Broken Pages Cost Citations. Redirects Do Not.
We tracked the same pages daily for a month to see what happens to a citation when an address moves or breaks. The answer reverses the usual migration anxiety.
Key Takeaways
- A redirected page keeps 94% of the staying power of a working page.
- A broken page keeps 59%, and is seven times less likely to hold a citation for fifteen days or more.
- Redirected pages show no decay across the month. Their share of citations holds steady.
- During a migration the redirect is not the risk. The page that quietly stops working is.
- Pages blocked to crawlers keep their citations normally. That is a reporting problem, not a performance one.
Why we looked into this
Every site migration produces the same argument between engineering and marketing. Marketing wants every old address redirected. Engineering points out that redirects add complexity, and that some of these pages have not been visited in years. Somewhere in the middle, a set of pages ends up returning an error instead, and nobody notices for months.
In traditional search that trade off is well understood. What nobody had measured is what it costs in AI answers, where the engine is not crawling your site on a schedule but reaching for a source at the moment it answers a question.
There is a reasonable theory that redirects would hurt badly here. The engine holds an address, follows it, finds it has moved, and might reasonably drop it in favour of something that resolves cleanly. We tested that theory, and also the alternative: that a broken page is the thing that actually costs you.
How we tested it
Because every question was asked daily for a full month, we can measure something a single crawl cannot: how consistently a page holds its place in an answer over time.
For every combination of a page, a question and an engine, we counted on how many of the 31 days that page was quoted. A page quoted on 28 days is a fixture. A page quoted once and never again is noise.
We then visited every quoted address and recorded what it actually returned: a working page, a redirect to somewhere else, or an error. Comparing staying power across those groups answers the question directly.
One group needed separating out carefully. A large number of pages return a block to automated visitors, which looks like a failure but is not: those pages work perfectly well for people and for AI engines. Counting them as broken would have contaminated the result badly, so we identified and excluded them. Their staying power turns out to be indistinguishable from healthy pages, which confirms the separation was correct.
What we found
The theory about redirects is wrong, and the alternative is right.
A redirected page holds its place almost as well as a page that works normally. It keeps 94% of the staying power of a healthy page, a difference small enough to be unremarkable.
A broken page keeps 59%. And when you look at the pages that hold a citation for a long stretch, the gap widens dramatically. 8.3% of working pages are still being quoted after fifteen days or more. For broken pages that figure is 1.2%, seven times lower.
The pattern over the month
We also checked whether redirected pages fade as the month goes on, which is what you would expect if engines were gradually dropping them. They do not. Their share of citations holds steady week after week, and they stop being quoted around two days earlier than healthy pages across a 31 day window. That is not decay, it is noise.
Why this happens
A redirect is a complete answer to the question the engine asked. It says the content exists and it is now here. The engine follows it, retrieves the content, and the citation survives with the address updated.
An error is not an answer. There is nothing at the end of it, no forwarding instruction and no content. The reference the engine holds is now pointing at nothing, and once that is discovered there is no reason to keep returning to it.
The distinction is not really about technical status codes. It is about whether the content still exists somewhere the engine can be told about. Redirects preserve that relationship. Errors sever it.
What this means for your business
The practical instruction reverses the usual anxiety about migrations. The redirect is not the risk. The page that quietly stops working is the risk, and it costs far more than most teams assume.
This reframes broken link cleanup, which is normally treated as housekeeping and scheduled accordingly, meaning it never gets done. It is not housekeeping. It is the difference between a page that keeps earning citations and one that is dropped seven times faster. That is a considerably easier case to take to an engineering team than the usual tidiness argument.
Three things follow for anyone planning a migration. Redirect everything that had any external presence, and do not agonise over the performance cost, because the data says the citation survives. Audit for errors afterwards with more urgency than for redirect chains, since errors are what actually destroys value. And check the pages that no longer have internal links pointing at them, because those are the ones that break unnoticed and they may still be earning citations you cannot see in your analytics.
One more point worth knowing. Pages that block automated visitors, which includes a large amount of legitimate content behind protection services, keep their citations perfectly well. If your site sits behind protection that blocks crawlers, that is not costing you AI visibility. It is only costing you the ability to audit your own presence, which is a reporting problem rather than a performance one.
We treat broken URL cleanup as visibility work rather than housekeeping at Novastacks.
Methodology
89 questions agreed with chief executives and business owners as ones their companies need to monitor, put to ChatGPT, Google AI Overview and Perplexity every day through August 2026, producing 61,536 citations. Staying power is measured per combination of page, question and engine across 31 days, which controls for how many opportunities each page had.
Every quoted address was visited and its response recorded. Pages returning an error and pages returning nothing at all are grouped as broken, giving 1,756 combinations. Redirects account for 247. Pages blocked to automated visitors are excluded and analysed separately.
Our visit to each page happened after the measurement month, so this compares addresses that redirect now against addresses that work now, rather than tracking a single page through a migration. All brands are anonymised.
Frequently Asked Questions
Do redirects hurt your AI citations?
Barely. A redirected page held its citation for an average of 3.98 days across a 31 day month, against 4.25 for a page returning normally. That is 94% of the staying power of a healthy page.
What happens to AI citations when a page breaks?
They fade fast. Broken pages held a citation for 2.51 days on average, and only 1.2% were still cited after fifteen days, against 8.3% of healthy pages.
Should I redirect old pages during a site migration?
Yes. On this evidence the redirect costs almost nothing while letting a page break costs a great deal. The risk in a migration is the page that quietly stops responding, not the redirect.
Does blocking crawlers hurt AI citations?
Not in our data. Pages behind protection services that block automated visitors held citations at 7.7%, essentially the same as healthy pages at 8.3%. It affects your ability to audit, not your visibility.
How was this measured?
Staying power is measured per combination of page, question and engine across 31 days, covering 61,536 citations. Every quoted address was then visited and its response recorded.
Is Your Site Quietly Losing Citations?
We check which of your cited pages still resolve, which redirect and which have broken, then show you what it is costing you in AI answers.