One of our clients stopped using AI written social posts three months ago. Engagement went from about one like per post to fifty five. I went looking for research that would explain it, and the research says the effect should have been tiny.
That gap is the whole story, and it points somewhere more useful than "AI bad".
What actually changed
The practice is an emergency animal hospital, open overnight, every weekend and every holiday. Independent, and in the same city since the late nineties.
Their social was running on generated content. Competent, on topic, seasonal. It was also getting close to nothing. Zero or one like, no comments, no shares, week after week.
We changed two things at once, which matters later.
The first was the pictures. Out went stock and generated imagery, in came photographs of their own people. An actual doctor sitting on the floor of a kennel with a litter of puppies. A technician holding somebody's cat. Their own rooms, their own cabinets, their own scrubs.
The second was the subject. Instead of general advice, the posts started saying the one thing about this practice that no competitor in their market can say: locally owned, independently operated, and not part of a corporate chain for nearly thirty years.
The numbers
Read the third column as shares, which is the one that costs a reader something.
- After hours care, with the doctor
- 55 likes, 7 comments
- 11
- Why locally owned matters
- 43 likes, 4 comments
- 1
- Locally owned, nearly thirty years
- 15 likes
- 0
- Any given post before the change
- 0 to 1 likes
- 0
Those are counts off the posts themselves, on our own client's account, reported by the team that runs it. First party numbers, not a study.
What the research says the penalty should be
Here is where it gets interesting, because somebody has measured this properly at enormous scale.
Carney, Riveros and Tully published "Made With AI" in the Journal of Consumer Research in 2026. Their first study looks at real platform data, not a survey.
1,135,817posts by 8,650 creators, with roughly 1.9 million comments analysed separately.
Holding views constant, what did an AI disclosure cost?
7%fewer likes, and about 7 percent less combined engagement, compared with posts carrying no AI disclosure.
Seven percent. Not ninety. The mechanism they identify is a drop in parasocial connection, which is the ordinary one sided bond a reader forms with whoever seems to be talking to them.
Separate work points the same way. Bruns and Meissner, in the Journal of Retailing and Consumer Services in 2024, ran three experiments and found that readers cannot reliably tell human and generated content apart by looking, yet still rate a brand that automates its content as less authentic.
Their most useful finding is the one nobody quotes. The negative reaction is weakened when the tool is used to assist a person rather than to replace them. Automating the whole job is what costs you. Using it to draft faster, with a human supplying the facts and the pictures, is not the same act and does not carry the same penalty.
So the AI was not the problem
Our client did not go up seven percent. They went up something closer to fiftyfold.
The published penalty for AI content cannot explain a change that size. Which means the generated content was not what was holding them back. Something else was, and the AI was just the most visible thing in the frame.
My read, and I want to be clear this is a reading rather than a measurement: the old posts were about a category. The new posts are about a specific practice, containing specific people, making a specific claim that is true of them and false of their competitors.
Generic content does not underperform because a machine wrote it. It underperforms because it would be equally true if you swapped the logo. Nobody shares a post that could have come from anywhere.
Why the locally owned post worked
The best performing posts here were not the clever ones. They were the ones stating a fact the practice can defend.
Independent for nearly thirty years is not a slogan, because in this category independence is genuinely rare. A 2025 SWOT analysis in Frontiers in Veterinary Science, drawing on Brakke Consulting's 2022 industry overview, puts corporate ownership at:
- 75%Specialty and emergency practices
- 50%Share of all national veterinary revenue
- 25%Primary care practices
- Specialty and emergency practices
- 75%
- Share of all national veterinary revenue
- 50%
- Primary care practices
- 25%
Three quarters of the category this practice competes in is corporate owned. So "locally owned" is not a warm sentiment. It is a statement about the other three options on the map at two in the morning, and it is checkable. No competitor can copy it by rewriting an About page.
Every category has one of these, and it is almost never the thing the website leads with.
- The dentist who still does their own lab work
- The optometrist whose entire team has been there nine years
- The chiropractor who is the only one in town treating a specific condition
- The practice whose owner still takes call personally
Notice none of those need a marketing department. They need somebody to notice they are true and then say them out loud with a photograph attached.
What this does not prove
I would be doing exactly what this blog exists to avoid if I let that sit without qualification.
This is one account. Three months. No control, and no counterfactual for what those same weeks would have looked like unchanged.
We changed the imagery and the subject matter at the same time, so nothing here separates the effect of real photographs from the effect of saying something specific. My guess is the subject matter did more of the work, and that is a guess.
Their follower count also grew across the same period, so part of the lift is simply more people seeing the posts, not a higher rate among the same audience. Engagement counts without a follower baseline overstate the change.
And a caution on the research: the Journal of Consumer Research finding measures what a visible AI disclosure costs. Our client never labelled anything. So that study is not a direct measurement of their situation, and I am using it for scale rather than as a like for like comparison. What it establishes is a ceiling. If disclosed AI content only costs single digit percentages, undisclosed generated content is not what takes an account to one like.
What to do with this
- Look at your last ten posts and ask which of them a competitor could have published unchanged. If it is most of them, that is the problem, and it is not a tooling problem.
- Find the one true sentence about your practice that your nearest competitor cannot say. Ownership, tenure, hours, a piece of equipment, a person. It only has to be true and checkable.
- Get photographs of your own people. A phone is fine. The doctor on the floor with the puppies outperformed everything produced to a brief.
- Check what your current service already includes before buying anything from anybody, including me. Social is bundled into more packages than people realise and very often has simply never been switched on.
- Give whoever writes your posts something to work with. Nobody outside the building knows that your technician has been there eleven years unless somebody tells them.
The uncomfortable version of all this: the content was never the bottleneck. The information was. A generated post and a written one are equally empty if nobody has told either of them what is actually true about you.