Nothing |
| "Hello everyone, today I want to tell you something" | Courtesy without topic | That it will take time |
| "X published 18,000 lines of its algorithm in May. Here’s what changes in your bio" | Topic, entity, figure, and scope | Fact + concrete promise |

Six checks you can manually do on any hook, without magic punctuation:

1. **Appropriate length**: neither telegram nor paragraph (the useful range is between about 40 and 200 glyphs).
2. **There is a concrete fact**: a figure, a date, a proper name.
3. **There is a clear promise**: what the reader will gain by continuing.
4. **Opens a loop**: there is something left to resolve.
5. **No empty jargon**: nothing like "game changer", "disruptive", "brutal".
6. **No filler formula** at the beginning: neither "I’m starting a thread", nor "in this thread I’m going to", nor "I hope this has helped you".

That is exactly the criterion with which the aiTwitter hook review marks a start: **six boolean criteria, without numerical score**, and a veto on empty formulas with **12 patterns in Spanish and 11 in English** that apply at the beginning of the hook and warn you of what gets trimmed. It’s a writing control, not a promise of reach: no one can promise you reach.

> The model does not reward enthusiasm. It rewards understanding what you are talking about, because that’s what it needs to decide who to show it to.

## How much does the first hour really weigh?

It weighs, but less than folklore suggests and with fewer official figures than are cited. What is documented is that a post receives distribution for approximately **24 hours** after being published (Postory, 2026) and that freshness is a strong signal: a post from 20 minutes ago competes better than one from 6 hours ago ([Que.es](https://www.que.es/2026/08/04/como-funciona-el-algoritmo-de-x/), August 4, 2026). The specific figures for "the first 15 minutes" circulating on LinkedIn **are not in the published code**.

![Timeline of luminous bars: the first are high and intense and the rest plummets until extinguished, with a ring of light marking the initial window](https://cdn.aistohire.com/i/5359-qcbh-pxco-neor-ijbf.webp)

What can be deduced from the system design is more interesting than a stopwatch:

- If the distribution window is one day, **the timing of publication matters less than the hook**: a start that is not understood does not improve by being published at 9:00.
- Asking for interaction explicitly is counterproductive: X has stated that **it does not count impressions of posts that attempt to manipulate the algorithm** by instructing users to interact ([Nikita Bier](https://x.com/nikitabier/status/2083713485947748631), 2026).
- Responses in which **the author replies** receive extra weight through their own scoring (Postory, 2026). Translated: the first hour is best utilized by responding, not begging.
- And reposts clearly weigh more than "likes", which are the lowest-ranking positive (Postory, 2026); in the same vein, a post with 50 responses and 10 likes reaches further than one with 500 likes and 10 responses (Que.es, August 4, 2026).

## Are external links still penalized on X?

Here is an honest disagreement worth knowing. The majority reading in the industry holds that **external links reduce reach** and that native content takes priority, and that the tactic of "the link in the first reply" is still valid (Franz Russo, November 2026; Que.es, August 4, 2026). But the analysis of the open code states the opposite: **there is no encoded penalty for links** in the repository, and the weight of the click (`CLICK_WEIGHT`) is positive (Postory, 2026).

Both can be true at the same time, and this is the most economical explanation: if the model optimizes **time in the feed** and probability of interaction, a post that takes the user out of X competes worse in those signals without the need for any anti-link rules. It is not a punishment, it is a consequence.

| Common belief | What the published code says | What to do |
|---|---|---|
| "Links are penalized by rule" | There is no encoded penalty; the click scores positive | Stop avoiding them out of superstition |
| "You must put the link in the reply" | External tactic, not a code rule | Try it yourself: it's cheap to measure |
| "Hashtags help reach" | They are no longer relevant in the new system | Use them only as human tags |
| "Posting a lot on the same day multiplies reach" | There is exponential decay by author position | Space out the posts |

Source of the central column: Postory (2026) and Franz Russo (November 2026).

## Checklist: how to rewrite your profile and threads for the 2026 algorithm

If you are only going to do five things, do these, in this order: change the name and bio to plain text with your real keywords, rewrite the hook of your latest threads with a concrete fact at the front, stop asking for interaction, space out your posts, and reply yourself in the first hour.

