Guide
How fast can an AI assistant stop citing the page it cited last month?
Fast enough that a measurement taken in July was already wrong by the middle of August. Reddit’s share of ChatGPT Search citations averaged 3.83% between 18 July and 7 August 2026, then fell to 0.52% by 17 August, after OpenAI changed how ChatGPT selects and queries sources on 8 August. That is an 86% drop inside ten days, and nobody whose page had been cited was told it was coming (Axios, 2026; Forbes, 2026).
The mechanism is straightforward once you see it. ChatGPT Search does not read a fixed index the way a search engine does. When a user asks a question, it runs its own fan-out of sub-queries, decides which domains to interrogate directly using site: operators versus which to search openly, and assembles an answer from whatever that pass turns up. On 8 August, OpenAI’s use of site: operators aimed at specific domains jumped from around 0.4% of fan-out queries to close to 17%, and kept climbing through the following week. Nothing about Reddit’s content changed. The route ChatGPT takes to find sources did.
Why a citation reading has a shelf life
This matters operationally, not just as trivia. A citation pattern is a reading of the system’s current source-selection behaviour, and that behaviour is set by the platform on its own release schedule, which nobody outside it can see coming. The 8 August change was not announced in advance; it only became visible afterwards, once someone was tracking the numbers closely enough to notice the date it happened. There is no reason to expect the next change to arrive any differently.
Which is the whole argument against measuring once. A business that ran a citation check in July and saw itself cited fairly often through Reddit threads about its category would, by mid-August, have been looking at a stale picture without anything on their own site having changed. The gap between what was true and what they believed was true opened up entirely on OpenAI’s schedule. That gap is why we run the check as a job rather than a one-off: an agent redoes the same set of buyer-question measurements every cycle, because a reading taken in July tells you what was true in July and nothing about what is true now.
The Reddit case is the clean, publicly documented example because a citation tracker happened to be watching the right thing at the right moment. Most source-selection changes are not reported anywhere, because most affected pages are not a platform the size of Reddit. A business’s own citations can move by the same mechanism, for the same reason, with nobody writing about it.
More detail on how quickly this moves is in How Fast Does ChatGPT Start Recommending You?
Common questions
Did ChatGPT stop reading Reddit, or just stop citing it? The reporting is specifically about citation share, meaning how often a Reddit page is named as a source in the visible answer. It is a measurement of what gets shown, not a claim about what gets read.
Is this specific to Reddit, or could it happen to any site? The mechanism, a change to how sources are selected and queried, applies to every domain equally. Reddit is the documented case because its scale made the drop visible to a tracker; a smaller site’s equivalent shift would look the same to that site and be invisible to everyone else.
How would a business know if this happened to them? Only by measuring again. There is no dashboard or notification from the AI platforms when their source-selection behaviour changes, so the only way to know the current answer is to ask the same questions again and check.
Courage Horizon sets up AI agents for UK small businesses: software that does one job, such as chasing unpaid invoices, and asks you before it sends anything. Being found when customers ask ChatGPT is part of that work. Get a free report on your own business.