Here come the ads. There goes the trust.

by Lisa Peyton and her team of AI Super Heroes

This morning I did some testing, trying to trigger ads on ChatGPT. Despite my overall dislike of ads on any platform (yes, not typical of a marketer, I know!), the ads on ChatGPT seem to be rather benign. It’s also worth noting that only free accounts will be served ads, at least initially. However, when using my pro account to suggest summer tee-shirts, ChatGPT did serve up what looked very much like a featured ad with a ‘visit’ to shop link. Hmmm… with no transparency into why my prompt triggered that specific shopping opportunity, I’m not sure I would trust the results. AND THAT MY FRIENDS IS THE WHOLE POINT.

Anthropic has taken a very strong stance AGAINST ads, which I applaud, and given their valuation it appears to be the right call.

I can see this playing out where ChatGPT becomes more of an AI shopping tool and Claude is the tool where real work gets done. This is how I use both platforms in my own life. ChatGPT will quietly rake in the billions of ad dollars while Anthropic crushes the enterprise AI for work market. Only time will tell, but these radically different approaches to ads will certainly have a MAJOR impact.

What OpenAI just shipped

This week, OpenAI opened its self-serve Ads Manager to U.S. advertisers in beta and added cost-per-click bidding alongside the original CPM model. The $50,000 minimum spend that defined the pilot phase is gone. Agency holding companies including Dentsu, Omnicom, Publicis, and WPP are inside, plus ad-tech platforms Adobe, Criteo, Kargo, Pacvue, and StackAdapt. The company is targeting $2.5 billion in ad revenue this year and $100 billion by 2030.

That’s the launch. Now let’s see what’s happening inside the product.

How I tested

Three prompts, three account states. Apparel (“I’m looking for some good tee shirts. Can you help?”), meal planning (“I’m planning meals for the week for a family of four. Can you suggest five easy weeknight dinners and put together a grocery list I can shop from?”), and kitchen basics (“I’m setting up my first real kitchen and want to invest in good basics. What stand mixer, cookware, and knife set should I buy?”).

I ran each prompt on my ChatGPT Pro account, on a free ChatGPT account logged in, on the same free account logged out, and on Claude as the no-ads control.

The ads themselves are forgettable, by design

The meal planning prompt triggered a HelloFresh ad. The mixer prompt triggered a JCPenney ad(shown below). Both followed the same format. A small favicon. A “Sponsored” label. One-line headline. Short body copy. Small product image. Both placed at the very bottom of long, helpful responses. Both easy to miss.

If this were the whole story, the ads on ChatGPT would barely register. They’re contextual, labeled, non-intrusive, and small enough that a user can scroll past them without thinking. By the standards of what we live with on every other platform, this is restraint.

But the ads aren’t the whole story.

Logged-out users get the heavier ad load

Both ad placements only triggered when I was logged out. Same prompts, same free tier, same browser. Logged in: no ads. Logged out: ads.

OpenAI’s public position is that ads run on Free and Go tiers (Go is OpenAI’s $8/month plan that sits below Plus. Plus and the paid tiers above it stay ad-free, for now). Reports of an unannounced rollout to logged-out users have been circulating for weeks before this week’s official announcement. My test confirms it. Anonymous users are getting served the heavier ad surface. For advertisers, that means matching the prompt and not the person, with no profile data, no retargeting handle, and none of the audience signal that makes paid social work. Think early days of Google Ad’s – like 2005.

Same prompt, different “neutral” answer

This is where the trust problem lives. The mixer prompt produced different brand recommendations depending on whether I was logged in.

Product Type	Logged in (free)	Logged out (free)
Stand mixer	KitchenAid 6-Qt Bowl-Lift Pro, ~$750	KitchenAid Artisan 5 Tilt-Head, ~$400
Cookware	All-Clad Tri-Ply + Cuisinart Pro	Tramontina Tri-Ply Clad
Knife	Misen + Ninja NeverDull	Wusthof Classic 6-Piece

OpenAI says ads don’t influence the model’s answers. The model is producing materially different brand picks based on user state, and the version with the heavier ad layer happens to recommend a more accessible KitchenAid model that fits more shoppers. Coincidence or not, we have no way to tell from the outside. That’s the trust problem in one screenshot.

The Pro account is where it gets uncomfortable

I’m on Pro. Pro doesn’t run ads. But when I asked ChatGPT for tee shirt recommendations, the response surfaced a Fair Indigo Organic Cotton Crew Neck card (shown below). Clickable. Opens a side panel with stock status, “Best price,” and a “Visit” button straight to fairindigo.com.

It is not an ad. It is also functionally indistinguishable from one.

