Signal
Audit position and ordering effects per engine
Where a product lands in the list an agent reads changes how often it gets inspected, by an amount that depends on the layout, the model generation and how much the model reasons, so you record your rendered position per engine rather than assuming it away.
The effect is well established and its size is not. Two controlled shopping studies disagree by an order of magnitude, and the difference tracks layout and model generation rather than either study being wrong.
Per-platform status
Where a platform documents relevant behavior, its status and source are shown. Undocumented platforms show "unknown" - honest unknowns, never invented.
| Platform | Status | Notes | Source |
|---|---|---|---|
| ChatGPT | Reported | A third-party measurement of the live surface reports that roughly 99.9% of feed-sourced product citations appeared in the first carousel position. That describes how the carousel is sourced and ordered, not evidence that position causes the choice. | Source (opens in new tab) · verified 2026-06-24 |
| Gemini / Google AI Mode | Unknown | Not documented by this platform. | - |
| Perplexity | Unknown | Not documented by this platform. | - |
| Copilot | Unknown | Not documented by this platform. | - |
| Claude | Unknown | Not documented by this platform. | - |
Evidence and sources
This signal is typed reported (Medium confidence).
- ACES (arXiv:2508.02630) (opens in new tab) · v3, 2025-12-17
- Does Rank Still Matter? Position Bias When AI Agents Shop on Our Behalf (opens in new tab) · v3, 2026-09-08
- Lost in the Middle: How Language Models Use Long Contexts (opens in new tab) · v3, 2023-11-20
- Position Bias Undermines Preference Consistency in Listwise LLM-Based Reranking (opens in new tab) · 2026-08-04
- Evaluating Position Bias in Large Language Model Recommendations (opens in new tab) · 2025-08-04
- ChatGPT Shopping deep dive (Profound) (opens in new tab) · 2026-06-24
How to act on this
Record four things beside every win or no-win: your rendered position, how many products the surface showed, the engine, and the model version with its reasoning setting if the engine exposes one. Position is a confound you control for, not one you engineer away.
Related checklist items
Notes
The two shopping studies disagree, and the disagreement is the useful part. ACES reports strong position bias in a two-dimensional grid, varying by provider and model version and persisting even in text-only interfaces. A 2026 study of a one-dimensional list of 100 hotel listings reports a weak, non-monotonic effect whose minimum sits in the middle of the page rather than at the bottom, and concludes that the attributes shown matter more than the placement. Different layouts, different model generations. The mechanism behind the middle-minimum is older than either: language models recall information best at the beginning and end of a long context and worst in the middle. Two further studies of LLM recommenders find that input order shifts scores and rankings, and that evening out exposure is not enough to make a ranking function consistent, which is why the honest advice is to measure and control for position rather than try to neutralize it. Every one of these is a controlled simulation, not a record of real sales.