ChatGPT Is a Search Engine Now

Update: the ad just entered the canvas

OpenAI is moving the ad closer to the thing the user is making.

On October 5, the company announced a new visual ad format that it will initially test while people generate images in ChatGPT. The U.S. test starts later this month with a small group of advertisers.

OpenAI says the ads will be labeled, kept separate from the generated image, and will not influence ChatGPT's answers.

Those are good boundaries.

They are also narrower than the product question.

An ad does not have to alter the pixels in the user's image to shape what the user imagines, considers, clicks, or buys next.

That is what makes this launch more important than a new rectangle in the interface.

Image generation is not an ordinary ad break

People use image generators to make things that do not exist yet.

A room they want to renovate. An outfit for an event. A birthday invitation. A product mockup. A wedding table. A logo. A garden. A trip. A version of themselves with different hair, clothes, furniture, or surroundings.

Those prompts can carry unusually rich commercial intent.

The user is not only searching for "green sofa."

They may be showing the dimensions of the room, the color of the walls, the style they prefer, the budget they mentioned earlier, and the outcome they are trying to picture.

OpenAI does not say that advertisers receive those conversations. Its public position is the opposite: conversations remain private, and ads stay separate from answers.

But the platform still has to decide whether an ad is relevant enough to show.

That decision is the part users need to understand.

"The advertiser cannot read your prompt" and "your prompt does not help determine the ad" are not the same promise.

OpenAI should make that distinction painfully clear.

The bigger launch is the measurement loop

The visual format gets the screenshot.

The measurement stack is the business model.

OpenAI is adding integrations with Hightouch, Tealium, and LiveRamp so advertisers can send conversion data from their existing systems into ChatGPT Ads. It is working with a long list of attribution providers across web and apps, plus partners for geo experiments, incrementality studies, and brand measurement.

The company's measurement post says its Pixel and Conversions API can report purchases, leads, installs, and sign-ups. It also says campaign optimization can learn from eligible conversions after both clicks and ad views.

That last part matters.

OpenAI says that in an early analysis, 52.7% of eligible one-day view-through conversions happened within an hour of the matched impression. That is an OpenAI analysis, not proof that the ad caused every purchase. A person can see an ad and buy for another reason. Attribution assigns credit. Incrementality tries to estimate causation. The two should not be presented as interchangeable.

Still, the architecture is clear.

Conversation creates a moment of intent.

The platform shows an ad.

The user may act somewhere else.

The advertiser sends the outcome back.

The ad system learns which future users look more likely to convert.

Very innovative. We have reinvented the advertising feedback loop inside the interface people use to think.

Digiday reports that advertisers had been asking OpenAI for better measurement, more formats, and more inventory. This release addresses the first two directly.

For advertisers, that is normal product progress.

For users, it means the useful transparency question is no longer only:

"Was this answer influenced by an advertiser?"

It is also:

"Why did I see this ad, which signals were used, what happened after I left ChatGPT, and how did that outcome change the next ad decision?"

Brand safety is not user accountability

OpenAI is also developing brand-suitability pilots with DoubleVerify and Integral Ad Science. The companies will evaluate how OpenAI applies its placement safeguards in a controlled environment without seeing private user conversations.

That is a reasonable privacy-preserving design.

But look closely at the job.

Brand suitability asks whether an advertiser is comfortable appearing next to a conversation.

User accountability asks whether a person is comfortable with the ad system interpreting that conversation at all.

Those interests overlap.

They are not identical.

OpenAI's ad policies say ads should stay out of emotionally reliant interactions, mental and personal health conversations, and other sensitive journeys. They also restrict categories such as political ads, gambling, alcohol, and many health or financial products.

Good.

Image generation makes those placement rules harder, not less relevant. A person can generate an image about body image, grief, illness, a child, a breakup, a new home, a financial goal, or a private identity without writing a neat category label into the prompt.

The safety system has to understand the task, not just scan for a forbidden word.

