Sentence length, not post length, predicts AI content reach

Across SocialNexis-managed accounts, two posts with the same character count, topic, and posting time pull different reach inside the first 60-minute window. The variable that splits them is sentence-length variation. Posts above 0.55 burstiness clear the bar for continued distribution. Posts below 0.35 stall.

Sentence length variation predicts LinkedIn AI content reach more than post length

Start with the finding, because it reframes everything that follows. On the accounts SocialNexis manages, the strongest predictor of whether a post spreads is not how long it is. It is how much the sentence lengths vary inside it. LinkedIn's feed ranking measures dwell time, the seconds a member spends viewing a post, as a primary positive signal. Posts that members scroll past in a beat get classified as skipped updates and handed a negative ranking score, which pulls distribution down directly.

We see the effect cleanly in controlled tests. Holding post length, topic, and posting time constant, and changing only the sentence rhythm, the bursty variants averaged 40 to 60 percent more impressions in the first 24 hours. Same argument. Same character count. Same slot in the day. The only difference was that one version mixed short sentences with long ones, while the other marched in near-identical lengths from open to close.

So the operative lever is sentence-length variance, not character count. A post of a given length can stall in its first cohort or travel well past it, and the thing that moves it is whether the sentences alternate. This is the part most length-focused advice misses. You can hit the recommended character band and still write every sentence at the same width, and the ranking model reads that flatness as a quality problem.

The readability numbers point the same way. In our own corpus, posts with sentences averaging 10 to 12 words scored about 30 percent higher on readability, and single-line paragraphs lifted reading time by up to 20 percent. Both feed dwell time scoring in LinkedIn's 2026 model. Short sentences and white space are not stylistic preferences here. They are inputs to the metric that decides reach.

Length still matters, but as a ceiling rather than a lever. In AuthoredUp's dataset of 372,126 posts from September 2025 through February 2026, the 2,001 to 2,500 character band peaked at a 2.67 percent engagement rate, the highest of any length bucket. That is the post-length optimum. Whether a given post reaches that ceiling or stalls below it is decided by rhythm. In our data, posts scoring above 0.55 burstiness consistently clear the 2 percent engagement threshold needed for continued distribution inside the first 60-minute window. Posts below 0.35 routinely stall at the first-pass cohort and get no further amplification.

We have a name for the failure mode internally: the flat-rhythm stall. A post goes out, picks up a handful of early views, then simply stops. Reach flatlines within the hour. When we pull those posts and measure them, the pattern is almost always the same low sentence-length variance, regardless of how strong the underlying point was. The idea was fine. The rhythm never gave a reader a reason to slow down, so dwell time stayed low, and the model read low dwell time as low value.

The rest of this guide covers how to measure that variance, the readability benchmarks that sit alongside it, and the rewrite we run before anything publishes. None of the mechanism is mysterious once you stop thinking in word counts and start thinking in rhythm.

Uniform sentence rhythm is the AI signal LinkedIn's 360Brew model scores

LinkedIn does not run a binary AI detector on feed posts. There is no green check or red flag that says human or machine. Its 360Brew model, a system with 150 billion parameters deployed in 2025, scores probabilistic quality signals instead, including lexical diversity and sentence-pattern uniformity, and treats those as proxies for low-value content. The output is a score, not a verdict.

Because the model issues a score rather than a label, the consequence is quieter than a ban and harder to notice. A post that shows statistically uniform sentence rhythm gets a depressed quality score, which trims its first-pass distribution regardless of topic or how much authority the author carries. You did not break a rule. The system just decided your post was probably not worth surfacing widely, and it did so before a single human reaction came in.

That suppression is invisible to the author and it compounds. Lower first-pass reach means fewer early likes, comments, and dwell signals. Fewer early signals mean the post never qualifies for re-amplification into broader audiences. The post does not get penalized in any way you can see in your notifications. It simply never leaves the runway, and you are left guessing why a piece you liked underperformed.

