ChatGPT Ads and eCommerce: What Merchants Should Actually Do Now

What Actually Shipped

OpenAI began testing ads in ChatGPT in the US in February 2026 and opened self-serve access in the spring, reaching nine markets by August. The European expansion then added 31 more in a single wave, initially through OpenAI's Ads Solutions team and agency and technology partners, with self-serve access following later.

Who sees them, according to OpenAI's own documentation: users on the Free and Go plans only, aged 18 and over, and both logged-in and logged-out users. Plus, Pro, Business, Enterprise and Education remain ad-free. Ads render below the response, labelled as sponsored, and OpenAI is emphatic that they run on separate systems from the chat model and cannot shape or reorder an answer. They are also blocked near health, mental health, political and other sensitive conversations, which is as much a brand-safety feature for advertisers as a user protection.

OpenAI splits its own progress into three phases, and the one now underway is the one to watch. The early platform phase is about advertiser value: expansion past 40 countries, auto-bidding, measurement partnerships, and what OpenAI calls agentic native experiences, meaning automatic campaign creation and sponsored agents.

Four developments matter more right now than the launch itself.

Product Feed Ads Arrived In June

This is the important one. You upload your product catalogue and OpenAI builds ad units from it, featuring, in its own description, "product images, titles, stars, prices, sales prices, and your brand." It supports Google-compatible product data feeds, so the file you already send to Google Merchant Center is largely reusable. OpenAI recommends validating with a sample of around 100 items, then refreshing at least daily. Items expire after two weeks, which is the detail to internalise: a stale feed means your ads stop. That makes the refresh an operational job rather than a marketing one, and a good argument for scheduled orchestration rather than a recurring reminder in someone's calendar.

Real Targeting Controls Landed

You can now upload your own customer and prospect lists as custom audiences, geo-target down to country, state, DMA and zip in the US, and choose which surfaces you appear on across iOS app, Android app and web. Bidding covers CPM, CPC and, since August, conversion-optimised CPC. Minimum daily budget in the US is $25, and OpenAI is running a $500 credit for new advertisers who spend $500.

Getting Started Got Much Easier, And Is About To Get Stranger

OpenAI is rolling out onboarding that takes one input, your website URL, and generates a complete first campaign in under a minute, all editable before launch. Further out it has demonstrated an Ads Manager agent that runs inside ChatGPT itself, briefed conversationally, that builds and optimises the campaign for you. The stated goal is that you never have to learn what an ad group is. For lean teams that is a real unlock. It also means your competitors' barrier to entry is about to drop to roughly zero, which will show up in auction prices.

In-Chat Checkout Got Walked Back

In March, OpenAI deprioritised Instant Checkout and pushed transactions toward merchant-run apps inside ChatGPT. Walmart's EVP of AI acceleration called the results disappointing, "with conversion rates three times lower for the selection sold directly inside the chatbot than those that require clicking out." Etsy said it did not drive large volume but that ChatGPT was a valuable discovery tool. The practical translation: discover in AI, buy on your site. Which puts the weight back on your own checkout experience, not the assistant's.

Almost Total Reach, Almost No Control

Ads serve the Free and Go plans. Go is not a free tier: it costs $8 a month, and OpenAI's pricing page carries the line "This plan may include ads" against it. No other tier does. Paying OpenAI does not make you ad-free. Paying OpenAI more does.

That leaves almost everybody. OpenAI reported more than 900 million weekly active users against 50 million subscribers in February 2026, and told eMarketer that 85% of its US users are eligible to see ads. The free tier is not a poverty filter either: Bank of America went through its own customers' transactions and found only 3% of US households pay for any AI service at all, so removing the payers shifts the income profile of everyone left by well under a percentage point.

So audience size is not the interesting question. What matters is that near-total reach arrives with three constraints attached, and each one breaks a habit.

You cannot buy frequency. Fewer than 20% of eligible US users are shown an ad on a given day. Reach is nearly universal, impressions are rationed, which is close to the inverse of the problem most media buyers spend their careers solving. Plan for coverage, not for weight, and do not expect to buy your way to share of voice.

You cannot choose who. There is no demographic targeting at all. And the only published audit of delivery found ads skewing toward lower-income accounts, with the odds of seeing one falling roughly 2% for every additional $1,000 of ZIP-code median income. Hold it loosely, since the accounts were synthetic, income was inferred from postcode and the result is borderline significant, but it is the only measurement anyone has. Your single corrective lever is first-party list inclusion and exclusion, which is why that item below carries more weight than its length suggests.

You cannot pick the moment. Ad load rises sharply with commercial intent: that same audit saw ads on 5.6% of interactions across a broad prompt mix, while a vendor study weighted toward shopping queries saw over 50%. But the ad follows the person rather than the prompt, so the impression itself can land in a conversation with nothing to do with your category. Creative has to stand on its own.

Almost total reach, no frequency control, no audience control. That is an unusual combination, and right now it is underpriced, because the market is reading the last two constraints as reasons to stay out rather than as the cost of getting in early.