And independent evaluation should eventually answer user-facing questions, not only advertiser-facing ones:

• How often do ads appear in contexts the policy says are sensitive?
• Can users inspect why a specific ad appeared?
• Can they disable ad personalization without losing the product?
• Are targeting categories and advertiser identities available in a public ad register?
• How are view-through conversions matched, retained, and used for optimization?
• Can researchers test placement behavior without receiving private conversations?

The DSA questions in the original article below just became much more concrete.

Separation is the floor

OpenAI keeps repeating three promises: answers remain independent, conversations remain private, and users remain in control.

It should.

Those promises now need operational definitions.

Independent from what: the ad creative, the advertiser's bid, commercial category preferences, or the platform's own incentive to create more ad inventory?

Private from whom: the advertiser, measurement partners, OpenAI's ad-ranking systems, or all of them in different ways?

Control over what: personalization, targeting, conversion matching, ad history, sensitive categories, and deletion?

A label tells the user which rectangle is sponsored.

It does not explain the system that selected the rectangle.

The ad can stay outside the generated image and still enter the decision.

That is why ChatGPT is a search engine now.

And increasingly, an advertising platform with a chatbot attached.

The Short Version

ChatGPT is not just a chatbot anymore.

Europe just made that explicit.

On August 31, the European Commission designated ChatGPT as a Very Large Online Search Engine under the Digital Services Act. Reddit and Roblox were designated as Very Large Online Platforms in the same move.

That sounds like regulatory filing cabinet music.

It is actually a product diagnosis.

The Commission says the services declared at least 45 million average monthly users in the EU, crossing the threshold for the DSA's highest tier. OpenAI's own DSA help page says ChatGPT search had about 159.1 million average monthly active recipients in the EU for the six-month period ending March 31, 2026.

The useful story is not:

"Europe regulates another American tech company."

The useful story is:

An AI assistant that answers questions, searches the web, recommends options, shows ads, shapes purchases, and influences decisions is no longer just a model in a chat box.

It is a distribution surface.

And distribution surfaces get platform rules.

Very normal. The text box has become a search engine with feelings.

The label matters less than the job

There will be a tempting argument about the label.

Is ChatGPT really a search engine?

Is it a platform?

Is it a chatbot?

Is it an operating layer?

Yes. Annoyingly.

The product does not fit cleanly into old internet categories because the category is changing under it. A normal search engine takes a query, retrieves links, ranks them, and sends users elsewhere. ChatGPT can retrieve information, summarize it, synthesize it, remember context, compare options, recommend a next step, and increasingly act on the user's behalf.

That makes it less like a list of links and more like a decision interface.

If I ask for the best CRM for a small team, a travel itinerary, a coding library, a meal plan, a tutor, a therapist directory, a financial explanation, or a product comparison, I am not only asking for information.

I am letting the assistant structure the choice.

That is why the DSA designation matters.

Regulators are not only looking at the model.

They are looking at the service.

The service has users. It has ads. It has search. It has ranking. It has moderation. It has reporting flows. It has minors. It has recommender-like behavior, even when the interface looks like one answer instead of a feed.

The model may be the engine.

The product is the road.

The DSA is asking platform questions

The Commission's VLOP and VLOSE guidance lists the boring things that become very interesting at scale.

Designated services have four months to comply with additional obligations. They need to assess and mitigate systemic risks connected to illegal content, fundamental rights, public security, elections, minors, public health, and mental and physical wellbeing. They need independent audits, compliance functions, transparency around ads and recommender systems, data sharing with authorities, and access for vetted researchers studying systemic risks.

That sounds like paperwork until you put it next to the product.

What is an ad in a conversation?

What is a recommender system when the answer is generated on the fly?

What counts as illegal content when the model paraphrases, summarizes, or combines sources?

What is a ranking system when the user never sees the ranked list?

How do researchers audit a product where every answer depends on context, memory, tools, model version, location, personalization, safety policy, and sometimes browsing?