AI-text detection runs on two measurable signals: perplexity, which measures how predictable each next word is, and burstiness, which measures sentence-length variation. AI text tends to score low on perplexity, meaning predictable, and low on burstiness, meaning uniform. Here is the wrinkle that matters for 2026. As of 2025, fine-tuned large language models can match human perplexity levels, which makes perplexity unreliable on its own. Burstiness is the more operationally durable of the two remaining signals, because sentence rhythm is harder for a default model to fake.

This is why SocialNexis's pipeline targets a burstiness score above 0.40 as a pre-publish gate. The check runs before any publish step. If a draft scores below that threshold, the system automatically rewrites it, inserting shorter punchy sentences between the longer analytical ones until the variance clears the gate. The rewrite happens at generation time, not as a manual edit a human remembers to do later. That difference, a gate versus a good intention, is most of what separates AI content that distributes from AI content that gets quietly suppressed.

The practical takeaway is that you are not trying to trick a detector. You are writing for a quality model that scores rhythm. Give it varied rhythm and it scores the post as the kind of writing a person spends time on. Give it a flat line of same-width sentences and it scores the post as the kind of writing people skim and skip.

What is the burstiness score, and why does it predict LinkedIn AI content reach?

Burstiness is the standard deviation of sentence lengths divided by the mean for a given text. That is the whole formula. Count the words in each sentence, find the average, find the spread around that average, and divide the spread by the average. Human writing typically scores 0.60 to 1.00 and higher. AI-generated text typically scores 0.15 to 0.30. The gap is wide because language models default to statistically uniform sentence structures, while people naturally vary their rhythm without trying.

The mechanism behind that gap is straightforward once you watch how a model writes. It predicts the next token from a learned distribution, and that process pulls sentence length toward a comfortable middle. Mixing short punchy sentences with longer complex ones is the marker of natural, engaging writing that holds attention. A flat sequence of medium sentences does the opposite. It reads as competent and forgettable, and readers move on quickly, which is exactly what dwell time scoring punishes.

An analysis of 200 academic text samples found burstiness was the single clearest signal separating human from AI-generated writing, more reliable than vocabulary analysis or perplexity alone. GPTZero, whose detector was built on perplexity and burstiness, documents the same gap from the practitioner side: human writing scores 0.60 to 1.00 and higher while AI text scores 0.15 to 0.30. Two independent lines of work landing on the same variable is a strong sign the variable is real.

The most rigorous version of this comes from IBM Research. DivEye, published in TMLR 2026, formalizes rhythmic unpredictability using first- and second-order derivatives of per-token surprisal, which is a precise way of measuring burstiness. It outperforms zero-shot AI detectors by 33.2 percent and boosts existing detectors by 18.7 percent as an auxiliary signal. When a method that measures sentence rhythm improves detection by that margin, it confirms sentence-length variation is the strongest single distinguishing signal in the text, not a soft secondary cue.

Our own tests isolate the same variable from the other side. Holding post length, topic, and posting time constant while varying only sentence rhythm, the bursty variants averaged 40 to 60 percent higher impressions in the first 24 hours. Because every other input was pinned, sentence-length variation is the reach predictor, full stop. The detection research and the distribution data describe the same property from two angles: one says rhythm separates human from machine, the other says rhythm separates reach from stall.

If you take one number from this section, make it the threshold pair. Below roughly 0.30 you are in AI territory and at suppression risk. Above 0.55 you are reliably in the band that keeps distributing. The space between is where most unedited AI drafts land, and it is the space the rest of this guide is built to get you out of.

The 35% reach penalty for LinkedIn posts with flat sentence rhythm

Flat rhythm carries a measurable cost. In AuthoredUp's analysis, LinkedIn posts written above a 10th-grade reading level receive approximately 35 percent less reach. The algorithm favors short sentences, in the 10 to 15 word range, and simple vocabulary, and grade level predicts distribution outcomes directly, not just a readability rating in some tool. A high grade level usually travels with long, uniform sentences, which is the rhythm the quality model marks down.

The scale of the problem is now a population effect, not an edge case. A 2025 Originality.AI study found that over 50 percent of LinkedIn posts were likely AI-generated, and those posts showed measurably lower engagement rates than human posts. That is the suppression mechanism working across the whole feed through dwell-time scoring, with no explicit AI label attached to anything. More than half the feed is competing with the same flat-rhythm handicap, which is also why a post that fixes its rhythm stands out so quickly.

