OpenAI's ChatGPT Work agent reshapes how sources get used

Months after rivals began shipping agentic features, OpenAI has moved decisively. The OpenAI blog announced ChatGPT Work, a mode that converts the familiar chatbot into a long-running agent: one that can open files, operate third-party applications, hold a task in memory across an extended session, and return a finished output rather than a conversational draft.

The gap between "chatbot" and "agent" sounds technical. The consequences for how brands appear in AI-mediated work are not.

From retrieval to execution

A standard ChatGPT session is retrievive. A user asks; the model answers, citing or synthesising sources at the moment of response. Brand visibility in that model depends on whether the model's training data and retrieval pipeline surface a given brand's content at the point of query.

ChatGPT Work changes the unit of interaction from a query to a project. The agent takes a goal ("produce a competitive analysis of cement pricing in Southeast Asia," or "draft a grant narrative aligned with UNDRR's Sendai Framework targets") and works toward it across multiple steps, tool calls, and file reads. The citations and sources that feed that process are selected not by a single retrieval event but by a series of agent decisions made over time, often without a human in the loop to redirect them.

That shifts the question for brand visibility from "does my content appear in this answer?" to "does my content get selected as a working source for this task?"

The distinction matters enormously for institutions whose authority is tied to long-form, structured material: a multilateral's technical reports, a financial services firm's research papers, an industrial group's standards documentation. These assets are poorly served by a retrieval model optimised for short answers. They are precisely the kind of material an agent running a multi-hour task might read, integrate, and build upon, if the agent can find and trust them.

The new selection criteria

In a single-turn retrieval environment, the factors that predict citation are reasonably well understood: domain authority, topical match, freshness, and the presence of structured, quotable claims. In an agentic environment, those factors still apply, but new ones emerge.

Agents operating autonomously tend to prefer sources they can parse reliably. Clean document structure, consistent metadata, machine-readable formats, and explicit provenance signals (who published this, when, under what authority) become functional requirements, not nice-to-haves. A PDF buried behind a registration wall, or a report whose authorship is opaque, is a source an agent may simply skip.

Institutions in the UN system and multilateral development banks produce some of the world's most authoritative data on poverty, climate risk, and labour markets. Much of it is formatted for human reading and discovery. If ChatGPT Work begins functioning as a research partner for the policy analysts and senior executives those institutions want to influence, the formatting and accessibility of those documents becomes a direct variable in whether the institution's analysis shapes the output or is invisible to it.

For financial services firms, the calculus is similar but the stakes are commercial. If a corporate treasurer's team uses an agentic assistant to model counterparty risk across a portfolio, the research sources the agent selects will influence conclusions. Being absent from agentic retrieval is not a neutral outcome.

Persistence changes the authority signal

One underappreciated feature of ChatGPT Work is session persistence. The agent can stay with a task for hours. That means it may return to a source, cross-reference it against others, or weight it more heavily because it proved reliable in an earlier step. Reliability across a session, not just relevance at a single moment, becomes an authority signal.

This rewards brands and institutions that publish consistent, internally coherent bodies of work. A series of reports that contradict each other, or that use inconsistent terminology across years, will perform poorly in an environment where an agent is actively checking its sources against each other. The ISO, IEEE, and similar standards bodies, which TCE serves, are structurally well positioned here: their outputs are by design consistent and cross-referenced. Most corporate content functions are not.

OpenAI has not disclosed which retrieval mechanisms or knowledge sources ChatGPT Work draws on, nor how it weighs proprietary integrations against open web content. That uncertainty is itself a signal. Brands that have concentrated their authority-building in closed environments, private intranets, gated portals, paywalled databases, face a visibility risk that was manageable when the primary interface was a human analyst. It is less manageable when the analyst is an agent that takes the path of least retrieval resistance.

The organisations that will be cited in AI-generated finished work are the ones whose knowledge is structured, accessible, and internally consistent. That is not a content strategy; it is an infrastructure question, and most content teams are not yet treating it as one.