Ben Kahan's team ran the numbers on ChatGPT advertising before most agencies had even logged into the Ads Manager. The verdict from his PPC Hero analysis is blunt: the channel works, but not the way the pitch deck suggests.
The scale is undeniable. ChatGPT now reaches 853 million monthly active users, with roughly 63% of them using the platform for research and product comparison. That's consideration-stage traffic at a volume that didn't exist eighteen months ago. The question Kahan's testing answers isn't whether to be there, but what the channel actually does once you spend.
The Targeting Reality
OpenAI's ad unit sits below the ChatGPT response in a clearly marked "Sponsored" box: one image, a headline, a short description, and a click-through. The format is clean. The targeting is not what most buyers assume.
OpenAI's documentation confirms the system is contextual only, matching ads to the user's prompt rather than to demographic overlays, audience segments, or geographic precision. There's no behavioral retargeting, no lookalike modeling, no CRM list matching. Relevance to the prompt is the sole criterion for delivery.
This creates a different optimization problem than most paid media teams are used to solving. You're not bidding on audiences. You're bidding on conversational intent, and the only lever you control is how tightly your creative maps to the prompts you want to appear against.
The Pricing Math
Kahan's testing found CPMs landing around $40, which sits in the middle of the range industry benchmarks now show: $25 for niche, uncontested prompts, climbing toward $60 for competitive categories. The pattern is consistent across advertisers: precision pays. The more specific your messaging to the prompt context, the better your efficiency.
The CPC model, which OpenAI enabled in April, recommends starting bids of $3 to $5 per click. Kahan's team steers away from it. The methodology behind OpenAI's click optimization is too opaque to trust, and bidding on CPM while letting clicks come organically has delivered a better effective cost per click in their testing.
This is a channel where the auction mechanics are still being built. The relevancy model OpenAI runs on top of the proto-auction produces two prices depending on competition, and the spread between them is wide enough to make or break a test budget.
The Measurement Gap
Here's where the CFO conversation gets uncomfortable. Ads Manager reports seven metrics: impressions, clicks, spend, CTR, average CPC, average CPM, and one rolled-up Conversions number. That's it. No query-level data, no demographic breakdowns, no placement reporting. This isn't a beta limitation; it's a structural, privacy-by-design boundary.
The Conversions API and pixel exist, but they're basic. There's no conversion optimization yet, meaning you can't ask the platform to chase outcomes. What you can do is measure the knock-on effects of those clicks: do people arriving from ChatGPT show higher intent than they do from native, display, video, social, or search, judged on what they do once they land on your site?
Kahan's early read is encouraging. Click-through rates are coming in comparable to native ads, which is a strong signal for a format this new. But the practical move is to build the evaluation framework before you spend, deciding up front how you'll compare ChatGPT intent against site behavior from your other channels.
Three Routes In, Three Price Tags
Kahan outlines three buying paths. The first is direct through OpenAI's Ads Manager, which opened to all US businesses in May with no minimum spend. The second is through Microsoft Advertising, which manages some of the inventory. The third is through programmatic DSPs that have integrated ChatGPT supply.
Each route has different fee structures, different levels of control, and different reporting granularity. The self-serve path is the cheapest but offers the least support. The managed paths add margin but bring optimization expertise that matters when the platform's own signals are thin.
For B2B teams, the DSP route may be the most practical. Brainlabs' programmatic practice already runs intent-based campaigns across multiple platforms, and integrating ChatGPT supply into an existing measurement framework is cleaner than standing up a new attribution silo.
The Incrementality Question
The core issue, as one measurement specialist noted on LinkedIn, is the conflation of intent with incrementality. OpenAI's pitch is that ChatGPT users are "super intentional," arriving mid-decision and already considering their options. That's true. But high intent drives both the perceived success of the ad and the conversion itself. Naive attribution will exaggerate the ad's influence.
The fix is the same one that works for any new channel: geo-matched holdout tests, budget-on versus budget-off comparisons, and a clear hypothesis about what lift you expect to see. Incrementality testing frameworks designed for ChatGPT are already emerging, and they're the only way to answer the question that matters: did this spend create conversions that wouldn't have happened otherwise?
The Pilot Design
Kahan's recommendation is to start with a 2-3 week test, CPM bidding, creative tightly mapped to specific prompt categories, and a pre-built measurement framework that compares ChatGPT arrivals against your other channels on site behavior, not just conversion rate.
The risks are clear: opaque auction mechanics, thin measurement, and a relevancy model you can't inspect. The mitigations are equally clear: small initial budgets, tight creative-to-prompt alignment, and incrementality testing before you scale.
ChatGPT advertising is real. The audience is there. The intent signal is strong. But the channel is young enough that the operators who build measurement discipline now will be the ones who can defend the spend when the CFO asks what it actually did.
