Article · content
AI has a recency bias, and your 2021 blog post is paying for it
AI engines prefer recent sources and they show the dates. When they quote your three-year-old price or your discontinued service, the customer arrives with the wrong expectation and you look careless. Stale content used to cost you rankings slowly. Now it costs you trust instantly.
A prospective customer rang a client of ours and asked why the price had gone up so much. It had not. ChatGPT had quoted a figure from a 2021 blog post titled "How much does X cost?", which was still live, still ranking, and three years out of date. The current pricing page was fine. The engine preferred the old post because it answered the question directly and the pricing page did not.
The stale page did not just fail to help. It actively misrepresented the business to someone who was ready to buy.
Why engines lean recent
Classic search could afford to be patient with old content; a well-linked page from 2019 could rank for years. AI engines answer questions, and an answer about price, availability, regulation or best practice has a shelf life. So they weight recency, and they show it: Perplexity and ChatGPT search display dates next to citations, and users see them. Several 2026 analyses found pages updated within the last month were cited around three times as often as stale ones. The exact multiple varies by study; the direction does not.
There is a second mechanism. Engines cross-check facts across sources. When your 2021 post says £400 and three competitors' 2026 pages say £650, the engine does not just prefer the newer figure; it downgrades its confidence in you as a source.
The particular damage stale content does in AI search
In classic search, a stale page mostly harmed itself: it drifted down the rankings and stopped sending traffic. In AI search it can harm the whole business, because the engine quotes it as the answer:
- An old price is quoted as your current price, and the enquiry starts with an argument.
- A discontinued service is recommended, and you have to turn the customer away.
- A superseded regulation or best practice is attributed to you, and you look out of date to a buyer who knows better.
- An old team member is named as your expert.
- A "2023 guide" ranks for the question and the engine cites it with the date visible, which is a small announcement that you have not looked at your own site in a while.
None of this shows up in a traffic report. The pages that do the damage often have almost no visitors. They just have the right question in the title.
The date-bumping trap
The obvious shortcut is to change the date. Do not. Engines compare versions and compare sources; a fresh date on unchanged text that contradicts newer facts elsewhere does not become true, and a pattern of cosmetic date changes is a trust signal in the wrong direction. Worse is the yearly-URL habit: "prices-2024", "prices-2025", "prices-2026" as separate pages, each thin, each splitting authority, the older ones still live and still quotable.
An update means the content changed: new figures with their source, new prices with the date they apply from, a corrected claim, a new example, a removed service. Then the date changes, visibly and in the schema.
What the discipline looks like
Inventory the site: every URL, its last real update, its traffic, whether it is a money page, whether any engine currently cites it. The top 20 to 30 pages are the refresh set; they get a quarterly review with a named owner and an immediate update when anything about the offer changes.
Everything else gets one of three treatments. Consolidate: several old posts on one topic become one current page, with redirects. Prune: outdated, thin pages that contradict the current offer are removed or set to noindex, with a redirect to the nearest relevant page. Leave: evergreen pages that are still accurate get a review date and nothing more.
Show published and updated dates on every page and keep dateModified honest in the schema. Update lastmod in the sitemap and use IndexNow so Bing, and therefore ChatGPT search, picks the change up quickly.
The reframe
Most businesses treat their site as an archive: everything published stays published. In AI search the site is a set of claims the engines will make about you, in your name, to customers you will never see. Every page is either current and correct, or it is a liability with a title that matches a question.
The client's 2021 post is now a redirect to the pricing page, which now answers the question in its first sentence. The engine quotes the right number. That is the whole job.
Frequently asked questions
How much does freshness affect AI citations?
+
Several 2026 analyses found pages updated within the last 30 days were cited roughly three times as often as stale ones. Perplexity and ChatGPT search both display dates beside citations. Treat those figures as directional, but the direction is consistent across every study we have seen.
Should I delete old blog posts?
+
Consolidate first. Merge posts on the same topic into one current page and redirect the old URLs. Delete or noindex only the pages that are outdated, thin and contradict your current offer. Every stale page that contradicts your current pricing or services is a source of wrong answers about you.
Sources
George McKenna
Co-founder, Emerging Digital Partners
George co-founded Emerging Digital Partners in Eastbourne and built the AI Search & SEO Audit you are on now. He spent around twenty years in the UK IT channel before that, most recently running solution sales teams, and now splits his time between building websites and search tooling for EDP clients and taking AI products to regulated industries through the ETT Group, where he is Chairman and CTO. He writes about answer engine optimisation the way he practises it: test it, measure it, fix what is actually broken.
Read next
No author, no proof, no chance: the trust gap on small business websites
Most small business websites ask to be trusted without offering a single checkable reason. Anonymous posts, stock photos, no team page, no dates, no credentials, no named clients. A human visitor forgives that. An AI engine deciding whether to recommend you does not, because it has nothing to weigh.
AI cannot quote a homepage that says nothing
'We deliver innovative solutions that empower your business.' An AI engine reads that and learns nothing it can use. The vague-copy problem was tolerable when a human read past it to the contact form. It is fatal when the reader is a machine deciding whether to recommend you.
You are measuring the wrong things
Most marketing dashboards still report rankings and sessions, and neither tells you whether AI engines recommend you. AI referrals land in Direct, citations never register as a visit, and nobody checks whether the crawlers can get in. Here are the four measures a 2026 dashboard needs.