Performing for the Algorithm
How opaque systems change our behavior without ever explaining what they want
A few weeks ago, I wrote a LinkedIn post about receiving a rejection letter. I’m 69 years old, I’ve self-published several children’s books, and I’d recently started querying literary agents for the first time. One of them sent me a very nice rejection. She’d enjoyed The Stinky Princess, but her client list was full.
I wrote about why this made me unexpectedly happy. At 69, I’d found another way to be a beginner. Getting rejected meant I was actually trying something I hadn’t done before.
The post took off, at least by my standards. It received 8,401 impressions, 101 engagements, 29 comments and a repost. More interestingly, only 24 percent of those impressions came from within my LinkedIn network. LinkedIn had decided to show the post to thousands of people who didn’t know me.
I was delighted, and then I made a mistake: I tried to figure out why.
Reverse Engineering the Oracle
LinkedIn gives you information about what happened to a post. It tells you how many impressions it received, how many people engaged with it and something about who those people were. So, I started developing theories about why this post had done so well. Maybe the personal story had helped? Maybe Thursday was a particularly good day for me to post? Maybe the image was the key? Maybe keeping the story native to LinkedIn rather than sending people elsewhere made a difference? Maybe the publishing hashtags helped reach an audience I hadn’t tapped before?
All perfectly plausible, and, as it turned out, not terribly useful.
I had another story involving one of my children’s books where a marketer contacted me to say how much she’d enjoyed the book, only to eventually admit she hadn’t read it.
I thought this was fascinating, particularly as an example of how AI can make counterfeit attention extraordinarily cheap, so I wrote a Medium article about it. Then I created a separate LinkedIn post, using what I thought I’d learned from the previous one to guide me.
My successful post had involved children’s publishing, and so did this one. It had told a personal story, and so did this one. I posted on Thursday again, used an illustration from the book, kept the story on LinkedIn rather than merely posting a link, and used relevant publishing hashtags.
The result was 209 impressions, 88 percent of them from within my network.
Hmm.
Clearly, More Research Was Required
Then came Goodreads.
The marketer had mentioned my “other seven books.” This puzzled me because I had written only four, so I went looking for the source and discovered that Goodreads thought I had written eight. It had merged some of my books with books by another Phil Gold, a distinguished Canadian physician and cancer researcher who has been complicating my search results for years.
I corrected the Goodreads record and wrote another article about what happens when incorrect information enters apparently authoritative databases. Then I turned that into another LinkedIn story, beginning with what I thought was a pretty good opening:
“I recently discovered that I had written eight books. This came as something of a surprise, since I thought I had only written four.”
It was another publishing story, another personal story and another slightly ridiculous situation. This time I changed the day and posted Wednesday afternoon, because maybe Thursday wasn’t the secret after all.
The result was 148 impressions, 87 percent from within my network.
At this point, I developed another theory: maybe the post was too long.
Notice what was happening. LinkedIn hadn’t told me my post was too long. It hadn’t told me that Thursday mattered or didn’t matter, that publishing stories worked, that personal stories worked, that images mattered or that hashtags helped. In fact, it hadn’t actually told me anything.
I was changing my behavior anyway.
Then I Stopped Experimenting
A few days later, I received a wonderful review of The Duchess Would Not Approve! from Midwest Book Review. I posted about it on LinkedIn, not because I was conducting another experiment but because I had good news and wanted to share it. I included an illustration from the book.
Then I noticed something I hadn’t seen before. In the corner of the image was a tiny “Cr” icon. Clicking it revealed Content Credentials showing that the image had been created using OpenAI’s image-generation technology.
Interesting. Could LinkedIn be treating AI-generated images differently?
I have no objection to people knowing that I use AI in my illustration process. I’ve written publicly about it. But if an AI provenance marker were being used as an automatic reason to restrict distribution, I’d have considerably stronger feelings about that.
For a few minutes, I thought I might have found something. Then I checked the previous posts.
The counterfeit-attention post had carried the Content Credentials marker and had performed poorly. The Goodreads post hadn’t carried one and had performed even worse. My wildly successful Stinky Princess post hadn’t carried one either, although the cover included AI-generated artwork that I’d subsequently assembled into a composite image in Adobe Illustrator.