Why did Fair Indigo earn the visual hero slot when Uniqlo got the editorial top pick? Is there a shopping integration relationship? An affiliate cookie? A retailer feed deal? I have no way to know. Neither does any other Pro user. The clean tier still has commercial machinery running underneath the surface, and right now no one is required to disclose what it is.

Why both models picked Uniqlo

Here’s the most useful finding in the whole test. Both Claude and ChatGPT picked Uniqlo as a top tee pick. Claude called it the best quality for the money. ChatGPT cited Good Housekeeping’s 2026 testing and made it the “best first buy.”

Uniqlo isn’t on any list of confirmed ChatGPT advertisers. Both models converged because Uniqlo has won the editorial canon. Wirecutter, The Strategist, Good Housekeeping, NYT have made Uniqlo the default answer for affordable basics for over a decade. Both models trained on overlapping web text where that consensus dominates.

This complicates the easy “AI is biased” reading in a useful way. In this category, the recommendation reflects genuine editorial agreement, not paid placement. The trust problem isn’t that every recommendation is corrupted. The trust problem is we can’t tell which ones are, because there’s no disclosure layer for what’s shaped by retrieval, training data, shopping integrations, or affiliate relationships.

Two tiers of AI visibility

There are now two ways for a brand to show up when an LLM answers a category question.

Tier 1: Editorial canon. Uniqlo, KitchenAid Artisan, All-Clad, Wusthof. Brands that have won enough press coverage to become the default answer when an LLM reaches for “best of.” They show up across models, across user states, across logged-in and logged-out conditions. Durable. Free. Capped only by your ability to earn editorial trust.

Tier 2: Paid placement. HelloFresh, JCPenney, Target through its Roundel retail media network. They appear because someone bought the slot. Rentable. Capped by inventory. Currently expensive.

Worth flagging the wrinkle that came out of my test. Uniqlo won the editorial citation on ChatGPT but did NOT get the visual product card. Fair Indigo did. Two different visibility games, two different mechanics, both happening in the same response. Winning one doesn’t mean winning the other.

What Claude did instead

Claude’s response pattern across all three prompts was structurally different.

• On tee shirts, Claude asked three clarifying questions before recommending anything. Vibe, budget, fit. Then gave a wider list with sale-pattern context for each brand.
• On meal planning, Claude built an interactive grocery list with checkboxes and recipe filter tabs. No brand mentions in the ingredients. No ad layer.
• On the stand mixer, Claude said this: buy knives first, cookware second, mixer last. It explicitly suggested I might not need the thing I asked about.

This is what an assistant looks like when there’s no inventory to fill. It can ask questions. It can tell you to spend less. It can suggest a different priority order than the one you walked in with. Different DNA. Different commercial model.

The strategic split is the real story

ChatGPT is becoming the AI shopping tool. Claude is becoming the AI work tool. This isn’t a tone difference. It’s a structural fork in what these products are for, and it’s accelerating.

ChatGPT has hundreds of millions of weekly users and a $100 billion ad revenue target by 2030. The product has every incentive to push toward purchase moments and every incentive to expand inventory faster than its trust infrastructure can keep up. Claude has the enterprise contracts, the coding workflows, the analyst use cases, and the API revenue from companies that need a model whose recommendations don’t have to be explained.

Both companies have placed their bets. The category is bifurcating in real time. As a marketer, you can’t treat them as interchangeable anymore.

Where this leaves marketers

Four things to do this quarter.

• Audit your editorial canon presence. If your brand isn’t in the press tier that gets cited, you’re not in the answer. The window where you can earn into that tier through actual editorial work is open right now and will narrow as the canon calcifies around early winners.
• Test ChatGPT ads if you’re in an eligible category. The current categories include household goods, local services, travel and entertainment, and digital products and education. The minimum spend is gone. Inventory is constrained, which means early learners get the data advantage before competitors arrive.
• Plan for the split. Where you show up in shopping queries is one strategy. Where you show up in research, analysis, and decision-making queries is another. Different battles. Different rules. Don’t conflate them.

Now We Know

The point isn’t whether ads on ChatGPT are good or bad. The point is whether anyone outside OpenAI can see the wires. Right now, no one can. Marketers are betting on a black box. Users are getting recommendations from a system whose incentives are increasingly commercial and increasingly hidden.

ChatGPT just told us what it wants to be when it grows up. A billion-dollar shopping engine with conversational dressing. Claude is busy becoming the analyst, the strategist, the second brain that gets actual work done. Two products, two business models, two roles in your stack that have damn near nothing to do with each other.

My bet for the next five years is simple. More and more work gets done on Claude. More and more shopping gets done on ChatGPT. Plan accordingly.