1. **Visible name**: brand or real name in plain text, within 50 characters. Decoration, outside or reduced to a separator.
2. **Bio**: 160 characters that say what you do with the words people search for, not with adjectives. At most, one stylized word.
3. **Hook**: number, date, or proper name in the first line, and explicit promise. No opening formulas.
4. **Thread body**: one idea per tweet, within 280 characters, knowing that a link counts as 23 fixed characters in a tweet.
5. **Rhythm**: one good thread a day over three mediocre ones, due to the decay of author diversity.
6. **Long version**: if you have access to long posts (up to 25,000 characters), repurpose the thread as native article instead of linking out.

## What part of this work can a tool do

The mechanical part: counting glyphs accurately, respecting X's limits, styling without breaking searchability, and auditing the hook with fixed criteria. aiTwitter, from AiToolsForApps, is built exactly on that, and works **without registration** from the first visit.

- **Bio and name**: 7 Unicode styles, 9 separators, and dual counter with `Intl.Segmenter` for the 160 characters of the bio, the 50 of the name, and the 15 of the @user.
- **Threads**: AI generation in a single call (hook, body, and closing) with 4 tones and three lengths —3-5, 5-8, or 8-12 tweets—, optional numbering, and the long post derived from the same result without a second call.
- **Your own text, free**: splitting a long text into tweets, editing it, re-splitting it, or deleting pieces does not consume AI.
- **Hook review**: the 6 boolean criteria and the veto of 12 empty formulas in Spanish and 11 in English.
- **No connection with X**: the tool does not use X's API or request your account; the profile measures it shows are marked as *unverified* when not measured on a real device, and aiTwitter is not affiliated with X Corp.

In guest mode there are limits per visit (2 threads with AI and 6 files), and writing the bio with AI or generating artistic background requires an account. You can claim everything you have created as a guest later.

## Frequently Asked Questions

### Does the X algorithm read my bio?

The published code does not document "reading the bio" as an explicit step, but the ranking is a transformer derived from Grok that interprets text, and among the signals it predicts are profile visits and following from the feed (Franz Russo, November 2026). Writing a comprehensible bio is, at the very least, the cheapest hypothesis; writing it in decorative glyphs is betting that someone will normalize your characters for you.

### Do rare Unicode letters harm my reach on X?

There is no official confirmation from X on this matter. What is documented is that they are characters distinct from normal letters unless NFKC normalization (Unicode UAX #15) is applied, and that screen readers announce them by their technical name instead of reading them. That is enough reason not to put your brand name entirely in decorative bold.

### How many characters do I have for the name, bio, and a tweet on X?

50 characters for the visible name, 15 for the @username, 160 for the bio, and 280 for a tweet, with up to 25,000 characters in long posts. In a tweet, any link counts as 23 fixed characters regardless of its length.

### Where is the code for the X algorithm and what does it include?

In the public repository [xai-org/x-algorithm](https://github.com/xai-org/x-algorithm), published on January 20, 2026, under the Apache 2.0 license and primarily written in Rust and Python. It includes the architecture (Home Mixer, Thunder, Phoenix, Candidate Pipeline) but **does not** include the trained weights or the training data (adySEO, 2026; Franz Russo, November 2026).

### What percentage of the "For you" feed comes from accounts I don’t follow?

Between 50% and 70% may come from unfollowed accounts after the 2026 change (Franz Russo, November 2026); other analyses estimate it at approximately half, conditioned by your latest interactions (Que.es, August 4, 2026). In any case, most of those who read you do not know you.

### Is it useful to ask for "like and share"?

No. X has stated that it does not count impressions of posts that attempt to manipulate the algorithm by instructing users to interact (Nikita Bier, 2026). It is one of the few things that is stated clearly.

### How many times a day should I post on X?

The code applies an author diversity multiplier that exponentially dampens your subsequent posts within the same window: the second one is worth quite a bit less and the third almost nothing (Postory, 2026). Spacing out posts is better than piling them up.

## Text has become work again

The honest summary of 2026 is that formatting shortcuts have lost value and writing has gained it: with a language model deciding who sees your post, a readable name, a bio that states what you do, and a hook with a specific fact are worth more than any typographic trick. None of this guarantees reach —X has not published the weights and no one can promise you that— but it does eliminate the known ways to reduce your opportunities. For the operational part —counting glyphs, respecting limits, styling without breaking searchability, and auditing the hook— aiTwitter, from AiToolsForApps, brings together those [text, header, and thread tools for X](/x-twitter-header-bio-threads-and-quote-cards-230) in one place and can be tried without creating an account.

> Write your name, bio, and the hook for your next thread on aiTwitter, with a real glyph counter, 7 Unicode styles, and hook review: no registration and no linking your X account.

