What are 'AI writing tics' and how do they affect content engagement?

AI writing tics are patterns in content, often generated by AI, that readers perceive as canned or performative. A study found these tics, especially conclusion signaling, correlate with lower engagement, leading to reduced time on page and higher bounce rates.

What did the study by Adam Gnuse involve?

SEO Content Manager Adam Gnuse analyzed over 1,000 content marketing URLs across 10 domains, normalizing the appearance of various writing tics per 1,000 words and comparing them to GA4 engagement metrics like engaged sessions.

Study Reveals AI Writing Tics That Reduce Content Engagement

Q: Which specific AI writing tic was identified as the biggest engagement killer?

The study found that 'conclusion signaling,' particularly repeated 'Conclusion' starters, had the strongest negative correlation with engagement rate. These sections often feel generic and announce the end of helpful content, causing readers to leave.

Q: How can content creators improve their conclusions for better engagement?

Instead of generic 'In conclusion' sections, creators should deliver what readers want: next steps, trade-offs, decision criteria, or a summary that feels earned and provides real value, rather than just repeating earlier points.

By Charles Ryder

I’ve watched the AI-content debate bounce between two useless extremes: everything written with AI is garbage versus AI can do the whole job. A new study highlighted by Search Engine Land lands in a more practical place. It suggests certain “AI writing tics” correlate with lower engagement—just not in the simplistic, TikTok-detective way people like to frame it.

The lesson isn’t “hunt for em dashes and delete them.” Readers respond to patterns that feel canned, prematurely tied up, or weirdly performative. When they check out, the impact is obvious: time on page drops, bounce rate rises, and organic performance can slide—especially after years of Google pushing “helpful” content and cracking down on scaled, low-value pages.

What the study actually did (and why it matters)

SEO Content Manager and analyst Adam Gnuse analyzed 1,000+ content marketing URLs across 10 domains, spanning industries like tech, ecommerce, healthcare, education, and analytics. He normalized the appearance of various “tics” per 1,000 words and compared them to GA4 engagement metrics—specifically engaged sessions (for example: 10+ seconds, a key event, or 2+ pageviews).

One thing I like about the approach: it wasn’t built as a gotcha. The dataset likely mixes human writing, AI writing, and hybrid drafts, because that’s what publishing looks like now. Gnuse also added literary controls like a 2021 novel and Shakespeare’s Hamlet to show that some “tics” aren’t inherently AI. They’re just… writing.

So no, this wasn’t another “AI detector” dressed up as research. It was an attempt to connect specific patterns on the page to actual reader behavior.

The biggest engagement killer wasn’t what most people expected

If you spend time around content folks, you’ve heard the usual villains: em dashes, “delve,” “intricate,” overly symmetrical phrasing, and stiff transitions.

But the strongest negative correlation with engagement rate in this study wasn’t punctuation. It was conclusion signaling—especially repeated “Conclusion” starters, which showed the most negative relationship (reported around -0.118 correlation in the write-up).

That number sounds modest until you remember what real marketing data looks like: noisy, inconsistent, and full of competing variables. A single stylistic habit showing up across industries and still emerging as the strongest negative signal is worth taking seriously.

Why “Conclusion” can backfire

Most “Conclusion” sections read like content clearing its throat. They announce you’re about to stop being helpful.

When the wrap-up is generic—“In conclusion, X is important, consider Y”—readers spot the template. They scroll, see the finish line, and leave. And a lot of those endings just repeat earlier paragraphs instead of delivering what readers actually want at the end: next steps, trade-offs, decision criteria, or a summary that feels earned.

The sneaky repeat phrase that drags performance

Another pattern with a larger-than-average negative effect was “not only…but also”—sometimes appearing up to 12 times in a single post.

The problem isn’t the phrase itself. It’s the repetition. When a page reads like it was assembled from a template, trust drops—even if the facts are solid. The language starts to feel like padding, and padding kills engagement.

This is where a lot of AI-assisted drafts stumble: they don’t have a “taste filter.” They’ll reuse the same rhetorical structure because it’s statistically safe, not because it’s the best way to say something.

The twist: em dashes aren’t the villain

Here’s the part that will irritate the internet: em dashes had a slight positive correlation with engagement.

That tracks with how people actually read. Used well, an em dash signals voice, rhythm, and comfort with nuance. Sure, models can overuse them. Humans can too. But readers don’t punish punctuation. They punish predictability and low-value structure.

Gnuse even joked that, as someone who genuinely likes em dashes, the data was “deeply validating.” I get it.

What’s neutral (and what not to obsess over)

A few tics that people love to tag as “AI” didn’t show meaningful impact here, including:

Sentence starters like “then,” “this,” or “that”
Common framing lines like “in this article”

That doesn’t make them great. It just means they’re rarely the lever that moves engagement. If your content is underperforming, deleting every “this” won’t rescue it.

The bigger point: engagement is the real battlefield

The most useful implication is also the simplest: stop writing to avoid detection, and start writing to earn attention.

Google’s recent wave of updates—Helpful Content, more emphasis on E-E-A-T, warnings about scaled content abuse—point in the same direction. If a page exists mainly to occupy a SERP slot, it eventually loses. Engagement metrics aren’t the only ranking factors, but they’re tightly linked to whether a page is satisfying intent, especially in competitive spaces.

Secondary commentary around this research points the same way: human-crafted pages often show better on-site behavior, with claims like lower bounce and longer session duration compared to AI-heavy content. Even if every number doesn’t generalize, the direction is consistent: formula writing makes readers leave.

What I’m changing in my own AI-assisted workflow

I’m not anti-AI. I’m anti-autopilot. Here’s how I’m tightening my process after reading this:

I stop announcing the structure and start delivering it.
Instead of “In conclusion” or a labeled “Conclusion” that restates the obvious, I end with a practical wrap: what to do next, what to ignore, and what a “good” decision looks like.
I actively hunt repeated rhetorical scaffolding.
Phrases like “not only…but also,” “it’s important to note,” and “in today’s fast-paced world” aren’t fatal—unless they show up like a drumbeat. Variety reads as human because it reflects real thought.
I keep the em dash if it’s doing a job.
Rhythm, contrast, a quick aside: great. Randomly sprinkling them in because the model likes them: gone.
I rewrite endings first.
A lot of AI drafts collapse at the finish. If I fix the ending early, the piece becomes more purposeful, and I’m less tempted to pad the middle.

Conclusion: use the study as a compass, not a blacklist

The value of this research isn’t a list of banned phrases. It’s the reminder that engagement drops when readers can see the template underneath the prose. If you use AI, the goal shouldn’t be to “sound less like AI.” It should be to sound like someone who knows what matters, has made real decisions, and is willing to be specific.

If you want a tighter way to spot and clean up engagement-draining tics—without sanding off your voice—build a consistent quality pass into your workflow with a tool like AIuthority. It helps turn AI-assisted drafts into reader-first publishing.