Technical rescue · the long form
Tens of thousands of URLs Google could not usefully read.
A brokerage’s own listings software had generated tens of thousands of near identical pages, sold homes, expired listings, every filter combination its search tool could produce, and buried the 237 pages the business actually runs on.
The scale is what makes it a systems problem
A handful of bad pages is a checklist fix. Tens of thousands is not, because a checklist has no way to tell a page worth keeping from a page the listings vendor spat out an hour ago. Google could not tell either, and treated the whole domain as noisy rather than trustworthy.
This is the point where most providers stop returning calls, because the honest first step is not writing anything. It is counting.
How the count becomes a decision
Separate the real site from the exhaust
Every URL pattern the listings tool could generate gets mapped, so the audit can say which shapes of page are the business and which are the tool's own filter combinations arriving as pages nobody asked for.
One decision per page, written down
Keep, noindex, redirect, or block from being generated at all. Not a batch action. A record a developer can run and a client can read afterward and see exactly what happened to what.
Nothing gets deleted in a panic
A page a competitor's tool made can still be a page a real buyer searched their way to. Deleting first and checking later is how a rescue turns into a second injury. Every move here is reversible.
The vendor's part gets written down and chased
Some of the faults only the listings vendor can fix, because the pages are generated by their system, not the site's own. That work gets packaged into an action sheet and driven to the people holding the keys, not left as a line item nobody owns.
The number that moved twice, and why both readings are honest
Referring domains for this site read 615 on 9 August 2026. The same instrument, Ahrefs, read 570 on 2 September. That is not the count falling because something went wrong. Ahrefs periodically drops domains from its own index that it decides have gone dead, and a chunk of the earlier 615 was made of exactly that kind of link. The 570 is not a worse number. It is a cleaner one, read on a later date, by the same tool, and it is the figure this site's own case study quotes because it is the more current reading.
The reason to write this down at all is that a number moving between two dates is the first thing a nervous client asks about, and the honest answer is almost always the instrument doing its job, not the work coming undone.
Nothing here runs on theory.
A site most operators would abandon can end up outperforming its own promise, once the count is honest.