Specific figures beat vague copy in every one of 60 GPT-5.4 trials

John Lawson published a controlled test on LinkedIn on 15 September, running 15 source pairs through GPT-5.4 across 60 calls. Each pair was word-for-word identical except that one version carried exact figures and the other carried vague descriptions. The specific version was cited 100% of the time.

Positions were swapped between runs to rule out order effects. Five control pairs with identical wording saw the first-shown source picked every time, which tells you position matters when nothing else distinguishes the sources. Once one source carries numbers, that signal overrides position entirely in this sample.

Lawson's Harrier 8 example makes the mechanism plain. "8mm heel-to-toe drop" and "260 grams in a UK size 9" are phrases a model can lift into an answer. "Fairly low drop" and "light for its size" are not.

What to check on your own pages

Before acting on this, pull ten pages that describe your product's specifications and read them the way a model would. Where you have written "lightweight" or "long battery life" or "fast charging", find the number that sits behind the phrase and put it in the sentence. If the number does not exist, that is the work.

We think the finding generalises beyond spec sheets to any claim a model might quote back. Prices, dates, sample sizes, version numbers, dimensions. Sixty trials on one model is a strong signal on one model, so run the same test on the engines you care about before rewriting at scale.

The reason to add the figures is not that the model prefers them. It is that a reader deciding between two shoes wants the drop in millimetres, and the page that gives it to them is the better page.