Who Owns the Pixel? ChatGPT and the Repricing of Advertising

OpenAI now has an advertising platform. A business can create an account, set a budget, upload an ad, launch a campaign in ChatGPT, and measure the result. A developer can do the same work through an API: create campaigns and ad groups, upload creative, retrieve performance, manage product-feed ads, and send conversions back from the advertiser’s website.

Look at those pieces from a distance. The platform owns the conversational surface where intent appears. It runs the auction and decides delivery. It provides the buying interface, the developer API, the measurement pixel, and the optimisation loop. Around that machinery sit ChatGPT’s research, image generation, software creation, browser workflows, apps, product discovery, and commerce systems.

The advertising industry spent more than a century separating those jobs. The publisher supplied attention. The agency understood the client, planned the media, and made the work. Research firms studied the audience. Ad-tech companies moved bids between buyers and sellers. The advertiser owned the product and, eventually, the first-party customer record.

ChatGPT places many of those functions inside one environment. The connections remain incomplete. Ads do not shape answers, app invocation is not a sponsored format, and the Ads API does not automatically turn a conversation into a finished campaign. Yet the direction is visible. Production is becoming easier. Buying is becoming programmable. The publisher can interpret the task, measure the response, and optimise the next impression.

This changes the price of expertise. Large teams, specialised interfaces, publisher access, production capacity, and campaign trafficking lose some of their scarcity. Judgment, product truth, independent measurement, cultural understanding, and the ability to build trustworthy systems become harder to replace.

The question behind the new channel reaches beyond where an ad appears. When one platform can help make the ad, sell the placement, select the audience, record the conversion, and participate in the purchase, who holds the useful knowledge?

Contents

• Advertising Began as Access
• The Stack Moves Inside the Platform
• One Task Replaces Ten Placements
• Everyone Can Make an Ad
• Who Owns the Pixel?
• The Publisher, the Exchange, and the Storefront
• The Agency Advantage Is Repriced
• From Share of Voice to Share of Choice
• Evidence and Direction
• A Practical Experiment
• Trust Sets the Limit
• The Next Reinvention
• Reference Reading

Advertising Began as Access

The early advertising agent sold space.

In the nineteenth century, publishers had pages to fill and businesses wanted to reach readers. Agents connected them. Some bought newspaper and magazine inventory in bulk, negotiated favourable rates, and resold the space to advertisers. Their advantage came from access, price, and knowledge of the publications.

That advantage did not stay scarce for long. Competition drove down margins. Advertisers became suspicious of agents recommending the publications that paid the best commission. Agencies needed a better reason to exist.

They began to add one.

The History of Advertising Trust records how J. Walter Thompson grew from a seller of advertising space into a business offering copy, layout, package design, trademark development, and market research by the 1890s. In the decades that followed, JWT built research operations, radio production, recording studios, and television expertise.

The agency became full service because access alone became cheap and contested. A recent UCLA Anderson analysis of agency evolution traces the same progression: media planning and research came first, creative production followed, and the client relationship expanded from isolated transactions into long-term advice.

Advertising kept repeating this pattern. Each technical era created a scarce capability. Companies formed around it. Technology then lowered the barrier, and value moved elsewhere.

Retail media offered a preview of the current change. The retailer became publisher, data owner, marketplace, and measurement provider. Academic research describes retail media networks through that combination of first-party data, retailer-owned surfaces, and closed-loop sales reporting. The seller no longer waited at the end of the funnel. It built the media environment around the shelf.

ChatGPT extends that convergence beyond the shelf. It begins while the customer is still defining the problem.

The Stack Moves Inside the Platform

The new OpenAI Ads developer hub is more revealing than a screenshot of an ad unit. It describes an operating surface.

The Advertiser API manages campaigns, ad groups, ads, files, and reporting. API keys are scoped to one ad account. Developers can create and update resources, change campaign state, upload creative assets, run asynchronous bulk jobs, and query insights at account, campaign, ad-group, and ad level.

The quickstart shows the whole minimum flow:

Confirm account access
        -> upload creative
        -> create campaign
        -> create ad group and context hints
        -> create ad
        -> retrieve insights

That is familiar ad-tech architecture. Its location makes it consequential. The API serves ads inside the same product where people can express a need, research options, compare trade-offs, create documents and software, invoke external apps, discover products, and move towards a transaction.

The wider product now covers much of the commercial path:

These surfaces are connected by product direction rather than one documented automatic pipeline. A person can use AI to draft copy, create an image, build a landing page, and write the code that calls the Ads API. OpenAI has not documented a single button that performs all of those steps and publishes the result. The important fact is that each former specialist activity now sits within reach of the same user and the same programmable environment.

Advertising was once assembled through a chain of companies. Here, much of the chain can be assembled through a sequence of calls.

One Task Replaces Ten Placements

Consider a simple request:

I want to open a small coffee shop in London. Help me get it ready.

The web usually breaks that goal into fragments. The customer searches for property, equipment, suppliers, finance, payment software, insurance, branding, and local marketing. Each fragment creates a page view, query, audience segment, or retargeting opportunity. Advertising markets grew around those observable pieces.

