Testing whether I should reuse old content
I fit a hierarchical Bayesian model to test whether retitling old content for AI citation works, and a robustness check collapsed the one clean result it found.
Ben Rees - 7 September 2026
Retitling old content to match how people phrase questions to AI systems is standard GEO advice. I've run it five times. It worked once. A proper statistical check just told me that "once" was doing all the work.
That's not a failure of the tactic. It's the difference between a story and evidence, and I only know which one I had because I went looking.
The one page that moved
On 2026-07-11 I retitled three bjrees.com pages to match three Scaling B2B Marketing queries that had gone nowhere for weeks: "how to build a B2B marketing engine from scratch," "B2B marketing strategy for scaling companies," and "challenges of scaling B2B marketing past series B." All three sat at zero appearances in Klara, the AI visibility tracker I've written about before.
Four weeks later, two of the three were still at zero. The third, an existing 10-point guide to scaling from SMB to enterprise, started appearing for "past series B" in Google AI Mode on 2026-08-08, then organic search the next day. Two consecutive positive readings, AI Mode first, is a real pattern, not noise. It's also one page out of three.
So I built a model instead of trusting the story
One clean win makes a good story. It doesn't prove a tactic generalises, and I've written before about AI visibility as a kind of Bayesian belief updating. So inside my project named Hammerstein, I built the actual model: a hierarchical Bayesian logistic regression of citation probability on query, category, platform, time, and intervention type, fit against 4,136 real capture observations across 63 queries since April. Four chains, clean fit, zero divergences.
Before trusting it on anything new, I checked it against facts I already knew. Branded SEO sits miles above everything else. Scaling B2B Marketing is the strongest non-branded category. Both came out exactly as expected.
The model said yes, convincingly
Then I asked the real question: does retitling move citation probability? The first answer was an emphatic yes. beta_retitle = +1.82, 94% credible interval [0.56, 3.10], 99.9% probability the effect is real and positive. About as clean a result as a marketer gets.
Then I removed the one page that actually moved
Only four retitle events had enough data to inform that estimate, and one was the "past series B" page, the only one that had actually ignited. So I ran the obvious check, a leave-one-out check: refit the model with that query's rows removed, and see whether "retitling works" survives losing its best example.
It didn't. beta_retitle drops to -0.50, [-2.50, 1.45], 31% probability of a positive effect. Statistically indistinguishable from the model's finding on new-content publishing: no credible effect at all.
Why one data point can do that
This isn't a bug in the fit. It's what a hierarchical model does when it pools across too few groups: with four retitle events, one outsized result gets amplified rather than averaged away. Tversky and Kahneman described the human version of this in 1971: people, including trained researchers, expect a small sample to look like the population it came from, and draw conclusions from four data points with the confidence that should require forty. My model didn't make that mistake on its own. I nearly did, by stopping at the first number instead of checking whether it held.
It's a specific case of an older problem. Ioannidis's 2005 paper on why most published research findings are false makes the point that a significant result usually isn't wrong because the maths is wrong. It's wrong because the maths gets trusted at exactly the moment it's least earned: early, on a promising, underpowered result, before anyone checks whether it holds.
What I'm actually doing next
Not retitling more pages expecting a general lift, and not writing this off as null either. What I have is precise: one confirmed case of a retitle moving citation probability, zero confirmed cases of anything else doing it. Two more retitles from mid-August are still pending, their real clock reset to 2026-09-04 by a sitemap bug that kept them uncrawlable until then. A check on 2026-09-18 either turns "one win" into a replicated effect, or leaves it exactly where it is. I'll know which before I touch another title tag.
Related reading
Your scaling story is your most AI-retrievable content and the most interesting
Your B2B scaling story is your most differentiated content asset. Here's why it stays invisible to AI systems and what to change.
Why AI systems cite some SMB-to-enterprise stories and ignore others
AI systems rarely cite customer case studies that only live on the vendor's own site. Citation comes from independent corroboration, not content quality - here's what to do about it.
My AI visibility went from zero to 26% in six weeks. The one category still at zero is the one this post is about.
Six weeks after publishing a zero-appearance rate across every AI platform, most categories moved. The one still at zero is AI Search Visibility itself.
Running AI locally doesn't change what you know. It changes what you can keep private.
Local models protect your data but can't change what external AI systems believe about your brand. That's still a retrieval and training problem.