The Maze: Writing another “best products” list may be a weaker bet for winning ChatGPT citations. In Peec AI's study of 180 million sources across the same one million tracked prompts, listicles lost 50.5% of their citation share after the GPT-5.6 rollout; comparison pages lost 32.1%. Peec AI measures brands' visibility in AI search. Its paired-week sample shows a sharp change in the kinds of pages cited, but not an official penalty against those formats. Nor does losing half your share mean losing half your customers.
Listicles lost half their slice, not half the entire pie. Their citation share moved from 15.77% before the rollout to 7.80% after it: a 7.97 percentage-point decline, expressed as a 50.5% relative reduction. That distinction changes the headline. A percentage-point loss describes movement within the whole; a relative loss compares the category with its own starting position. Neither tells a retailer how many visits or purchases disappeared. Listicles still earned citations in the second week, so declaring the format dead outruns the evidence.
Comparison pages weakened too, but by less. Pages built around one product or company versus another fell from 9.08% to 6.17% of citations, a 2.91-point decrease and a 32.1% relative drop. Two familiar routes into product discovery therefore lost weight together. Yet this is a comparison of page formats, not a ranking of content quality or commercial returns. It does not distinguish a useful hands-on product comparison from a thin page written to attract an AI mention. Cutting both indiscriminately would turn a measurement into a content policy it cannot support.
A larger search pool changes the denominator. The same analysis records 12.48 retrieved sources per chat before the change and 25.85 afterward, while cited sources increased separately by 25%. Retrieved pages are candidates the system fetches; citations are links included in its answer. Background searches, or fan-outs, also increased. More searching can dilute a format's share without a matching fall in its raw count. The reported totals and shares do not justify combining different aggregation methods into a precise estimate of lost listicle citations. Measure the counts directly.
The useful response is to diversify, then measure. Searches containing “top” lost 75% of their share and 36.5% in average number per chat; “best” grew in absolute terms while losing share. That is a warning against treating any one search phrase or page template as a durable distribution channel. Domain-restricted searches also became more common, but an anti-spam explanation remains a hypothesis. The study does not disclose country coverage, classifier accuracy or buyer outcomes. A week on either side of a rollout cannot establish a permanent rule for every market.
Why it matters: Retailers and publishers should audit dependence on listicles and comparisons while keeping useful product evidence easy to find on their own sites. Track fetched pages, actual citations, relevant visits and purchases separately. The practical risk is a content budget tied to yesterday's citation pattern. The opportunity is a broader set of useful pages whose value survives a change in the model. This sample supports testing that approach; it does not promise first-party winners or prove an intentional crackdown.