The clearest window into the mechanism comes from a place you would not expect. In Stanford research, AI detectors misclassified 61.3 percent of non-native English TOEFL essays as AI-generated. The essays were written by humans. They got flagged because constrained vocabulary and syntactically uniform sentences mimic AI output's low-burstiness signature. The detector was not reading for meaning or origin. It was reading for rhythm, and uniform human writing tripped the same wire that uniform machine writing does.

That false-positive result is the proof that the signal is structural. The system is not detecting AI in any semantic sense. It is detecting the rhythm of AI: uniform sentence lengths, predictable word choices, and low variance in syntactic complexity. A human who writes that way gets read the same as a model that writes that way. This is liberating once you internalize it, because it means the fix is mechanical and available to anyone willing to vary their sentences.

Pulling our own corpus against these thresholds, the cliff is consistent. Posts scoring below 0.30 burstiness exhibit the rhythm that both external AI detectors and LinkedIn's internal quality model treat as a suppression signal. The performance drop is not gradual across the range. It clusters at the 0.35 threshold, where posts reliably tip from continued distribution into the first-cohort stall. AuthoredUp's 35 percent reach penalty for high-complexity text and our burstiness cliff are two readouts of the same underlying flatness.

Why RLHF-aligned AI tools make the LinkedIn sentence length problem worse over time

The default outputs of ChatGPT, Claude, and Gemini are RLHF-aligned, and alignment makes this problem worse, not better. RLHF-aligned models paradoxically produce longer, more repetitive text that is easier for detectors to flag. The alignment process introduces stylistic regularities that flatten burstiness and strengthen the very signal LinkedIn's quality model scores. The more polished the default output looks, the more uniform its rhythm tends to be.

The reason is baked into how alignment works. RLHF optimizes for human preference ratings, and raters reward consistent tone, clear structure, and even pacing. Those preferences push the model toward a narrow stylistic range: readable, helpful, evenly weighted, and low in burstiness. The training that makes a model pleasant to read in a chat window is the same training that flattens its sentence rhythm into something a feed ranking model marks as low value.

We can see the drift in our own measurements. Off-the-shelf outputs from current RLHF-aligned models produce flatter burstiness curves than older GPT-3-era outputs did. The raw material got smoother and more uniform over successive model generations, which means the AI content suppression problem is getting harder over time for anyone relying on default outputs without post-processing. Teams that copied a 2022 workflow into a 2026 model and wondered why reach dropped are running into exactly this.

The pipeline compensates with a post-processing step that injects sentence-length contrast patterns, targeting Flesch-Kincaid Grade Level 6 to 8 while hitting a burstiness target above 0.50 at the same time. This addresses the rhythmic signal the quality model scores, rather than swapping out a few obvious AI clichés and hoping the rest passes. Cleaning up surface word choice does nothing for burstiness if every sentence is still the same length.

There is a strategic point underneath the tactics. Because perplexity and burstiness are both becoming less reliable in isolation as detection technology advances, generic advice to write more conversationally is wearing thin. Tuning burstiness as a specific, measurable pre-publish target is more durable, because it is tied to a number you can compute and gate on rather than a vibe you hope you captured. A target survives model upgrades. A vibe does not.

LinkedIn AI content sentence length benchmarks: Flesch-Kincaid and burstiness targets

The readability benchmarks for LinkedIn content are concrete, so use the numbers. Aim for a Flesch Reading Ease score of 60 or higher and a Flesch-Kincaid Grade Level of 8 or below. These thresholds tie directly to both user engagement and search-driven visibility, which is why they are worth treating as gates rather than nice-to-haves. They are also easy to check, and most teams never check them.

Top creators sit even simpler than that. AuthoredUp's analysis of 20 LinkedIn influencers found the optimal Automated Readability Index range is 1 to 5, which reflects the plain, conversational writing the best-performing accounts use consistently. That range looks shockingly low to anyone trained to write formally. It is a deliberate choice, and it is the floor those creators write to on purpose, post after post.

