OpenAI Ads has no search-terms report: you can’t see which prompt triggered your ad

The question with no answer

A running-shoes store runs ads inside ChatGPT and wants the obvious thing: which shopper questions surfaced its card. On Google Ads that's the search-terms report. Ask the OpenAI Ads insights API to break performance down by prompt and it replies with the only dimensions it knows:

segments[]=prompt  →  400: expected one of: "product", "country", "device"

Product, country, device. That's the entire menu. There is no prompt field, no query field, no conversation field — and no search-terms endpoint to fall back to.

What OpenAI Ads exposes instead

The insights API returns performance — impressions, clicks, spend, conversions, ROAS — across the campaign, ad group, and ad hierarchy, sliced by at most one of product, country, or device. That's the whole reporting surface. The one thing it never returns is the input: the words someone typed into ChatGPT before your ad appeared.

Targeting works by description, not by term. An ad group carries a context_hints field — a plain-English account of the conversations you want to reach, like "shoppers comparing trail-running shoes for wet terrain." You don't bid on keywords; you describe intent and the model decides when your ad fits. So the keyword, the match type, and the query report all simply don't exist here.

Why there's nothing to see

Google at least has a query concept. Even after it hides its rare search terms — 68% of one real campaign's spend sat in an unexpandable "Other search terms" row — the search still exists as a discrete, billable event, and negative keywords act on it at matching time. OpenAI Ads has no query event to hide. There is no search term because there was no search — there was a conversation, and you don't get to read it.

That is a deeper opacity than Google's. Google withholds the rare tail of queries; OpenAI withholds the input entirely, because the match happens inside the model rather than in a keyword auction.

What you can still do

The trigger is invisible, so work the two ends you can see: the description going in, and the outcomes coming out.

1. Treat context_hints as the only steering wheel. It is the sole lever over what your ads match. Write it precisely, change it deliberately, and keep the versions, so a shift in results ties to a specific edit instead of a guess.
2. Split ad groups to A/B the descriptions. You can't read the prompts, but you can run two context_hints variants and let conversions say which framing the model matched better. Inference stands in for the search-terms report.
3. Instrument conversions before you scale. Wire up the pixel or Conversions API so every campaign reports conversions and ROAS. When the input is a black box, cost per conversion is the only honest read on whether spend is working.
4. Lean on the three real segments. Product, country, and device are the breakdowns that exist. Use them for the slicing the prompt dimension can't give you.

How the check actually ran

The evidence above is the OpenAI Ads API answering for itself: ask it for a prompt breakdown, and it enumerates the three segments it supports. On an account that runs on Adjar, the parts you can control are plain-text config — the context_hints you steer with live in version control, so an A/B is a reviewable diff and a rollback is one command:

[[campaigns.ad_groups]]
name = "Trail running — wet terrain"
max_bid_usd = 1.20
context_hints = [
  "Shoppers comparing trail-running shoes for wet, muddy terrain; grip and waterproofing questions.",
]

The report Adjar pulls back carries conversions and ROAS per campaign, not a query column — because the platform has no query to give. You measure the outcome, tune the description, and repeat. That is the loop OpenAI Ads leaves you.