Obility Editorial · · 4 min read
Language models do favour a recent date, at least in the lab
A common piece of AEO advice this year is to change the “last updated” date on your articles, because AI assistants supposedly prefer fresh pages. There is real evidence behind part of that claim. In September 2025, researchers from Waseda University and the Hong Kong Polytechnic University published a paper on arXiv testing whether language models favour newer documents when they rank search results.
They took passages from two standard TREC retrieval test sets, asked models to order them by relevance, then repeated the task with a line such as “Published on: 2025/01/01” added to the front of each passage. Nothing else about the text changed. All seven models they tested moved recently dated passages up: GPT-3.5-turbo, GPT-4, GPT-4o, two Llama 3 models and two Qwen 2.5 models. The ten top results became up to 4.78 years newer on average, and single passages moved by as many as 95 places. In a second test using pairs of passages that human judges had rated equally relevant, adding an old date to the preferred one and a new date to the other reversed about a quarter of the smallest Llama model’s choices. The largest Qwen model reversed about 9 percent. Bigger models were less affected, and none were immune.
The experiment does not describe ChatGPT or AI Overviews
Read the method before acting on the headline. The dates were plain text placed directly in front of short passages, spread one year apart from 1926 to 2025, and fed to models that are now one or two generations old. No production answer engine was tested. OpenAI, Google, Anthropic and Perplexity do not publish whether a page’s date reaches the model that chooses sources, in what form, or how much weight it carries. The paper shows a tendency worth knowing about. It does not show that editing a date on your site changes what ChatGPT cites.
Vendor experiments do not settle it either. OtterlyAI changed only the visible “last updated” date on ten of its own articles in August 2026 and reported that their AI citations rose 31.5 percent over the next 30 days, while two control articles declined. Its own write up calls the result directional rather than conclusive, because citation trends were already moving before the change. Ten pages and one month cannot separate a date effect from ordinary drift in AI answers.
Google already treats a cosmetic update as a warning sign
Google is the one engine operator that has written about this directly. Its guide to creating helpful content lists warning signs that a site is writing for search engines rather than people, and one of them asks whether you are changing the date of pages to make them seem fresh when the content has not substantially changed. Its documentation on byline dates says its systems do not depend on a single date factor, and instead estimate when a page was published or significantly updated from several signals. AI Overviews and AI Mode draw on that same index. A date that disagrees with the page around it is something Google says it is built to look past.
There is also a cost on the reader’s side. A buyer who sees “updated this month” above 2023 prices and a discontinued plan trusts the page less. An assistant that takes the date at face value may repeat those old prices as current, which is a worse outcome for you than not being cited.
Make the date true by changing what it describes
The useful version of a refresh starts with the facts buyers ask about. Check prices, plan names, integration lists, screenshots, comparisons and any statistic with a year attached. Add answers to questions buyers have started asking since the page was written. Then update the date, because it now describes a real change. If the recency tendency in that paper does carry into production systems, a page that has genuinely changed gets the benefit honestly, and it survives any check that compares the date with the content.
Keep the mechanics clean. Google’s byline guidance asks for a clearly labelled visible date, such as “Updated” or “Last updated”, and for the visible date to match datePublished and dateModified in your structured data. It also asks you not to use future dates. Keep the original publication date and add an updated date beside it rather than replacing it, so readers can see both.
Measure a refresh like an experiment
If you want to know whether refreshing pages helps you, test it on your own prompts. Fix a set of buying questions, record which pages each engine cites for a few weeks, then update one group of pages and leave a comparable group untouched. Compare how often each group is cited over the following weeks, not what one answer said on one day. Obility’s Answer Engine Insights is built around this kind of comparison: prompts are grouped by audience, buying stage and market, the wording and context stay attached to each observation, and a gap can be turned into a content review with a follow-up measurement.
Know the limits of what that tells you. A refresh changes many things at once, so a lift shows the update worked, not which part of it did. Engines also need to fetch the page again before any change can count, and none of them say when that will happen. What you can control is whether the page is accurate when they do.