SocialNexis targets Flesch-Kincaid Grade Level 6 to 8 combined with a burstiness score above 0.50, and the important detail is that these are independent gates. A post can hit Grade Level 7 while scoring 0.20 burstiness, if every sentence happens to be the same length and uses simple words. Simple vocabulary does not guarantee varied rhythm. Both gates have to pass before anything publishes, because each one catches a failure the other misses.

The reach math reinforces the targets from the other direction. Posts above a 10th-grade reading level receive approximately 35 percent less reach in AuthoredUp's data, while posts targeting a 4th-grade reading level outperform high-complexity text when topic and audience are held constant. Lower grade level is not dumbing the content down. It is removing the friction that keeps a reader from getting to your actual point, which is where the dwell time lives.

Both metrics are computable before you hit post, with free tools. Hemingway Editor reports grade level in real time as you write. Burstiness takes a simple calculation: the standard deviation of per-sentence word counts divided by the mean. The whole pre-publish check runs in under two minutes once it is a habit. That two minutes is the practical line between posts that distribute and posts that stall, and almost nobody spends it, which is part of why the fix works so well.

How to tune sentence burstiness in an AI-generated LinkedIn post before publishing

Start with the first three sentences, because they do more work than the rest combined. Our data shows the opening lines carry disproportionate weight for both the see more click signal and the algorithm's initial quality classification. The pattern that works is mechanical: a short sentence under 8 words, then a longer context sentence of 18 to 25 words, then another short sentence. That sequence builds the burstiness signature inside the above-the-fold preview, which primes both the reader and the ranking model before the post is even expanded.

Then compute the score for the whole draft. List the word count of each sentence, calculate the mean, calculate the standard deviation, then divide the standard deviation by the mean. Target above 0.50. If the result lands below 0.40, the post needs at least 3 to 4 sentence splits or merges before it is ready. Do not guess at this. The number tells you exactly how much work the draft still needs, and an unedited model draft will almost always land below the line.

The rewrite is structural, not semantic, and that is what makes it fast. Break a 20-word sentence into a 6-word sentence and a 16-word sentence. Merge two adjacent 8-word sentences into one 18-word sentence. The argument stays identical. The word count barely moves. Only the rhythm changes. You are not rethinking the post, you are re-cutting it, and you can do a full pass in a couple of minutes once the splits become reflex.

Pair the rhythm change with visual breaks. Single-line paragraphs increase reading time by up to 20 percent, so put each short sentence on its own line with a paragraph return. The white space and the short sentence reinforce each other, and together they pull total dwell time toward the 30 to 45 second range that signals continued distribution. Rhythm on the page and rhythm in the layout are the same lever pulled twice.

The target for a publish-ready post is specific enough to gate on: burstiness above 0.50, Flesch-Kincaid Grade Level between 6 and 8, and average sentence length between 10 and 15 words. This is the same gate SocialNexis runs automatically, where any draft scoring below 0.40 burstiness gets rewritten with shorter sentences inserted between longer ones before publishing. Whether a pipeline does it or you do it by hand, the rule holds. Posts targeting a 4th-grade reading level consistently outperform high-complexity text when the other variables are held constant, and varied rhythm is how you get there without flattening the substance.

Frequently asked questions

Does sentence length variation (burstiness) affect LinkedIn reach more than total post length?

Yes. SocialNexis A/B tests holding post length, topic, and posting time constant show bursty variants averaging 40-60% more impressions in the first 24 hours. Total post length affects reach within a character-count range (the 2,001-2,500 character band shows peak engagement at 2.67% in AuthoredUp's data), but within any length bracket, burstiness is the variable that determines whether the post reaches its ceiling or stalls.

What is the optimal burstiness score for a LinkedIn post to avoid AI-content suppression?

SocialNexis targets a burstiness score above 0.50 for publish-ready posts and treats 0.40 as the minimum acceptable threshold: posts below that are automatically rewritten before publishing. Posts below 0.35 stall at the first distribution cohort in SocialNexis's data. For reference, human writing typically scores 0.60-1.00+. AI-generated text from RLHF-aligned models typically scores 0.15-0.30.