An agent can keep the larger task in view while the tools change.

Classic digital path

Impression -> Click -> Landing page -> Research -> Comparison -> Conversion

Agent-mediated path

Intent -> Requirements -> Evaluation -> Recommendation -> Approval -> Action

A keyword captures a phrase. A task may contain company size, geography, incumbent suppliers, budget pressure, deadlines, files, preferences, and the definition of a useful result. This gives the platform a richer matching surface than the advertiser can inspect.

ChatGPT Ads use the context and intent of the current conversation alongside the ad, landing page, and advertiser-provided context hints. Those hints can describe relevant conversations, topics, or keywords. They guide delivery without operating as exact-match search terms.

The difference appears in two possible signals:

coffee shop payment system

and:

I am opening a 20-seat coffee shop in London.
I need card payments, basic loyalty, and accounting integration.
The staff should learn the system in one afternoon.

The second signal contains the beginnings of a brief. The platform can interpret it. The advertiser receives aggregated reporting rather than the private conversation.

This asymmetry defines the channel. The platform knows the task. The advertiser must infer which situations produced the outcome.

Everyone Can Make an Ad

For most of advertising history, production imposed a natural limit. Making a television commercial required writers, art directors, producers, directors, editors, studios, and media commitments large enough to justify the work. Digital tools lowered the cost. Generative systems push it down again.

A small business can now use AI to research a category, draft a proposition, produce copy, generate images, build a landing page, create tracking code, and prepare campaign files. An engineer can use the Ads API to publish and monitor the campaign. A merchant can connect a product feed instead of producing one ad for every item.

The surrounding tools do not guarantee a good campaign. They remove much of the logistical argument for a large one.

This is where the new platform reaches beyond media buying. It gives the advertiser access to capabilities that agencies once gathered under one roof: research, writing, design, production, software, trafficking, and reporting. Other AI platforms can join through files, APIs, browser workflows, or MCP-based tools. The user can assemble a temporary production company around one task.

The gain is real. So is the cost.

When production becomes abundant, more weak ideas survive long enough to enter the auction. Variation can become duplication at scale. Brand rules drift. Claims escape review. Landing pages multiply faster than analytics teams can validate them. An API can publish thousands of technically valid ads whose only shared quality is that nobody had to argue hard enough for them.

Large agencies once defended their size through production capacity and operational reach. Those advantages shrink when a small team can generate assets and call the same campaign endpoints. Yet abundance creates work of another kind: deciding what deserves to be made, maintaining coherence, proving incrementality, protecting the brand, and stopping the machine when volume starts to imitate learning.

The scarce resource moves from output to judgment.

Who Owns the Pixel?

The pixel offers a way to see where power settles.

The OpenAI Ads Measurement Pixel is installed on the advertiser’s website. It records events after someone clicks an ad in ChatGPT. The browser SDK can capture a privacy-preserving referral identifier from the landing-page URL, store it in a first-party cookie, and attach event time and source information. Advertisers can send standard events such as content views, checkout starts, orders, leads, and registrations. Optional identifiers must be hashed; raw email addresses and raw customer IDs are prohibited.

The Conversions API provides the server-side path. Browser and server events can share an event ID for deduplication. OpenAI recommends server-side measurement where reliability matters.

Once the event exists, it can do more than close a report. Conversion-optimised campaigns can use a selected tracked event to optimise delivery while billing on valid clicks. The signal travels from the advertiser’s property back into the platform’s decision system.

So who owns it?

The pixel belongs to an advertiser account. The learning it enables is distributed. The brand supplies the commercial outcome. The platform decides how that outcome influences future delivery. The agency may build the measurement and interpret the result, but it cannot inspect the conversation that produced the click.

This is a familiar bargain. Digital advertisers have spent years placing platform code on their properties in exchange for attribution and optimisation. ChatGPT raises the stakes because the platform may hold a more complete expression of intent than a search query or an audience segment.

First-party data becomes more valuable at the same moment that the platform’s context becomes less visible. The brand needs its own record of qualified demand, conversion quality, margin, cancellation, return, activation, and retention. Otherwise the platform learns from the outcome while the advertiser learns mainly from the dashboard.

The Publisher, the Exchange, and the Storefront

ChatGPT is its own publisher in the practical sense: the ad appears inside a surface OpenAI operates. OpenAI also supplies the buying interface, runs the auction, controls delivery decisions, exposes the management API, and provides the measurement tools.

Advertising history gives us reasons to examine that concentration carefully. Google’s combination of publisher tools, exchange infrastructure, and advertiser technology became central to a major antitrust case. A US court found unlawful conduct in parts of that stack; Google has disputed the decision and pursued appeal. The case does not establish anything about OpenAI’s system. It demonstrates why ownership across several sides of an advertising market affects transparency, bargaining power, and independent measurement. Associated Press

ChatGPT adds a further role: it can help organise the customer’s task.

