How Often to Refresh Content for AI Search

The standard advice is to refresh every page on a 90 day clock, and the largest study of AI citations says that clock is mostly wasted motion. We have published more than 250 articles here and tracked which ones AI engines keep quoting back. Below, what 16.975 million cited URLs show about content age, the 4 page types that genuinely decay, and the schedule we run instead of a calendar.

How Often Should You Refresh Content for AI Search?

That is the whole answer, and almost every guide on this topic gets to a different one because it starts from the wrong question. The common framing is how often a page should be updated. The useful framing is what on the page has stopped being true.

Those two questions produce very different work plans. The first produces a spreadsheet of 200 URLs and a quarterly reminder. The second produces a list of 30 pages that carry a number, a price, a benchmark, or a claim about how a tool behaves, and a plan to fix those before anything else gets touched.

Content refresh

A substantive revision of an already published page that changes what the page says, not just when it says it was said. Substantive means new or corrected figures, a section added to answer a question the page was missing, claims removed that are no longer accurate, or a rewritten summary block. Changing the visible date or the dateModified property without changing the text is a date change, not a refresh, and retrieval systems treat it as such.

Content decay

The gradual loss of accuracy or usefulness in a published page as the facts around it move. Decay is driven by what the page claims rather than by how old it is, which is why a 4 year old definition can be perfectly current while a 6 month old benchmark post is already wrong.

If the wider frame is new to you, start with what generative engine optimization is and GEO vs SEO differences explained, which cover how the target moved from a ranked list to a generated paragraph.

Does Content Age Actually Affect AI Citations?

The number that should stop you is 1,064 days. That is close to 3 years. An engine picking sources for an answer is not reaching for last month's post, it is reaching for the page that answers the question cleanly, and on average that page was published a long time ago.

The Ahrefs study of whether AI assistants prefer fresh content is the largest public dataset on this, and the second half of it is the part that gets skipped. Measured by publish date, AI citations run 25.7 percent fresher than organic. Measured by last update, the gap shrinks to 13.1 percent. Updating narrows the difference rather than eliminating it, which is what you would expect if freshness is one input among many rather than a gate.

The per-engine split matters more than the average, and it is the single most actionable finding in the dataset. ChatGPT cited the newest material at 958 days average age. Perplexity sat at 1,166 days. Google AI Overviews came in at 1,432 days, which is effectively identical to organic results and means AI Overviews shows close to no freshness preference at all.

Read that as a budgeting instruction. If most of your visibility comes through AI Overviews, a fast refresh cycle buys you very little and the same hours spent on answer capsules or entity consistency will move more. If you are chasing ChatGPT specifically, recency is worth something, though 958 days is still not a 13 week window.

Ahrefs also found that on both Perplexity and ChatGPT, newer content tends to be cited earlier in the reference list. Position inside an answer is worth having, so recency is not worthless. It is just a tiebreaker between pages that already answer the question, and no amount of it rescues a page that does not.

Why Does the 13 Week Rule Keep Getting Quoted?

The 13 week figure is everywhere this year. Trace it back and it moves through secondary sources rather than landing on a published dataset with a stated sample size. Quattr's write-up of the content decay cycle is one of the more careful versions, and it attributes both the half-under-13-weeks figure and the often quoted 3.2x multiplier to third parties rather than to its own analysis. Salespeak's piece on content freshness is the origin most of the chain points at.

None of that makes the figure wrong. It makes it a figure you should hold loosely and never build a quarterly budget on, per the sourcing standard we apply to every number we publish.

The honest reconciliation is that citation freshness is a property of the question, not of the web. Ask an engine what SPF means and it will quote a stable explainer written years ago. Ask it which cold email tool is best in 2026, or what the current reply rate benchmark is, and it will reach almost exclusively for recent pages, because the older ones are visibly stale to the model.

• Time-anchored queries carry a year, a version, or a word like current, latest, or now. These pull recent pages hard, and a page that has not been touched in 18 months is unlikely to be selected.
• Benchmark and pricing queries behave the same way without the explicit date, because the answer is a number that everybody knows moves.
• Definitional and mechanical queries barely care. What domain reputation is has not changed, and the page that explains it best wins whether it was written last week or in 2022.
• Comparison queries sit in the middle. The category is stable, the specific products are not, so a 2 year old comparison that names a product which has since changed pricing gets passed over.

Which means the first refresh decision is not a cadence at all. It is sorting your library by which of those 4 buckets each page sits in, and there is no tool that does it for you.

