ChatGPT is flagging 'inauthentic mentions' in ranked lists
A classifier note stripped a citation from a first-place listicle entry, and it lines up with a citation drop measured three weeks earlier.
What happened
On 5 September 2026, Metehan Yeşilyurt posted a screenshot of ChatGPT appending a note about "inauthentic mentions" to a list-format answer. He says he lost the citation despite ranking himself first in the list. He reports the pattern is rare so far, alongside A/B testing activity he attributes to GPT-5.6 and an Astra 6 variant, plus classifier updates visible in Statsig. The evidence is one screenshot from one operator.
What we think is going on
Three weeks earlier, Tomek Rudzki measured listicle citations roughly halving after GPT-5.6 shipped. That was a share-of-voice reading across a large sample. Yeşilyurt's screenshot is a single case, but it shows a mechanism that could produce exactly that drop. A classifier is reading ranked lists, deciding some entries look self-promotional, and suppressing the citation while the text still renders.
We think these are two views of the same shift. Rudzki saw the outcome in aggregate. Yeşilyurt caught one enforcement event. Neither on its own is proof. Together they suggest the model is not derating the listicle format, it is derating a subset of listicles where the author is also the top-ranked entity.
The practitioner implication is uncomfortable. A list you wrote that ranks you first is the exact artefact the classifier appears trained to flag. Softening the ranking language without changing the underlying piece is the move most likely to get caught, because the classifier is looking at the structure, not the adjectives.
What to do about it
Before changing anything, verify the finding on your own surface. Pull the last 90 days of ChatGPT citations for any listicle you own where you appear in the ranked positions. Compare citation rate before and after roughly 15 August, when Rudzki's drop began. If your own numbers moved, you have a local signal worth acting on.
For pieces that rank your own product or client alongside competitors, the honest fix is to stop publishing them in that form. A ranked list authored by an interested party is a weak source regardless of what any classifier does this month. Write the comparison as analysis a reader would pick without you in it, and let the ranking fall out of the evidence.
Do not rewrite existing listicles to sound more neutral while keeping the same self-first order. That is the pattern the classifier note describes. If a piece is genuinely the best answer to the query, restructure it around the reader's decision rather than a ranked table with you at position one.
What we're watching
Whether the "inauthentic mentions" note spreads beyond self-promotional lists is the read. If it stays narrow, Rudzki's halving still has a second driver to be found. If it appears on third-party listicles too, the format is being repriced inside the model and the fix has to be structural. We want a second operator to reproduce the screenshot on a query where they have no commercial interest.