The six-day citation collapse that shows why one AI engine is not a GEO strategy

Between 8 and 14 August 2026, one source's share of ChatGPT Search citations fell 86% in six days, from an average of 3.83% down to about 0.5%, according to Promptwatch data reported by Search Engine Land and Axios on 20 August 2026. The source was Reddit. Nothing about Reddit's content, structure or crawlability changed in that window, which is the part worth sitting with.

We track this kind of shift because our own work tells clients to earn AI citations. This one is a useful stress test of that advice, not a story about Reddit.

What actually happened between 8 and 14 August

According to Promptwatch tracking data, Reddit held a steady average of 3.83% of ChatGPT Search citations from 18 July through 7 August 2026. On 8 August, the same day Promptwatch says ChatGPT Search changed its query fan-out behaviour, that share began sliding into the mid-2% range. According to Promptwatch, by 14 August it had fallen below 1%, settling around 0.5%, an 86% drop from the July to early August average.

Why it happened is still not confirmed

Promptwatch was explicit that its data shows when the shift occurred, not why, and called the finding provisional while it keeps monitoring. Its working theory, cited in the same coverage, is that ChatGPT Search began using "site:" operators more aggressively around 8 August, targeting specific domains directly rather than running broader web queries that had previously surfaced Reddit threads. OpenAI, for its part, said it had not changed how it selects sources.

That gap between a plausible mechanism and a confirmed one is normal for this category. Nobody outside the platform can see the ranking logic, so every external measurement is inference from outputs. Treat the fan-out explanation as the leading theory, not an established fact.

Our reading: citation share is instrumentation, not a KPI

Most of the coverage we read framed this as a diversification lesson: do not depend on one platform, spread your presence across engines. That is true and also too comfortable. It leaves the underlying assumption standing, that citation share is a number worth building a strategy, a dashboard or a client report around in the first place.

Our view is that this event argues against that assumption directly. A source did nothing differently and still lost most of its visibility on one engine in six days, for a reason its own publisher cannot confirm and the platform denies causing. According to this Promptwatch reading, if a number can move 86% for reasons nobody outside the platform can verify, it is not a metric with the kind of stability a target or a bonus structure needs. It is a smoke detector: useful for noticing when something changed, useless as the thing you are trying to optimise.

We have said before that crawlability, structure, schema, freshness and original data are the levers that earn citations, because they are levers a site owner actually controls. Citation share is the read-out those levers produce, filtered through a ranking system nobody outside the platform can see. Confusing the read-out with the lever is the mistake this event exposes.

What we tell clients to do with this

Keep measuring citation share, the same way you would keep a smoke detector even though it does not put out fires. We have written about why AI referral traffic is worth tracking at all despite its small volume, and the same logic applies here: it is a signal, not the goal.

Build the assets a platform cannot zero out overnight. That means the structural work in our own GEO audit checklist, making sure your most citable pages are not invisible to the crawlers that feed these engines in the first place, and treating any single engine's current behaviour, ChatGPT Search's fan-out logic included, as something that can change without notice and without your content changing at all.

And be honest in how you report this internally. A citation-share chart that dropped 86% for reasons the source cannot confirm is not evidence the underlying GEO work failed. Reporting it as though it is, in either direction, teaches a board or a client the wrong lesson about what the work is actually for.

Three questions before a citation number goes in a deck

• Can you name the platform, the date range and the exact percentage move, rather than a rounded impression of "citations are up" or "citations are down"? If not, the number is not solid enough to act on yet.
• Has anything about the underlying page, its structure, schema, freshness or the crawler access it grants, actually changed in that window? If the content is untouched and the number moved anyway, the platform moved, not your work.
• Would the number survive being wrong about the cause? According to Promptwatch, Reddit's share fell 86% and nobody, including the platform that produced the fall, can confirm why yet. A metric you cannot explain is not one to attach a target or a bonus to.

None of that means ignore the number. It means read it the way you would read a smoke detector going off: investigate, do not assume the reading itself tells you what to fix.

The six-day drop is six days old. What it should change is not which engine you chase this month, but whether citation share sits in your strategy deck at all, or in the instrumentation panel underneath it. Our AI Search Visibility Field Manual sets out the levers we think are worth building a strategy on instead.