OpenAI reveals how ChatGPT selects and serves ads

OpenAI has just told advertisers something it has spent two years declining to tell publishers: how the model decides what to surface, and on what timetable. Per Search Engine Journal, OpenAI has published details of a custom LLM orchestrator that reads multi-turn ChatGPT conversations, nominates candidate ads for an auction, and discards any placement that fails to resolve before a response deadline. This is not a minor technical disclosure. It is the first real blueprint of a live, latency-governed relevance system inside a conversational model, and it tells brands more about how ChatGPT ranks things than any SEO guidance OpenAI has issued to date.

The deadline is the story

Strip away the auction mechanics and one detail stands out: ads get dropped if they miss a response deadline. ChatGPT is not running a leisurely retrieval-and-rank pipeline the way a search engine can. It is composing an answer in real time, and anything, ad or citation, that cannot be resolved fast enough gets excluded regardless of relevance. That is a fundamentally different selection logic from Google's auction stack, where latency is a ranking factor among many. Here it looks closer to a hard cutoff: fast and plausible beats slow and perfect.

That matters beyond advertising because the orchestrator OpenAI describes is almost certainly adjacent to, or shares infrastructure with, the retrieval system that decides which web sources get cited in an organic answer. OpenAI has never published equivalent detail on citation selection. But the ad disclosure confirms what third-party audits have long inferred: ChatGPT's answer construction is a race against a clock, conducted turn by turn, with sources and placements nominated and then pruned based on what resolves in time. A brand's content being accurate, authoritative, and well-structured is necessary but not sufficient. It also has to be fast to retrieve and cheap to parse, or the orchestrator will simply move on without it.

What multi-turn nomination means for institutional content

The more consequential detail for the audiences that matter to financial institutions, multilaterals, and large industrial groups is the multi-turn piece. OpenAI says the orchestrator interprets the conversation, not just the final prompt, when nominating candidates. That is a meaningfully different retrieval target than the single-query model most brands still optimise for. A pension regulator asking ChatGPT about capital buffers, then following up on Basel implementation timelines, then asking which institutions have published position papers, is building a conversational context that the orchestrator reads cumulatively. Content built for one-shot queries, the classic FAQ page or glossary entry, is poorly suited to that. Content built as a connected sequence, the kind that answers a follow-up before it's asked, is better suited to nomination at turn three or four, which is often where institutional and B2B queries actually resolve.

This is the part advertisers have been told explicitly and publishers have had to reverse-engineer. Media buyers now know that nomination happens per-turn and that a placement can win visibility in turn two after losing it in turn one. Nobody has told publishers or brands that the same cumulative logic applies to citations. It almost certainly does, because OpenAI is not plausibly running two separate conversational-context engines side by side, one for ads and one for organic answers. The ad disclosure is the closest thing to a technical map of ChatGPT's internals that any outside party, including SEJ, has managed to extract.

The asymmetry that should worry comms and policy teams

OpenAI is being transparent with advertisers because advertisers pay for transparency: auction mechanics must be disclosed to some degree or buyers walk. No equivalent commercial pressure exists for citation transparency, which is why organisations spending heavily on authoritative content, white papers, position statements, research summaries, the staple output of a CGAP or an ISO, have no comparable account of how that material gets surfaced or dropped. The ad stack disclosure is a signal of what that system probably looks like underneath, but it is not a substitute for OpenAI publishing it directly.

The practical implication is not to buy ChatGPT ads to reverse-engineer the ranking logic, though some agencies will try exactly that. It is to treat conversational persistence as a design requirement for owned content: structure material so it survives and strengthens across a multi-turn exchange, not just a single query, and assume that whatever gets measured for speed in the ad stack is being measured for speed in citation retrieval too. The institutions that wait for OpenAI to publish that second blueprint will be optimising against yesterday's model by the time it arrives.