Get Phil Gold’s stories in your inbox
Join Medium for free to get updates from this writer.
Perhaps Content Credentials mattered. Perhaps they didn’t. I added another theory to the pile.
The book-review post eventually received 78 impressions, 97 percent from within my network. For those keeping score, my attempts to understand LinkedIn had now produced 8,401 impressions, followed by 209, then 148, then 78.
I was getting remarkably bad at this.
The Algorithm Doesn’t Have to Explain Itself
Somewhere in the middle of all this, I realized that the most interesting part of my experiment wasn’t whether I could figure out LinkedIn’s algorithm. It was what the algorithm was doing to me.
The first-order effect of an algorithmic feed is straightforward. There’s far more content available than any of us could possibly consume, so a system decides what to show us. That’s useful, and probably unavoidable.
But there’s another group affected by that decision: the people who want to be shown.
The second-order effect is that once creators, businesses, musicians, authors or job seekers realize that an algorithm stands between what they produce and the people they hope to reach, they begin adapting to it. The strange part is that we don’t even need to know how the algorithm works. We just need theories about how it works.
A post performs well, so we examine it for clues. Was it the subject, the length, the time of day, the image or the hashtags? Did people spend longer reading it? Did early comments trigger wider distribution? Did the system classify it differently? Was Mercury in retrograde? (I haven’t tested that last one yet.)
Then we change something and try again.
The system has changed my behavior without ever explaining its own behavior to me. The algorithm never issued an instruction; it never said, “Phil, we’d like you to write shorter posts.” But after watching a long post that I thought was genuinely funny receive 93 impressions, guess what I’m planning to try next?
A shorter post.
We’re All Doing This
LinkedIn is only one example. Musicians think about what Spotify’s recommendation systems might favor. Writers think about what Amazon might surface. Businesses structure websites around what they believe Google wants. Job applicants add keywords because they know software may evaluate their résumés before a person does.
Entire professions have grown around trying to understand these systems. Some of the resulting advice is based on good evidence, some on experience, and some is folklore repeated so many times that it begins to sound like established fact.
The situation becomes even more interesting because the systems themselves keep changing. Suppose enough people become convinced that an algorithm rewards a particular behavior and start behaving that way. Suddenly the platform fills with content optimized around that belief, which changes the environment in which the algorithm is operating.
The people designing the algorithm respond. Users observe the new results, develop new theories and adapt again.
Technology changes behavior. Human behavior changes the environment. Then technology adapts to the changed environment, and around we go.
The Frustrating Part
I understand why LinkedIn doesn’t publish a precise recipe for getting broad distribution. If everyone knew exactly what the system rewarded, we’d optimize for the recipe rather than for the people reading our posts. In fact, we’re already doing some of that without knowing the recipe.
But there’s a considerable distance between publishing the recipe and providing almost no meaningful explanation at all, and that’s the part I find frustrating.
My rejection post received 8,401 impressions, 76 percent from people outside my network. A funny story about discovering that I’d apparently written four books I knew nothing about received 148 impressions, 87 percent from people already connected to me.
I can see what happened, but I have almost no idea why LinkedIn had decided to show the first post to thousands of people who didn’t know me.
And because distribution matters to me, I adapt. That’s what people do. We’re remarkably good at spotting patterns, including patterns that may not actually be there, particularly when the outcome matters to us.
The algorithm is trying to learn what people will do. Meanwhile, people are trying to learn what the algorithm wants them to do. Neither side has complete information about the other.
What Have I Learned?
Not how to succeed on LinkedIn, obviously. If anything, my experiments have become progressively more impressive demonstrations of how not to reproduce an 8,401-impression post.
But I have learned something more interesting. We tend to think of algorithms as systems that observe, classify and predict human behavior. That’s only half the relationship.
Once we know the algorithms are watching us, we start watching them back. We study their decisions, develop theories and change what we do. Then they observe the changed behavior and adjust again.
Technology starts the change. People decide what happens next. And sometimes the technology gets another turn after that.
This week, I’m going to try a shorter LinkedIn post. I have absolutely no evidence that this will work, but apparently LinkedIn has trained me to try it.
I’ll let you know what happens.









