OpenAI Put an Ads App Inside Your Shopify Admin. Install It, Cap the Test, and Refuse the Default Billing.

On September 16, 2026, OpenAI published a post naming Shopify its first ecommerce partner for ChatGPT Ads and HubSpot its first CRM partner. For Shopify merchants the practical output is a free app in the Shopify App Store, developed by OpenAI, that connects or creates a ChatGPT Ads account from inside the admin, syncs your product catalog, installs a pixel, and runs campaigns without you ever opening Ads Manager.

US merchants have it now. International availability in markets where ChatGPT Ads run begins September 23.

The recommendation

Install it and run a bounded test this month, because auction pressure only goes one direction from here and because the international opening on September 23 adds bidders. But do three things before you spend anything.

One, start on cost per click, not on the impression-billed conversion optimization that OpenAI made generally available in the same announcement and now recommends as its performance model.

Two, set your attribution columns to a 7-day click, 0-day view window. The windows became configurable in this release. The default credit model on a brand new channel with no independent measurement will flatter it.

Three, run the permissions past whoever owns privacy at your company before you hit install, because this app reads more customer data than most merchants assume an ads app reads.

Then size the test to answer a question you can actually answer. More on that below.

What the app actually does

From the Shopify App Store listing, reviewed this week: developer OpenAI, price free, release date September 3, 2026, which is thirteen days before OpenAI announced it. It is filed under Ads, works with Checkout and Shopify Admin, and supports English only.

The listed functionality:

• Connects an existing ChatGPT Ads account or creates a new one from Shopify
• Syncs product inventory automatically between the store and ChatGPT
• Builds ads from the product catalog
• Runs campaigns using a Shopify pixel and conversion events
• Reports spend, click-through rate, conversion tracking and cost per acquisition
• Targets by audience segment, custom audience, device, location, platform and product category

Two details worth pausing on.

First, the pixel. A reviewer asked on September 11 whether the app used server-to-server tracking. OpenAI replied the next day that the app automatically installs a Shopify server pixel, which the company equates with a server-to-server Conversions API integration. That is a real measurement advantage over a browser pixel, and it is also a server-side data connection you are switching on by installing an app.

Second, the permissions. Per the listing, the app reads customer names and email addresses, IP addresses, browser and operating system details, browsing behavior and a client ID cookie. It reads the store owner’s contact details, product listings and collections. It can edit marketing events, web pixels and server pixels.

That is not an unusual permission set for an ads app. It is an unusual permission set to grant without reading it, which is what a one-click install from the admin encourages. If you sell into the EU or UK, note that the European Commission designated ChatGPT a very large online search engine on August 31 after OpenAI declared 159.1 million users, and that the international app rollout starts September 23. Get your data processing position settled before the app reaches those markets, not after.

As of September 17 the listing carried three reviews and a 4.0 rating. Two five stars and one one star, from a US retailer on September 11 saying the app did not work. Three reviews is not a signal. It is a reminder that this is day-one software.

Who actually sees your ad

This is the part most coverage skips, and it determines whether the channel is worth your time.

ChatGPT ads serve to logged-in adult users on the Free and ChatGPT Go tiers. Plus, Pro, Business, Enterprise and Education do not show ads.

Read that as a targeting constraint, because that is what it is. The users who pay OpenAI twenty dollars a month or more are the users your ad cannot reach. If your customer skews toward higher income professional buyers, a meaningful share of them are sitting behind a paid tier and are structurally unreachable on this surface. If you sell a considered purchase to a broad consumer audience, the constraint matters much less.

Ask the question in your own numbers before you budget: what fraction of my buyers are plausibly paying for ChatGPT right now, and am I comfortable buying reach that excludes them?

The format, and why your PDP does the work

The ad unit is small. Advertiser name, favicon, a headline of roughly 16 characters, a description of roughly 32, an image and a landing page, appearing as a card at the foot of an answer or, since August, in a carousel. Sensor Tower estimated the multi-product carousel accounted for roughly 23 percent of US desktop ChatGPT ads between August 15 and August 30. That is a vendor panel estimate, not an OpenAI figure.

Sixteen characters of headline means your creative lever is close to zero. What you are buying is a click into your own product page, and your PDP does all of the converting.

There is direct evidence for why the click-out path is the one on offer. OpenAI launched Instant Checkout with Stripe in September 2025, letting shoppers buy from Shopify merchants inside the chat. It shut the product down in March 2026 after Walmart disclosed that conversion rates for products sold inside the chatbot ran three times lower than for purchases that clicked out to its own site. Product feed management moved into Ads Manager in June, supporting up to one million SKUs per advertiser.

So the platform tried in-chat buying, measured it against click-out, and retreated to feeds and ads. Take that as your own planning assumption. Budget your effort into landing page fit and feed hygiene, not into ad copy you have sixteen characters to write.

The math, by AOV

Public CPC figures for this channel are agency and vendor reported rather than published by OpenAI. Reported entry bids run roughly three to five dollars per click, with CPMs reported in the twenty-five to sixty dollar range, on a relevance-weighted second-price auction. Treat those as a planning band, not a rate card, and replace them with your own numbers after week one.

Run break-even at the midpoint, four dollars per click.

Brand A: $90 AOV, 40 percent contribution margin. Contribution per order is $36. At four dollars a click you get nine clicks per order before you are underwater, which requires an 11.1 percent conversion rate. Cold traffic from a chat surface will not do that. At a more believable 2.5 percent conversion rate your CPA is $160 against $36 of contribution, a 4.4x loss on every order.