How does LinkedIn's algorithm detect uniformly structured AI-generated posts without an explicit AI label?

LinkedIn's 360Brew model (150 billion parameters, deployed 2025) scores probabilistic quality signals, including lexical diversity and sentence-pattern uniformity. It does not issue a hard AI label. Posts exhibiting statistically uniform sentence rhythm receive a depressed quality score that reduces first-pass distribution regardless of topic. The suppression is invisible and self-reinforcing: lower first-pass reach means fewer early engagement signals, which prevents re-amplification.

What Flesch-Kincaid grade level should a LinkedIn post target to maximize organic reach?

Target Flesch-Kincaid Grade Level 8 or below and Flesch Reading Ease 60 or higher. Posts above a 10th-grade reading level receive approximately 35% less reach, while posts targeting a 4th-grade reading level outperform high-complexity text. SocialNexis targets FK Grade Level 6-8 as a pre-publish gate, paired with a burstiness score above 0.50, because a post can meet the grade-level target while still failing on sentence rhythm.

Why do AI-generated posts get lower reach on LinkedIn even when they are not flagged as AI?

Because LinkedIn's quality model scores the rhythm of AI content, not its topic or authorship. Uniform sentence lengths produce low dwell time: readers scan and scroll past, generating a negative ranking signal. A 2025 Originality.AI study found over 50% of LinkedIn posts were likely AI-generated, with those posts showing measurably lower engagement rates, confirming that algorithmic suppression operates through dwell-time scoring across the full LinkedIn feed, not through explicit labeling.

What is the difference between post length and sentence-length variation as LinkedIn reach predictors?

Post length sets the opportunity ceiling. The 2,001-2,500 character range shows the highest engagement rate (2.67%) in AuthoredUp's dataset of 372,126 posts, so length determines the distribution band you are competing in. Sentence-length variation (burstiness) determines whether you reach that ceiling. Two posts identical in length, topic, and posting time can diverge 40-60% in impressions based on burstiness score alone.

How does sentence rhythm (alternating short and long sentences) affect LinkedIn dwell time scoring?

LinkedIn's feed ranking measures dwell time as a direct positive signal: time spent viewing a post increases its distribution score. Short sentences create visual white space and reading pace that pulls the reader forward. Long sentences deliver substantive content. The alternation prevents eye fatigue and increases the probability a reader reaches the end, pushing total dwell time toward the 30-45 second range that triggers continued distribution.

Can tuning the burstiness of an AI-generated LinkedIn post prevent algorithmic suppression?

Yes, based on SocialNexis controlled observations. Holding post length, topic, and posting time constant, bursty variants of AI-generated posts averaged 40-60% higher impressions in the first 24 hours. The rewrite is structural: split long sentences, merge short ones, target a standard deviation above 0.50. The argument and word count stay nearly identical. Burstiness tuning addresses the rhythmic signal LinkedIn's quality model scores, not surface-level word choices.

What writing patterns does LinkedIn's 360Brew model score as low-quality AI content?

Based on published model descriptions and observed suppression patterns: statistically uniform sentence lengths, low lexical diversity, and predictable word-level choices that produce low perplexity scores. These patterns correlate with low dwell time. The model scores these signals probabilistically. Posts exhibiting multiple low-quality patterns in combination receive the steepest distribution penalties, even when no explicit AI label is applied.

How do Flesch Reading Ease scores and sentence burstiness interact to predict LinkedIn post reach?

They are related but independent gates. Flesch Reading Ease scores grade level and word complexity. Burstiness scores sentence-length variance. A post can score Flesch Reading Ease 65 (acceptable) while scoring 0.20 burstiness (suppression risk) if every sentence is similar in length but uses simple vocabulary. SocialNexis runs both checks before publishing: Flesch-Kincaid Grade Level 6-8 and burstiness above 0.50. Both must pass independently.

Sources and further reading

• LinkedIn Engineering: how dwell time signals feed ranking
• GPTZero's explainer on perplexity and burstiness in AI text detection
• Basani and Chen (IBM Research): rhythmic unpredictability as an AI detection signal