What a Real Model Shift Looks Like in the Data

In mid-August 2026, several outlets reported that ChatGPT had sharply reduced how often it cites social and community sources. Search Engine Land, citing analyses from Petra Labs and PromptWatch, and a follow-on piece in Inc., described Reddit's share of ChatGPT citations falling roughly 86% between August 14 and 17, with YouTube and Instagram down as well. The reports framed it as ChatGPT-specific: Claude, Gemini and Perplexity were said to be steadier over the same days.

We track citation source type by provider every day, so we checked it against our own data before repeating it. Here is what we found, stated the way it came out rather than the way it would have been more convenient.

What we checked

Monroya logs every source an AI model cites in a tracked answer and classifies it by type (community, review, media, reference, vendor and so on), by provider, and by buyer-journey stage. That log exists precisely so a shift like this is visible without anyone having to guess.

We compared two windows in the same dataset:

• Before: August 4 to 13, 2026
• After: August 14 to 25, 2026

Same accounts, same prompt set, same classification rules. We looked at the community source type overall and at reddit.com specifically, split by provider.

What the data actually shows

Reddit specifically, inside ChatGPT: 1.5% of citations before, 0.7% after.

So the direction matches the reports. ChatGPT's community share did roughly halve. But three things stop us from calling it confirmed:

The sample is thin. Those percentages rest on 21 community citations before and 21 after, from six and then five distinct accounts. Our research reporting requires a minimum citation count and a minimum number of contributing accounts before a slice gets labeled as a finding rather than a signal. This slice does not clear that bar. A handful of prompts landing differently would move the share by a full percentage point.

It is not ChatGPT-specific in our data. Gemini's community share fell further over the same window, from 4.0% to 2.1%, on a larger base. If a platform-specific policy change were the whole story, Gemini should not be moving in the same direction by a larger margin. Claude went the other way, up from 2.1% to 2.7%.

Our Perplexity collection has a gap. There are no Perplexity citation rows in the before window, so we cannot compare it at all. That is a limitation in our own pipeline, not a finding about Perplexity.

Our honest verdict: directional, not confirmed. Something moved in the right direction for ChatGPT. Our sample in this window is too thin to independently confirm an 86% collapse, and the parallel move in Gemini argues against the "one model changed" framing. We logged it in our model-behavior log with exactly that caveat, because a non-confirmation is a real data point too.

Why this matters even without a confirmation

The interesting part is not whether the specific number holds. It is that a company doing nothing wrong, changing nothing on its site, could have seen its ChatGPT visibility move in that window purely because the model changed what it trusts.

That is the whole argument against treating AI visibility as a one-time audit. An audit run on August 10 and never repeated would describe a world that stopped existing four days later. The only way to tell "our content got worse" apart from "the model changed its sourcing" is to have been measuring continuously on both sides of the change.

It also shows why a single blended visibility score hides more than it reveals. In this window, one provider's community share halved, another's fell by more, and a third's rose. Averaged together, that reads as roughly flat. Nothing happened, apparently. Plenty happened.

What to actually do about it

If your content strategy leans on community visibility, and Reddit in particular, three practical steps:

1. Look at your own per-provider breakdown, not an aggregate. Your mix is not the market's mix. A drop concentrated in one model is a different problem than a drop across all of them.
2. Compare against a real before window. A share number with nothing to compare it to cannot tell you whether anything changed.
3. Do not rip out a channel on one report. Community sources still carry weight elsewhere. In our own data Claude cited community sources more after August 14, not less.

This does not invalidate Reddit as a channel. It argues for knowing which model your Reddit visibility is actually earning citations in, and watching that number over time instead of assuming it holds.

Methodology

Source: Monroya's citation log, which records every cited source returned in a tracked answer, classified by source type, provider and buyer-journey stage at the moment of the scan. Windows compared: August 4 to 13, 2026 (10 collection days) and August 14 to 25, 2026 (12 collection days). Providers: ChatGPT, Claude, Gemini, Perplexity. Community share is community-type citations divided by all citations for that provider in that window. Confidence labeling follows the same minimum-citation and distinct-account gating used in our published citation research: slices below the threshold are reported as directional signals, not findings. The ChatGPT and Gemini community slices in this analysis fall below that threshold, and are labeled accordingly. Perplexity is excluded from the before/after comparison due to missing collection in the earlier window. External claims are summarized from public reporting and are not reproduced here.

Related reading

• Reddit is Your New SEO: How Community Content Drives B2B AI Visibility
• How LLMs Actually Weigh Reddit Content (And How to Know If Your Comment Worked)
• Why Reddit and G2 Are Winning the AI Share of Voice Battle
• What Sources Is AI Actually Citing When It Answers Questions About My Industry?

Want to see your own per-provider citation mix, before and after a shift like this? Start a free trial or run a free visibility check.