Position bias
Position bias is a language model's habit of using information at the start or end of what it reads better than information in the middle.
Position bias is the tendency of a language model to make better use of information at the start or the end of what it's given than information in the middle. If the fact the model needs sits halfway through a long page, or halfway through a long pile of search results, it's more likely to miss it.
For a marketer, the practical lesson is simple: put the thing you most want an AI platform to repeat where it's easiest to find.
Where the idea comes from
The best-known test was a 2023 research paper, Lost in the Middle. The researchers gave models long inputs with the answer placed at different points. The models did best when the answer was near the beginning or the end, and noticeably worse when it was in the middle.
Models have changed since then, and newer ones may handle long inputs better. Putting your main point first costs nothing either way, and people skim the same way.
What it means for your pages
When a platform searches the web, it reads several pages and often only parts of them. If your page opens with three paragraphs of background and states your price in the sixth, the model may never weigh it.
For example, Harrow & Pine's delivery page could start with "We deliver across the UK in two working days" and explain the details after. A page that buries that sentence under its company history gives the model less chance to pick it up.
So lead with the answer. Put the key fact in the first lines of the page and at the top of each section, and keep each section about one thing.
Not the same as your position in an answer
Position bias is about where information sits in what the model reads. It's a separate idea from position in AnswerOrbit, which is the order in which brands are named in the answer the model writes. A page that leads with your key fact may help your position, but the two are measured in different places.
The Page checks in Recommendations include a check called An opening that answers, which looks at whether a page answers the prompt in its first lines.