Which Pages Actually Need Refreshing, and Which Do Not?

Here is the cadence we actually run, sorted by what drives the decay rather than by page template.

The last row is the one that saves the most time. Teams put a page on the refresh list because it is old, run the update, and get nothing back, because the page was never eligible for a citation in the first place. It had no extractable block, no direct answer, and no reason for an engine to choose it. Adding 200 fresh words to that page changes nothing. The diagnosis for it is in why AI answers cite some companies and not others.

The second most expensive mistake is refreshing case studies. A case study is a dated record of something that happened, and its value comes from being specific and verifiable. Updating the numbers in a client result to make the page look current turns a receipt into marketing copy, and readers can feel it even when they cannot name it.

What Counts as a Substantive Refresh?

Google has been explicit about this for years. Its Search Central guidance on helping Search know the best date for a page says to give an article a fresh date when it has been substantially changed, and not to artificially freshen a story without adding significant information or some other compelling reason. Search Engine Land's guide to byline dates covers what Google does and does not use them for.

The mechanical reason a date change does nothing for AI answers is simpler than the policy reason. A retrieval system splits your page into fragments and embeds the text of those fragments. If the text did not change, the embedding did not change, and the same fragment competes for the same questions exactly as well as it did yesterday. You moved a timestamp and left the content untouched, so nothing downstream has anything new to work with.

What we change on a real refresh, in order:

1. Every number, and the source behind it. If a figure cannot be traced to a publisher, it comes out rather than getting hedged. A soft claim with no source is worse than no claim, because it is the sentence an engine will not quote and a reader will not trust.
2. The answer capsule. This is the block most likely to be lifted verbatim, so it gets rewritten first and tested by reading it with the rest of the page hidden. If it does not stand alone, it is not finished.
3. Questions the page never answered. Pull them from real conversations, not a keyword tool. The follow-up questions buyers ask on calls are the ones missing from almost every page in the category.
4. Claims that quietly went false. These are the dangerous ones, because nothing flags them. A tool changed its free tier, a platform retired a feature, a benchmark moved, and your page still states the old version with total confidence.
5. Internal links to work published since. A page written 18 months ago cannot point at the 40 articles that came after it, and those links are how a library reads as one connected corpus instead of 250 loose pages.
6. The dateModified value, last. After the content moved, not instead of it.

One warning on scale. If you are refreshing 40 pages, resist the urge to run the same 3 edits across all of them. ZipTie's content refresh strategy for AI citations makes the point well: a batch refresh that touches every page the same way produces a set of pages that all read like each other, which is the opposite of what you want when engines are selecting one source out of many.

How Do You Build a Refresh Schedule That Is Not Just a Calendar?

The tiering pass takes an afternoon and it is the only part of this that has to happen before anything else.

1. Export every URL with its publish date and last update date. No judgment yet, just the list.
2. Tag each page by decay driver using the table above. Live numbers, product facts, platform rules, a year in the title, or none of these.
3. Mark which pages are currently cited. Run your question set through the assistants and record which URLs come back. This is the field that turns the list into a priority order.
4. Refresh cited pages that carry live numbers first. You are defending a position you already hold, which is cheaper than winning a new one.
5. Then uncited pages that carry live numbers. These are wrong on the internet under your name, which is a reason to fix them even if no engine ever quotes them.
6. Leave the stable tier alone until something breaks. Put a 12 month verification check on it and spend the recovered hours on new pages, which is almost always the better return.

The ratio is worth naming because it is what keeps the work finishable. On our own library of 250-plus articles, the tier that genuinely needs quarterly attention is small. Most of what we publish is mechanical: what a term means, how a system works, what the steps are. That material does not rot on a schedule, and treating it as though it does is how a content team ends up with no capacity to publish anything new.

There is a second-order benefit to publishing new work instead. Every new article is another page competing for its own set of questions, and it can link back into the older library, which is the mechanism that made ranking in AI search work for us at all. A refresh defends one page. A new page adds one and strengthens several.

Keep the machine-readable layer in step with all of it. If a refreshed page now states a different number from your llms.txt or your LLM info page, you have handed an engine two versions of the truth with no way to choose between them, and it will pick whichever it saw last. The same applies to schema markup on podcast pages and to the sitewide entity blocks covered in the generative engine optimization checklist.

How Do You Know a Refresh Worked?

Most teams cannot answer this question, which is why refresh programs get cut in the second quarter. The work is real, the reporting is not, and anything that cannot be shown to work eventually stops being funded.

