ChatGPT Ads Now Has Real Measurement. How to Run a First Test You Can Judge

What OpenAI announced, and what it leaves open

OpenAI's latest ChatGPT Ads update, published October 5, does two things that matter to an ecommerce team. It adds a visual ad format built around product imagery, which starts testing with a first group of US advertisers later in October inside ChatGPT's image generation. And it opens up measurement: a pixel and a Conversions API, named attribution partners including Triple Whale, Northbeam and AppsFlyer, geo-based incrementality tests with Haus and Measured, and a one-day view-through window on top of click attribution.

OpenAI also repeated the line that matters for anyone comparing paid and organic: ads carry labels and have no influence over ChatGPT's answers. Buying ads does not change whether the assistant recommends your product when a shopper asks.

Three caveats before anyone reallocates budget. The visual format is a test, not a launch. The performance figures OpenAI shared are early and partner-reported: a 15.3% lower cost per acquisition than blended paid search for one brand, 93% new visitors for another, as covered by Unite.AI. And view-through conversions are reported as a separate campaign metric. OpenAI's Conversions API documentation says cost per acquisition, bidding and conversion optimization stay click-through based, and view-through availability depends on the account.

Why the first test usually fails

At the events we attended in September, more than one ecommerce leader described the same experience: a few thousand dollars on ChatGPT ads, one order, test over. In every case the problem was not the channel. The feed was accepted with thin descriptions, the products were chosen because they were easy to export rather than because they carried margin, and the only measurement was last click in a platform that could not see the conversation.

With this update the measurement excuse is gone. What is left is the work on your side: which products, what the feed says about them, and how you will know whether the spend paid.

How to run a first test that can actually be judged

1. Pick products a visual ad can sell

Choose fifty to one hundred products with margin to spare, deep stock through the test window, and imagery that shows the product in use. OpenAI describes the new format as images of product inspiration, product usage, or the experience a product makes possible. A product shot on white will be at a disadvantage against a competitor's lifestyle image. If your best sellers only have cutouts, fix that for the test set before you start.

2. Get the feed accepted and keep it servable

ChatGPT product ads run from a product feed that follows OpenAI's product feed specification. The blockers we see most are missing descriptions, descriptions that repeat the title, missing merchant information such as a return policy, and price or availability that disagrees with the product page. Clear them in the system that generates the feed, not by hand in a spreadsheet, so the fixes survive the next refresh. A product that drops out of the feed mid-test is a hole in your data, not a learning.

3. Set up measurement before the first dollar

Install the pixel and send purchases through the Conversions API, which OpenAI calls the more reliable of the two. If you already run Triple Whale or Northbeam, use their integration so ChatGPT Ads lands in the same view as your other channels. Pick the click attribution window that matches your buying cycle, and read view-through as its own line. Tag the landing pages so sessions from the ads do not disappear into Direct in GA4. The setup is the same one we described for measuring AI referral traffic in GA4 and Shopify.

4. Decide in advance what a win looks like

Write down three numbers before launch: the share of orders from new customers, cost per acquisition against your blended paid search, and the organic appearance rate for the same products in ChatGPT's answers, tracked separately. The first two tell you whether the ads work. The third stops you from crediting ads for recommendations that come from your product information. If the budget is large enough, ask your measurement partner for a geo holdout. OpenAI's own incrementality work is early-stage, so treat partner-reported lifts as hypotheses for your category, not benchmarks.

What not to do

Do not run the test on the whole catalog. Do not change the product pages and the ad copy and the feed in the same week, or you will not know which one moved the number. And do not treat an ad impression as a substitute for being recommended. The product information that makes a feed servable for ads is the same information the assistant reads when it answers an unpaid question, which is why we treat the two as one piece of work.

This is the ChatGPT Ads work Paz does for lean ecommerce teams. A Paz expert chooses the products, clears what blocks them from the feed, writes ad titles and descriptions from your product data and reviews per-product impressions with your media buyer, who keeps the budget and the bids. You approve every draft, the platform publishes and tracks. If you want your first test set up so it can be judged, talk to us.