OpenAI's ad business hit a $1 billion annualized revenue run rate by the end of August 2026, per Reuters. That's a ten-fold jump from the $100 million pace reported in March. Money is flowing in. The controls advertisers need to spend it responsibly? Still catching up.

The "Conversation Safety" Problem Has No Precedent

Brand safety in search and social is a placement problem: which page, which feed, which video. ChatGPT ads introduce something different. Advertisers and brand-safety analysts are framing it as a "conversation safety" challenge, because ads can appear adjacent to sensitive prompts and to AI-generated content that may itself be wrong. A 2026 study found that 71.9% of U.S. ChatGPT answers carried at least one ad, with 74.6% of commercial questions and 69.5% of informational questions receiving a placement. Broad coverage, narrow visibility into what surrounds each impression.

OpenAI says ads are clearly labeled, separated from answers, don't influence responses, and are excluded near sensitive topics like health, mental health, and politics. Ads appear for Free and Go users; Plus, Pro, Business, Enterprise, and Education tiers remain ad-free. Those guardrails sound reasonable on paper. The gap is in verification: advertisers can't independently confirm what prompts triggered their placements or what the surrounding conversation looked like.

What Advertisers Want (and Don't Yet Have)

Prompt-matching transparency. Advertisers want to see which prompts triggered their ads. Ad selection is primarily contextual (not deep personal profiling), which reduces some privacy concerns but raises a different one: without prompt-level reporting, there's no way to audit suitability after the fact. Brand safety is increasingly a workflow with pre-campaign screening, real-time monitoring, and incident metrics like violation rate and time to resolution. You can't run that workflow on a channel that won't show you where your ads appeared.

Granular exclusion controls. Analysts argue the current tooling lacks the exclusion targeting and placement transparency advertisers expect from mature search or social platforms. OpenAI's self-serve Ads Manager (beta) offers CPC bidding and measurement tools, but the exclusion capabilities don't yet match what most media buyers consider table stakes. Experts in sensitive categories (medical, legal, financial) flag elevated reputational and compliance exposure from adjacency to those prompt types.

Incrementality measurement. OpenAI's current reporting includes impressions, clicks, spend, CTR, average CPC, average CPM, and conversions. That's a standard vanity stack. For teams accustomed to multi-touch attribution and holdout-based lift testing, the gap is obvious: you can't prove a ChatGPT placement generated pipeline that wouldn't have existed otherwise. Without the ability to run holdouts or compare exposed vs. unexposed cohorts at the prompt level, incrementality remains a guess.

How to Run a Conservative Pilot Without Getting Burned

The guidance from brand-safety experts converges on one theme: start tight. Define suitability rules before you spend. Build a sensitive-topic exclusion list specific to your category. Substantiate every claim in your ad copy (conversational context makes unsupported claims more visible and more damaging).

Concrete steps for B2B teams. First, map the measurement gap: reconcile OpenAI's reporting against your internal attribution requirements and document what's missing. You'll need that gap analysis to set expectations with leadership. Second, design a holdout before you launch. If the platform won't run one for you, build a synthetic holdout by restricting spend to specific time windows and comparing pipeline outcomes. Imperfect, but better than treating platform conversions as proof. Third, assign an owner for prompt-adjacency QA. Someone needs to review placement context on a regular cadence, even if that means manually checking a sample of impressions.

Programmatic display brand-safety violation rates dropped from 5.8% in 2023 to 3.4% by 2026, according to industry reporting. That improvement took years of tooling, verification vendors, and advertiser pressure. ChatGPT ads are at the beginning of that curve.

The Bet You're Making Either Way

OpenAI expanded from a U.S. test in January 2026 to 31 European markets, the UK, Mexico, Brazil, Japan, and South Korea by late August. The scale is real. Some advertisers will accept immature controls in exchange for learning on a channel where buyers increasingly start their research.

That's a legitimate trade-off, as long as you name it. The risk isn't that ChatGPT ads don't work. The risk is that you can't yet prove they work, and you can't yet verify they're safe. Two different problems, two different fixes, neither fully available.

OpenAI went from zero to $1 billion in ad revenue in roughly 200 days. The measurement and safety infrastructure that would justify that spend? Still somewhere around day one.