Intent orchestration, consumer world models, and agentic recommendation for multi-turn LLM experiences.
Generative engine optimization is information retrieval's next chapter, not a departure from it. Where a user once received a ranked list, they now receive a single synthesized, cited, conversational answer — and the questions of what gets retrieved, what gets cited and what gets recommended have all moved inside the conversation.
This workshop is about the part of that shift where information seeking turns into a transaction. How should retrieval and synthesis work faithfully across turns? How is a user's intent tracked as it changes? When a system stops retrieving and starts recommending, what has happened to its objective? And how would anyone know whether any of it worked?
We are interested in the last question most of all. The commercial claims in this area are currently running well ahead of the evidence, and a workshop is a reasonable place to close some of that gap.
Methods and metrics for generative engine optimization; LLM citation as a relevance signal; rethinking relevance for generated answers; retrieval-augmented generation for multi-turn recommendation.
Intent modelling across turns; feedback-based orchestration; relevance versus reasoning in conversational search; recommendation across e-commerce and offline journeys.
Digital-twin user representations; long-term and cross-session memory architectures; parameter-efficient personalization; persona and consumer simulation.
Agent and tool selection; DAG-based workflow orchestration; standardized lifecycle simulation; what an orchestration policy optimizes across a whole session, and what it costs to run.
User simulation and dataset construction; benchmarks for LLM conversations; customer-experience measurement; multi-turn conversation evaluation, and whether the constructs being measured are measurable at all.
Every submission names one of three categories. All three are reviewed on their own terms; none is a lesser class of contribution.
Prior and concurrent work. Submissions must not have been published in, or accepted to, a peer-reviewed venue. Two deliberate exceptions:
A preprint on arXiv or a similar server does not count as prior publication.
Work currently under review elsewhere may be submitted as a non-archival contribution — indicate this at submission. It is reviewed on the same terms and presented at the workshop, but omitted from the ACM proceedings, so it creates no conflict with the venue you submitted to. A workshop is a place to discuss work in progress; we would rather see it early than not at all.
Negative and null results are explicitly in scope. A well-designed study finding that an intervention did not work, or that a reported effect does not survive a competitive setting or a second domain, is as welcome here as a positive result — and given the state of the published evidence, more useful. Reproductions and failed replications are equally welcome, and a negative result that takes three pages should take three pages.
This was written into the call before any committee member agreed to serve, and reviewers are instructed accordingly.
By submitting to GUIDE 2026 you acknowledge that you and your co-authors are subject to all ACM Publications Policies, including ACM's Publications Policy on Research Involving Human Participants and Subjects. Alleged violations of that policy, or of any ACM Publications Policy, will be investigated by ACM and may result in a full retraction of the paper in addition to other penalties.
That policy is not boilerplate for this workshop. Several of our topics — consumer simulation, satisfaction measurement, user studies of conversational assistants — involve human participants directly, or involve claims about people derived from their behavioural traces. If your submission rests on either, read the policy before you write, not after you are accepted.
Please ensure that you and your co-authors obtain an ORCID iD, which is required to complete the publishing process for an accepted paper. ACM collects ORCID iDs from all published authors to improve author discoverability and attribution.
Further members to be announced. Each listed member reviews a small, named load; the committee is being assembled for reading depth rather than length of list.
This workshop is organized from industry: the organizer runs Guyu AI, a company working in this area. That is stated here rather than discovered later, and it is the reason the programme committee is weighted towards researchers with no commercial stake in these methods working — including several whose own published results are sceptical that they do.
Conflicts of interest with the organizer's company are registered in the submission system, and reviewing assignments exclude them. The negative-results clause above is a commitment made to committee members in writing before they agreed to serve.
Qiankun Zhao — Founder & CEO, Guyu AI (Beijing). PhD, Nanyang Technological University; postdoctoral research at Pennsylvania State University; previously Microsoft Research, AOL, and Telefónica.