Very relaxing questions. Perfect for a Tuesday meeting with Legal.

But these are exactly the questions serious AI products now have to answer.

The DSA is not only asking whether ChatGPT said a bad thing.

It is asking whether ChatGPT, as a system, can understand and reduce the risks created by its own scale.

That is a different bar.

Ads make the classification feel obvious

The timing is funny.

On August 18, OpenAI said ChatGPT Ads would expand across Europe. OpenAI said the expansion covers 31 European markets, that ads appear only for Free and Go users, and that Plus, Pro, and Enterprise remain ad-free. It also said ads do not influence ChatGPT's answers, advertisers do not see conversations, and ads are labeled and separate from answers.

Those are the right promises.

They are also the promises of a company that knows it now operates a decision surface.

OpenAI's ads post says people come to ChatGPT while planning trips, choosing software, furnishing homes, learning concepts, or starting hobbies. It says marketers can reach people while they explore, evaluate, and make decisions.

Exactly.

That is the regulatory point.

A search engine with ads is not just infrastructure.

A social feed with ads is not just hosting.

And an AI assistant with ads is not just a model answering questions.

It is an attention and decision market.

OpenAI can say, sincerely, that advertising does not influence the generated answer. But the deeper question is not only whether the ad changes the model output. It is whether the product as a whole changes what users consider, trust, click, buy, ignore, or believe.

That is why old platform rules keep finding their way into new AI products.

The interface changed.

The power did not disappear.

Update: the ad is becoming an agent

OpenAI just made the classification feel even less theoretical.

On September 16, OpenAI introduced new ChatGPT Ads experiences built around something it calls Sponsored Agents. After clicking an ad in ChatGPT, a user can choose to start a clearly labeled conversation with a business-sponsored agent. OpenAI says the sponsored conversation is separate from ChatGPT's independent answers and separate from the original conversation the user started.

That distinction matters.

It also does not make the product simple.

The old ad unit was:

Here is a labeled promotion next to the answer.

The new ad unit is closer to:

Here is a sponsored conversation that can help you evaluate, compare, ask follow-up questions, and then leave for the business's website.

That is not just advertising as placement.

It is advertising as an interactive sales surface.

For users, that may be useful. If I am comparing tables, insurance plans, project-management tools, or local services, a sponsored agent can answer practical questions faster than a static landing page. What size is it? Does it work with my current stack? Can I use it in France? What happens if I return it? Is there a cheaper plan? The best version of this is less banner ad and more helpful product specialist.

But that is exactly why it needs scrutiny.

When advertising becomes a conversation, the boundary between recommendation, sales assistance, customer support, and persuasion gets blurrier. A sponsored agent does not need to secretly influence the normal ChatGPT answer to influence the user's decision. It can influence the decision after the click, inside a branded conversation that still lives in the user's AI interface.

OpenAI is also turning ads into a workflow product for marketers. The same announcement says advertisers can use natural-language prompts in ChatGPT Work to create, update, and analyze campaigns with an Ads Manager plugin. Ads Manager will suggest copy and imagery from a landing page and campaign objective. Advertisers can opt into text customization that adapts headlines and descriptions to conversational context and translates ad copy into the user's preferred language.

That is a lot of agency around the ad itself.

The buyer does not only receive an ad.

The advertiser gets an agentic campaign tool.

The user gets a sponsored agent.

The platform sits between the two, shaping how commercial intent is detected, routed, labeled, measured, and followed up.

The HubSpot and Shopify integrations make the direction even clearer. Businesses can manage ChatGPT Ads from the tools where customer relationships and product catalogs already live. That is not OpenAI putting a billboard on a chatbot. It is OpenAI connecting the ad surface to CRM, ecommerce, campaign management, product feeds, and lead follow-up.

In other words, ChatGPT is not becoming "search with ads" in the old blue-link sense.

It is becoming conversational commerce infrastructure.

