Are you able to run a social media account solely utilizing AI?
I not too long ago took a while off work and thought I’d use the chance to reply this query.
See, it’s laborious developing with new content material persistently, however I’m a little bit of a purist relating to what I share on social media and the way I share it. We’ve even lined this earlier than whereas addressing how the Buffer Content material workforce makes use of AI in our artistic course of.
However I assumed, for science, I’d do that experiment and reply one of many AI questions that creators may be deeply interested in: how does AI content material on LinkedIn carry out?
What I did and the experiment parameters
I centered on a weeklong interval, scheduling all content material by Buffer. This method allowed me to investigate the efficiency of AI-generated content material, which I did utilizing the stats LinkedIn makes obtainable.
For this experiment, I centered my efforts solely on LinkedIn, the place I’ve been constructing my private model, over a one-week interval from November sixth to November twelfth, 2023, and I used LinkedIn’s platform analytics to trace and compile the efficiency information of the AI-written content material.
To generate the posts, I used three main AI instruments to not appear biased – Buffer’s AI Assistant, Claude AI, and ChatGPT. Additionally, these instruments particularly gave me the most effective likelihood of making content material to a sure stage of high quality I hoped for.
To maintain the experiment parameters strict, 100% of the content material was AI-generated based mostly on varied prompts. The one exception was some gentle proofreading earlier than publishing.
I assigned each the duty of drafting content material for a set variety of days:
- Buffer’s AI Assistant – three days of LinkedIn posts
- Claude AI – two days of LinkedIn posts
- ChatGPT – two days of LinkedIn posts
Whereas I attempted to verify every instrument created a set quantity of content material for the times I used to be assigned, I used a mixture of all of the instruments to refine the content material. So, there’s nobody instrument I’d say is healthier, however one widespread thread throughout all of the instruments was that the extra context I provided, the higher the content material I received.
By way of prompts, I equipped the AI instruments with:
- Authentic prompts I crafted particularly for my audience and messaging targets.
- An current high-performing content material immediate shared by Mike Cardona’s 90-Day Content material Library immediate.
- Prompts I had beforehand created for Buffer’s Assistant (these are often featured in our Social Media publication).
I made positive all of the prompts mirrored my content material pillars of non-public model constructing, profession development and AI, so I may keep on model.
Whereas there was no true scientific or data-driven foundation for this experiment (I really simply wished to see if my viewers would discover), listed below are among the boundaries I set for myself:
- All content material must be completely AI-generated – solely proofreading could be performed manually
- I can tweak prompts infinitely to get higher outcomes, however nothing extra
- I can hop in to have interaction with feedback
- All content material is scheduled in Buffer (*wink wink*) and sorted with our new Tags function.
These boundaries helped me pressure myself away from perfectionism, which allowed me to save time and work shortly. Nevertheless it additionally restricted the creativity and private perspective I may put into the content material, a significant limitation of AI-only content material.
The objective with these managed parameters was to check how an viewers would reply in the event that they acquired every week of content material from my account written totally by AI, with minimal human oversight. Listed here are the outcomes.
The outcomes
Now, I must share one factor about me: my information evaluation expertise aren’t the strongest. So I’ve needed to flip to AI at this stage of the experiment as effectively.
Upfront evaluation of content material efficiency
Listed here are the outcomes of the AI-generated content material from November sixth to twelfth, 2023.
- Complete impressions for that week: 9,624
- Common day by day impressions: 1,375
- Complete engagements for that week: 151
- Common day by day engagements: 22
Total, whereas the engagement charges may doubtlessly be larger, the AI-written posts fared effectively, and all my objections are extra concerning the high quality of the content material.
The impressions and complete engagement numbers point out an engaged viewers for content material written mechanically with minimal oversight. Monitoring this over an extended interval may present perception into actual efficiency traits. However for the week I centered on, the posts achieved strong metrics.
Now, let’s dive deeper into the information.
Impressions and engagements
Over the 7-day experiment, the AI-generated content material garnered vital visibility totaling 9,624 impressions, producing 151 person interactions by likes, feedback, and shares registered as engagements.
On a mean day by day foundation, this broke all the way down to:
- 1,375 impressions
- 22 engagements
In comparison with my general LinkedIn averages, this week massively over-indexed for visibility and response:
- Common impressions for Nov 6-12 are about 11% larger than a typical week over the earlier 3-month interval.
- Common engagements for Nov 6-12 are over 75% larger than a typical week.
Based mostly on this, we will assume that the AI-generated content material resonated by way of sheer attain and uncooked interactions generated based mostly on elevated volumes from historic baselines.
At an mixture weekly stage, reaching practically 10,000 impressions demonstrates a significant scale of discovery. And whereas I want engagement was larger (don’t all of us), crossing 150 actions or practically 25 per day is a powerful baseline response indicating the AI-produced posts intrigued my viewers.
Engagement charges
We will additionally study person conduct by the engagement charges, also called the ratio of interactions to impressions.
Over the seven-day stretch, the posts achieved a mean 1.57 % engagement price, which is taken from the 151 complete engagements generated divided by the 9,624 mixture impressions.