Brand B: $300 AOV, 45 percent contribution margin. Contribution per order is $135. At four dollars a click you can absorb 33 clicks per order, which is a 2.96 percent conversion rate. At three dollars a click the bar drops to 2.22 percent. At five dollars it rises to 3.70 percent. Those are numbers a good PDP can hit on qualified traffic.

The conclusion is not subtle. At reported CPC levels this is not a Meta prospecting substitute for a low ticket brand. It is a channel for considered purchases with high AOV, or for subscription and replenishment models where you can underwrite a first-order loss against known LTV. If you are in the first group and you want to test anyway, test it as a retargeting and branded-intent channel where conversion rates are three to five times higher, not as cold prospecting.

Size the test to a question you can answer

A $2,000 test at four dollars a click buys you 500 clicks. At a 2.5 percent conversion rate that is roughly twelve orders. Twelve orders will not tell you your true conversion rate or your ROAS. Anyone who reports ROAS off twelve orders is reporting noise.

What 500 clicks does tell you, reliably, is traffic quality. Read these instead:

• Actual CPC against the reported band
• Click-through rate on the card
• Bounce rate and time on page versus your Meta and Google benchmarks
• Add-to-cart rate, which is the earliest honest signal of intent
• Product mix, meaning which SKUs the feed surfaced and whether they were the ones you wanted

If add-to-cart rate on ChatGPT traffic lands within range of your paid social benchmark, scale to a budget that can actually read conversion rate. If it lands at a fraction of it, you have your answer for a few hundred dollars instead of a few thousand.

The three settings, in detail

Billing. The September 16 advertiser email, published by Search Engine Roundtable, confirmed that conversion optimization with impression billing is now generally available and is OpenAI’s recommended performance model. It optimizes for conversions from both clicks and views while charging on impressions. Understand what that combination does on a new channel: you pay for every impression, and the system takes credit for conversions from people who merely saw the ad. View-through credit arrived on August 19 with a fixed one-day window and was explicitly kept out of bidding, billing and CPA math. One month later it is in all three and recommended. Start on CPC, get a clean cost per click and a clean cost per acquisition, and only consider impression billing after you have a baseline to compare it against.

Attribution windows. Now configurable to 7, 14 or 30 day click and 0 or 1 day view, applied across every conversion column. A 30-day click window on a channel you launched two weeks ago will credit ChatGPT for purchases that Meta, Google and email actually drove. Set 7-day click, 0-day view, and compare it against the same window everywhere else.

Platform targeting. Split five ways in this release: Android app, Android web, desktop web, iOS app and iOS web, with matching reporting segmentation. Segment from day one. Mobile app traffic and desktop web traffic on a chat surface will not behave the same way, and if you do not split them at the start you will spend two weeks averaging two different channels together.

One more, if it has reached your account: an Event Quality Score under Tools then Conversions grades eligible Pixel and Conversions API data sources from one to ten on setup quality and returns recommendations. It mirrors Meta’s event match quality score. Check it before you scale, because a poorly matched event feed will make a viable channel look dead.

Sponsored Agents, and what to ignore for now

The same announcement introduced Sponsored Agents, an ad unit where a click opens a separate, clearly labeled conversation with an agent sponsored by the advertiser. It is in test with select US advertisers only.

It is the most interesting thing in the release and the least actionable. OpenAI has not said how it is priced, which model runs the agent, who writes its instructions, or what the advertiser receives from the transcript. Nor has it said what the reporting unit is. A conventional ad produces an impression and a click. A conversation produces neither in a form that fits a column.

Park it. When pricing and measurement are published, it will be worth a second look, particularly for considered purchases where a pre-sale question is the real conversion barrier. Until then it is a pilot, and pilots are not a media plan.

On the numbers everyone is quoting

The channel’s scale figures are getting repeated without caveats, so here are the caveats. OpenAI’s ad pilot began February 9, 2026. It reportedly passed $100 million annualized within six weeks. Digiday reported a $1 billion annualized run rate on August 31, derived from roughly $83 million of monthly revenue. Similarweb reported on August 17 that 26 percent of ChatGPT responses shown to Free and Go users carried a sponsored placement.

Every one of those is either self-reported by OpenAI or estimated by a vendor from a panel. None has been independently verified. They tell you the channel is real and growing fast. They tell you nothing about whether it works for your SKUs.

What to do this week

1. Install the ChatGPT Ads app on one store, not your whole portfolio. Read the permission list first and get privacy sign-off if you sell into the EU or UK.
2. Confirm the catalog sync pulled the right products and that your feed titles read like something a shopper would search, not like internal SKU nomenclature.
3. Confirm the server pixel installed and is firing purchase events.
4. Set CPC bidding, a 7-day click and 0-day view attribution window, and split platform targeting five ways.
5. Run your AOV and contribution margin through the break-even math above before you set a budget. If your contribution per order is under $50, plan this as a retargeting test, not a prospecting test.
6. Budget for traffic quality signals, not for a ROAS verdict. Read add-to-cart rate.
7. Set a calendar check for September 23 if you sell internationally, and watch your CPCs when the additional markets open.

Nothing here is legal or tax advice, and the privacy question in particular is one for your own counsel.

New channels are cheapest and least competitive in the first ninety days, and most expensive to enter after everyone in your category has figured them out. That argues for testing now. It does not argue for testing sloppily, and the defaults OpenAI shipped on September 16 are defaults that will make a mediocre channel look good for at least a quarter.

If you want a second set of eyes on whether this channel fits your margin structure, or you want the test built and read properly alongside your existing Meta and Amazon spend, book a call and we will work through your numbers.