That gives regulators and buyers a more concrete question than "are ads labeled?"

Are sponsored conversations logged and reviewable? Can users tell when the agent is optimizing for the business rather than the user's best option? What claims can a sponsored agent make? How are regulated categories handled when a product question turns into health, finance, housing, employment, or education advice? Can advertisers test aggressive scripts? Can OpenAI audit what a sponsored agent said after a dispute? Can minors enter these flows? How does personalization interact with targeting?

The answers may be reasonable.

But the questions are no longer speculative.

The assistant is becoming a place where the ad can keep talking.

Builders should stop pretending the category is neutral

This matters beyond OpenAI.

If you are building an AI product, your regulatory category is not determined by what you call it in the pitch deck.

"Copilot" does not make it harmless.

"Assistant" does not make it private.

"Agent" does not make it enterprise software.

"Search" does not make it objective.

"Recommendation" does not make it a minor feature.

The category comes from the job the product does in the user's life.

Does it retrieve information from the web?

Does it rank options?

Does it recommend products, people, content, software, travel, health resources, financial actions, jobs, or schools?

Does it show ads?

Does it shape decisions for minors?

Does it host user-generated content, bots, tools, stores, apps, or public outputs?

Does it act for the user?

Does it combine search, memory, personalization, and action?

Then you may have built more than a chatbot.

Congratulations. You have product-market fit and a governance problem.

The AI Act is not the whole story

One mistake is to treat this as "AI regulation" in the narrow sense.

The EU already has the AI Act. That matters for model providers, general-purpose AI, high-risk use cases, documentation, evaluation, and deployment duties.

But the ChatGPT DSA designation shows that the AI Act is not the only regulatory map.

AI products also inherit platform law, consumer law, privacy law, competition law, advertising law, child-safety rules, sector rules, and eventually procurement rules.

That is not because regulators are confused.

It is because AI systems are swallowing old product categories.

A chatbot becomes search.

Search becomes recommendation.

Recommendation becomes advertising.

Advertising becomes consumer protection.

Memory becomes privacy.

Plugins become platform access.

Agents become software operators.

The clean "model layer" story was useful for investors and architecture diagrams.

It is less useful once the product is deciding what the user sees and does.

What teams should ask

If your company builds or buys AI products, ask less:

What model powers it?

Ask:

• What role does this product play in the user's decision?
• Does it retrieve, rank, recommend, or advertise?
• Are ads separated from answers in a way users actually understand?
• Can the system explain why a source, product, vendor, or option appeared?
• Does the service have appeal and reporting flows for illegal or harmful outputs?
• Can researchers or auditors study systemic risks without seeing private user data?
• What happens when the same prompt produces different answers because of memory, location, model version, tools, or personalization?
• Are minors using the product, and does the system know?
• Does the product need a public ad repository, transparency reports, or recommender disclosures?
• If an enterprise embeds the assistant, who owns the compliance surface: the model provider, the application vendor, or the customer?

This is not as satisfying as a benchmark chart.

It is also much closer to the real work.

The bottom line

The EU did not make ChatGPT important with this designation.

Users already did that.

OpenAI did not become a search company because Brussels used the VLOSE label. It became search-like because people use ChatGPT to explore the web, compare options, summarize sources, ask follow-up questions, and make decisions. The Commission is now attaching old platform obligations to the new interface because the interface has reached old-platform scale.

That is the lesson.

The next generation of AI products will not be governed only by model cards, safety evals, and API terms.

They will be governed by the jobs they take over.

If the assistant behaves like search, it gets search questions.

If it behaves like a recommender, it gets recommender questions.

If it behaves like an ad platform, it gets ad-platform questions.

If it behaves like an agent, it gets operational-risk questions.

The chat box was a wonderful disguise while it lasted.

But the disguise is coming off.

ChatGPT is a search engine now.

More importantly, it is a decision layer.

And decision layers do not get to stay weightless forever.