Breaking down day by day engagement price supplies further context:
- Finest Performing Day: November sixth at 3.5 % price
- Worst Performing Day: November twelfth at 1 % price
- Remaining days ranged between one to 3 %
The most effective performing day was a Monday, and the worst was a Sunday, so the downward pattern isn’t worrying and matches with expectations of LinkedIn content material efficiency.
From this evaluation, I can inform that scheduling posts earlier within the week could possibly be higher for engagement.
Precise content material efficiency
Now, transferring on from the averages and aggregates of the entire week, one main notice from the efficiency of content material throughout the week is that actionable recommendation immediately serving to readers succeed at one thing carried out dramatically higher.
Digging deeper, the very best traction submit from November sixth overlaying actionable on-line writing ideas noticed 60 person interactions measured in opposition to 1,699 impressions for a 3.5 % engagement price.
Comparatively, November twelfth’s lower-performing submit was extra conceptual/philosophical as an summary of AI branding fundamentals and noticed solely 10 engagements from 967 views – a one % price.
Analyzing probably the most and least participating items by matter reveals that my viewers seems to strongly desire instantly relevant “how-to” enhancements. Regardless of its informational worth, forward-looking thought management usually overwhelms or loses parts of audiences.
This pattern recurred all through the week, with sensible skill-building content material considerably outperforming refined however extra passive consumption items.
The clear takeaway facilities on bite-sized, tactical content material higher commanding viewers funding – aligning rationally with their rapid development wants.
Time Collection Evaluation
I received ChatGPT to make a chart plotting day by day impressions all through the week.
Some key observations and takeaways from this evaluation:
- Peak days: There was a major uptick within the engagement at first of the week, with Mondays and Tuesdays exhibiting the very best ranges of interplay.
- Mid-week traits: A noticeable dip occurred mid-week, significantly on Wednesday and Thursday, indicating much less viewers exercise throughout today.
- Weekend insights: Regardless of a common notion of weekends being much less favorable for engagement, our Saturday posts carried out comparatively effectively, though a drop was noticed on Sunday.
What went effectively
So, let’s speak about the good things from this AI content material experiment. After I dove into the numbers and the ups and downs of the week, a number of cool issues actually stood out.
First off, the AI was operating the present right here with only a trace of a human contact by prompts and context sharing. This gave me a contemporary have a look at how content material lands with out the added private perspective. And shock, shock, it seems that AI can churn out stuff that not solely grabs consideration but additionally will get individuals speaking and interesting. Fairly cool.
Once we stack this week’s numbers in opposition to earlier ones, it is clear that AI is not only a one-hit surprise as a artistic assistant. We’re speaking constant affect, pulling in views and interactions past what we often see. Nonetheless, this is not simply random luck however a mix of some issues:
- The belief I’ve constructed with my unique content material performed an enormous position within the efficiency of the AI content material. I sometimes don’t publish on daily basis of the week, however once I do, I get engagement. That’s a results of belief constructed over time with my viewers. My recommendation: give attention to constructing that belief.
- A deep understanding of what’s prone to resonate with my viewers by content material pillars. I didn’t simply choose the random concepts I received from the AI instruments, I made positive to refine the content material until it matched what I knew individuals would count on from me.
Now, let’s speak matters. The most well-liked submit provided sensible recommendation centered immediately on the reader – how one can enhance their writing expertise as a creator. The least widespread took a unique, broader angle discussing AI purposes for private branding, ending up extra conceptual and summary for the typical reader. Some key takeaways:
- Posts offering tangible ideas, methods or recommendation for readers scored a lot larger engagement than big-picture assume items
- Actionable content material serving to customers make progress resonated greater than thought leadership-style concepts
- Practicality over philosophy when aiming to drive interactions
This means focusing content material on bite-sized, sensible takeaways readers can instantly apply will reliably yield larger engagement. Whereas extra conceptual or forward-looking themes could lose or overwhelm some customers regardless of being intellectually fascinating.
What didn’t go effectively
Prompting AI instruments is extra an artwork than a science, which implies there’s no exact solution to get it to really “sound human” except you intrude and edit the content material it generates.
For instance, once I would share a immediate, the primary reply would virtually all the time be extraordinarily flawed. Some widespread errors had been repetition and pointless lists. AI instruments even have a bizarre behavior of capitalizing in bizarre locations – and I don’t write like that. I may all the time share further prompts to get the outcomes nearer to sounding like me, nevertheless it wasn’t excellent.
Conclusion
So, sure, I revealed AI-generated content material for every week straight, and nobody seen. The truth is, my engagement stayed the identical and was even higher in some instances.
My subsequent transfer is all about fine-tuning. Listed here are some subsequent steps I’d take away from this experiment:
- My content material pillars work greatest once they comply with the actionable recommendation route, so I’ll prioritize that content material on LinkedIn to any extent further.
- Lengthy-form content material is a winner – all of the posts had been over 350 phrases, and the efficiency wasn’t harm by size.
- That is extra of a private factor, however I’ll all the time tweak the AI voice and elegance to match mine. It was uncomfortable to note issues I’d have eliminated/modified if I hadn’t set such construction parameters.
For those who’re like me and have constructed up belief together with your viewers, battle with consistency, or simply need extra methods to border your concepts, letting AI take a swing at increasing your attain looks as if a no brainer.