The Design Decision Most Coverage Has Missed

ChatGPT ads are not contextual ads. This is the part worth reading twice.

Be precise about what that means, because the section above said ads show up when the conversation is commercial, and the two facts sit alongside each other rather than against each other. The conversation decides whether an ad fires. The person's accumulated intent decides what the ad is for. Two separate systems, and only the first one resembles anything you have bought before.

Speaking at an ad solutions webinar in August, OpenAI's ads product team described the platform as returning ads "relevant to the person and not strictly just the prompt." The example they gave is instructive: a user spends a few days narrowing down a minivan purchase, fails to find one in their price range, and moves on. Two weeks later, mid-conversation about a weekend trip to Tahoe, they see an ad from a nearby dealership with that exact model, trim and colour in budget. The current conversation is about Tahoe. The ad is about the minivan.

The conversation is about the trip. The ad is about the minivan.

The reason this works is the shape of the input. A search box gets a keyword. ChatGPT gets the whole brief: who you are travelling with, what you are trying to achieve, what your constraints are. OpenAI's own comparison sets three separate Yosemite searches against one sentence that carries the trip length, the group size, the activities and the budget worry in a single breath. Every one of those is a targeting signal that a keyword never captures.

OpenAI calls the effect collapsing the consideration layer: six stages of a hiking boot purchase that used to be spread across review sites, forums, retailer pages and a friend's opinion, compressed into one conversation a brand can contribute to at any point.

That is closer to retargeting against an intent graph than to keyword or contextual matching, and it has three consequences.

First, your ad has to work out of context. It may surface in a conversation with nothing to do with your category, which puts unusual weight on a very small creative unit.

Second, the consideration window compresses, then persists. Research that took weeks now happens in one conversation. But the intent signal outlives that conversation.

Third, attribution gets harder. A click arriving two weeks after the research, from a conversation about something else, will not resemble anything in your existing channel model.

The Numbers, Honestly

Two bodies of evidence point in opposite directions, and both are worth knowing.

Start with where ChatGPT actually sits in a purchase. OpenAI's own breakdown of commerce-related activity on the platform puts product discovery at 31%, single-product evaluation at 23% and product comparison at 22%, against 13% for purchase itself. Treat those as indicative rather than exact, since the categories it lists add up to more than 100%, but the shape is the point: the heaviest use is in the middle of the funnel, where consideration happens, not at the transaction. Which is precisely where a product feed ad can do something useful, and precisely where attribution is hardest.

The encouraging side. Adobe, analysing over a trillion visits to US retail sites, put AI-referred traffic up 138% year on year in May 2026 and roughly 14x above its October 2024 baseline. That traffic converted 54% better than non-AI traffic and delivered 53% higher revenue per visit, a reversal from a year earlier when it converted about half as well. Shopify's Q1 2026 first-party data agrees: AI-referred orders up nearly 13x, 14% higher average order value, conversion roughly 50% above organic search on product detail pages.

On advertising specifically, OpenAI has begun putting names to results. Newegg reported 3x return on ad spend across campaigns over a 28-day period, and 7x during its two-week Fantastech Sale. The most useful of the quotes it published came from Garima Singh, Senior Manager of Paid AI Media at VistaPrint, who said "the majority of traffic ChatGPT drove was new visitors," which is a claim about incrementality rather than volume.

The sceptical side. Those figures were presented by OpenAI at its own sales webinar, with no methodology attached, and they are the wins rather than the average. eMarketer's August assessment was that the ad infrastructure remains basic next to Google and Meta, with underdeveloped performance insights, and advised advertisers to press for real ROAS proof. Its forecast still puts the entire US standalone-chatbot ad market below $1 billion in 2026, against OpenAI's $2.5 billion US projection. On the organic side, Contentsquare put AI-referred traffic at 0.2% of total visits in 2025; that figure is dated in a fast-growing channel, but Shopify is still blunt that organic search refers more sessions than all AI platforms combined.

One widely quoted criticism deserves a closer look, because it is usually read backwards. An SE Ranking study of 50,006 commercial prompts found 14.35% of ads scored no more semantically related to their own query than to a randomly paired one. On the design described above, some share of those are the system working exactly as intended: an ad matched to the person rather than the prompt is supposed to look unrelated to the prompt. The real problem is that a semantic-similarity test cannot tell a well-targeted intent-graph ad from a badly targeted one, and neither can you. Until OpenAI reports on the intent match rather than the query match, roughly one ad in seven is unauditable from the outside. That is a transparency gap, not a relevance failure, and it is the version of this criticism worth taking to a media meeting.

There is also a shopper-side risk that will never show up in a performance dashboard. Asked in January 2026, 63% of US adults said ads appearing in AI search results would make them trust those results less. If that holds, part of the cost of being in the answer is paid in how the answer itself is received, and no advertiser gets a report on that. It is also the stated reason Perplexity walked away from advertising entirely.

Read that as a channel that has graduated from experiment to genuine test line, with intent quality that looks excellent and measurement that does not yet let you prove it.