- [X (Twitter) — Measured header, a bio that actually fits, and threads with hooks](https://aitoolsforapps.com/x-twitter-header-bio-threads-and-quote-cards-230)

## About the platform

**X (Twitter) — Measured header, a bio that actually fits, and threads with hooks** — an app by AiToolsForApps. Available at: https://aitoolsforapps.com/x-twitter-header-bio-threads-and-quote-cards-230/blog/x-algorithm-2026-bio-name-thread-hooks

### Other languages

- `es`: https://aitoolsforapps.com/es/x-twitter-cabecera-bio-hilos-y-citas-230/blog/algoritmo-x-2026-bio-nombre-ganchos-hilo
- `fr`: https://aitoolsforapps.com/fr/x-twitter-banniere-bio-threads-et-citations-230/blog/algorithme-x-2026-bio-nom-accroches-fil
- `it`: https://aitoolsforapps.com/it/x-twitter-copertina-bio-thread-e-citazioni-230/blog/algoritmo-x-2026-bio-nome-ganci-thread
- `pt`: https://aitoolsforapps.com/pt/x-twitter-capa-bio-threads-e-citacoes-230/blog/algoritmo-x-2026-bio-nome-ganchos-fio
- `de`: https://aitoolsforapps.com/de/x-twitter-header-bio-threads-und-zitate-230/blog/x-algorithmus-2026-bio-name-thread-hooks
- `bg`: https://aitoolsforapps.com/bg/x-twitter-koritsa-1500-500-bio-s-realen-broyach-nishki-i-tsitati-230/blog/algoritam-x-2026-biografiya-ime-zaglaviyata-nishtki
- `ja`: https://aitoolsforapps.com/ja/x-twitter-header-bio-counter-thread-generator-230/blog/x-algorithm-2026-profile-name-thread-hook
- `nl`: https://aitoolsforapps.com/nl/x-twitter-header-1500-500-bio-met-echte-teller-threads-en-citaten-230/blog/x-algoritme-2026-bio-naam-thread-hooks
- `ko`: https://aitoolsforapps.com/ko/x-twitter-header-bio-thread-230/blog/x-algorithm-2026-bio-name-thread-hooks
- `sv`: https://aitoolsforapps.com/sv/x-twitter-header-1500-500-bio-med-riktig-raknare-tradar-och-citat-230/blog/x-algoritm-2026-bio-namn-trad-krokar
- `da`: https://aitoolsforapps.com/da/x-twitter-header-1500-500-bio-med-aegte-taeller-trade-og-citater-230/blog/x-algoritme-2026-bio-navn-traadkroge
- `no`: https://aitoolsforapps.com/no/x-twitter-toppbilde-1500-500-bio-med-reell-teller-trader-og-sitater-230/blog/x-algoritmen-2026-bio-navn-tradkroker
- `id`: https://aitoolsforapps.com/id/x-twitter-header-1500-500-bio-dengan-penghitung-nyata-utas-dan-kutipan-230/blog/algoritma-x-2026-bio-nama-pembuka-utas
- `cs`: https://aitoolsforapps.com/cs/x-twitter-zahlavi-1500-500-bio-se-skutecnym-pocitadlem-vlakna-a-citace-230/blog/algoritmus-x-2026-bio-jmeno-hacky-vlaken
- `zh`: https://aitoolsforapps.com/zh/x-twitter-header-bio-thread-tools-230/blog/x-suanfa-2026-ziji-mingcheng-chuanwen-kaichang
- `ar`: https://aitoolsforapps.com/ar/x-twitter-ghlaf-1500-500-wnbdha-badad-hqiqi-wslasl-waqtbasat-230/blog/khawarizmiyat-x-2026-nabdha-ism-khattafat-silsila
- `pl`: https://aitoolsforapps.com/pl/x-twitter-naglowek-1500-500-bio-z-prawdziwym-licznikiem-watki-i-cytaty-230/blog/algorytm-x-2026-bio-nazwa-zaczepy-watkow
- `fi`: https://aitoolsforapps.com/fi/x-twitter-1500-500-kansikuva-bio-oikealla-laskurilla-ketjut-ja-lainaukset-230/blog/x-algoritmi-2026-biografia-nimi-ketjukoukut
- `he`: https://aitoolsforapps.com/he/x-twitter-kotrt-1500-500-bio-am-monh-amiti-shrshorim-vtsitotim-230/blog/algoritm-x-2026-bio-shem-ptichim-shrashor