The loop we run is deliberately boring. Fix a list of 20 to 40 questions a real buyer would type. Run them across ChatGPT, Perplexity, Google AI Mode, and Claude on a set cadence. Record 3 fields per run, and treat the quoted sentence as the primary field rather than the yes or no.

That third field is the one everybody skips and the only one that makes the next edit an edit rather than a guess. Knowing you appeared is a score. Knowing which 50 words got lifted tells you exactly what the engine considers quotable in your category, and after a few months that log is more useful than any tool you could buy. AirOps has a good writeup on measuring whether a refresh improved LLM visibility if you want a second framework alongside it.

Set the timing expectation before you start or the program dies of impatience. Crawling, indexing, and re-embedding all lag, and a model that already has a version of your page cached will keep quoting it for a while after you change it. A fair first read is 4 to 8 weeks out. Our full setup is in how to track AI search visibility, the brand-level version is in how to audit your brand in ChatGPT, and the reason we do not judge any of this on sessions is in LLM citation vs SEO traffic.

Scale is the reason the measurement is worth building at all. Semrush's expanded 2026 AI Visibility Index analyzed 126 million US AI search prompts and found ChatGPT citing an average of 15 sources per response. That is 15 slots on a single answer, and a refresh that moves you from slot 12 to slot 3 is worth more than most teams assume. Platform-specific patterns are in how to get cited by ChatGPT and Perplexity, how to get cited by Perplexity, and AI Overviews for B2B.

One caution on attribution. If you refresh a page and also add schema in the same week, you cannot tell which one moved the result, and it is probably neither. Ahrefs matched 1,885 pages that added JSON-LD against roughly 4,000 control pages and found no meaningful citation uplift, as Search Engine Roundtable reported. Change one variable at a time or you learn nothing from either.

What Does This Look Like on a Podcast Library?

This is where the whole argument stops being theoretical for us. Every client show produces recordings full of specific, first-hand answers from people who actually run the thing being discussed, and that is exactly the material engines want to quote. In raw transcript form, almost none of it is usable.

An episode page from 2 years ago does not need a fresher date. It needs the buried answers surfaced. The real answer is at minute 34 inside a sentence that starts with "yeah, so", and no retrieval system is going to find it there. Pull the questions the episode answered, write each as a heading, write a 60 word answer underneath that stands alone, and a wall of transcript becomes 6 retrievable sections each competing for a different query. The mechanics are in turning episodes into answer content, podcast transcripts for AI search, and how to repurpose podcast episodes.

The one thing on an episode page that does decay is the guest's own claims. If a guest quoted a number in 2024, that number is now 2 years old and should carry the date it was said rather than being presented as current. That is an accuracy fix, and it is the only reason we touch an old episode page at all. How to get your podcast cited by AI assistants covers what that looks like across a full library.

Worth being straight about where this sits for us. The search footprint is the second return, not the first. The invite is the acquisition channel. We invite a client's ideal buyers onto their own show by email, the recording is the conversation, and our commitment is 30 recorded conversations with those buyers in 90 days or their money back. Editing is included, the client owns the show and every recording, and invites go out by email only. The citable library that builds up afterward is a second return on work that was already done, which is the honest way to frame it. More on that in what reverse outbound is and what a podcast does for your search footprint.

If you run this alongside cold email, the two disciplines share a habit worth naming. Both reward specificity, both punish generic work, and both take weeks to read properly. If the sending side is newer to you, start with what email deliverability is, SPF, DKIM, and DMARC explained, how to warm up a new email domain, what domain reputation means, and deliverability monitoring. Getting your ideal customer profile right is the input to both.

The Part Nobody Wants to Hear

A refresh program is the easiest content work to sell internally and the hardest to justify afterward. It is legible, it fills a quarter, it produces a spreadsheet that goes green, and most of it does not change a single thing an engine decides. The 90 day clock is popular because it removes the need to think about any individual page, and that is precisely the thinking the job requires.

The unglamorous version is this. Sort your library once by what would make each page wrong. Fix the pages carrying numbers, because those are wrong on the internet under your name whether or not anybody is reading them. Leave the stable ones alone and stop feeling guilty about it. Then spend the recovered hours writing pages that answer questions you do not currently answer at all, because a page that does not exist cannot be cited and a page that does not decay does not need you.

Most competitors will run the clock anyway. They will spend the quarter updating explainers that were already correct, log the work, and see nothing move, and they will conclude that content refresh does not work for AI search. It works fine. It just only works on pages that were decaying, and the gap between those two ideas is where the next year of visibility gets decided.