Brands currently have three distinct routes into that task:

1. Paid placement

Ads appear below eligible conversations and remain separate from answers. OpenAI supports CPM and CPC objectives through a relevance-weighted, second-price auction. Ads Manager reporting includes impressions, clicks, spend, CTR, average CPC or CPM, and conversions.

2. Earned discovery

ChatGPT can retrieve, interpret, and compare a brand or product because it appears relevant. OpenAI describes shopping results as organic and separate from ads. Product and merchant metadata, price, availability, quality, and user context can influence the result.

Our earlier work called this the two-layer discovery problem: a brand needs to be legible to inference systems and available through structured, protocol-ready data.

3. Executable participation

Apps in ChatGPT can retrieve information, render interfaces, and take permitted actions. A useful brand can become a capability: check eligibility, configure a product, reserve a table, calculate a quote, or manage an account.

No public documentation currently describes sponsored app invocation or paid placement inside an agent’s plan. The three routes remain separate. Their proximity still changes commerce.

The product-feed advertising workflow makes that proximity concrete. A merchant supplies current titles, descriptions, prices, availability, images, and destination URLs. A campaign selects an eligible product from the feed. A template populates the ad. Product-level insights report what received impressions and clicks.

The catalogue becomes creative input. The same product truth can support organic discovery, paid delivery, comparison, and checkout. Media, merchandising, and commerce begin to work on the same data surface.

The Agency Advantage Is Repriced

Self-service and APIs can look like a direct route around the agency. OpenAI’s own rollout is more complicated. The company named Dentsu, Omnicom, Publicis, and WPP among the agency partners helping businesses buy ChatGPT ads. It also named technology partners that bring the inventory into tools advertisers already use. OpenAI retains control of delivery. OpenAI’s May announcement therefore supports direct buying and intermediated buying at the same time.

The agency is still in the room. Its old advantages no longer carry the same price.

Publisher deals remain relevant across the wider media market. Within ChatGPT, OpenAI owns the inventory and controls delivery, so external scale has less room to create a unique supply advantage. Agencies can still earn preferential support, beta access, operational influence, and aggregated knowledge. Those benefits are weaker foundations than they once were because the platform can expose more functionality directly over time.

The durable agency becomes an independent intelligence and control layer. It knows which data the client should retain, which claims require evidence, which experiments can produce causal learning, and which platform recommendation conflicts with the client’s economics. It builds systems that survive a change of publisher.

An agency that mainly operates another company’s interface becomes easier to compress. An agency that remembers what the platforms cannot know becomes harder to remove.

From Share of Voice to Share of Choice

Traditional media metrics begin with appearance: reach, impressions, share of voice, rank, and click-through rate. An agent-mediated market introduces several decisions before the click. The system retrieves options, interprets them, constructs a comparison, forms a recommendation, and may invoke a capability.

A useful working model is Agent Share of Choice. It is our proposed measurement framework; OpenAI does not currently expose it as a report.

No single dashboard can observe the whole path. Early stages require controlled task testing. Paid response comes from Ads Manager. Transaction and completion belong in first-party systems. Our Agentic Commerce Readiness Stack explores how simulation, observed validation, memory, and Bayesian updates can combine those partial signals without pretending they are equivalent.

This is another opportunity for agencies and practitioners. The platform can report its own delivery. It cannot provide an independent view of the brand across every agent, publisher, retailer, and customer outcome.

Evidence and Direction

The argument becomes stronger when shipped capabilities remain separate from plausible development.

Available now

• A public OpenAI Ads entry point and beta Ads Manager
• CPM and CPC campaign objectives
• Context-hint matching and a relevance-weighted auction
• An account-scoped Advertiser API
• Programmatic campaign, ad-group, ad, file, state, and insights operations
• Asynchronous bulk campaign operations
• Product-feed advertising and product-level insights
• JavaScript Pixel and server-side Conversions API measurement
• Conversion-oriented optimisation for eligible accounts
• Organic product discovery, ChatGPT apps, MCP integration, and commerce surfaces

Direction suggested by the architecture

• Creative generated by AI and passed directly into campaign APIs
• Automated production from a website, catalogue, or product brief
• Agent-managed testing and budget recommendations
• Closer links between paid discovery, product selection, apps, and checkout
• More campaign work handled by small advertiser-side technical teams
• Agency services moving from execution towards systems, evidence, and governance

Still unsupported as a current claim

• One automatic OpenAI workflow that creates the strategy, site, creative, campaign, and budget
• Sponsored app invocation or paid inclusion inside an agent’s recommendation
• Full attribution across conversation, organic recommendation, app use, and checkout
• Universal access to every Ads API capability
• The disappearance of agencies or independent ad technology

The future sections of the stack may connect. They may also remain deliberately separated because privacy, competition, and user trust require boundaries. Strategy should prepare for convergence without reporting it as finished.

A Practical Experiment

The appropriate response is neither panic nor a reorganisation slide. Run a bounded experiment that leaves useful infrastructure behind.

For Marketing Practitioners

1. Map ten customer tasks. Write situations with goals, constraints, trade-offs, and a desired outcome. Begin with real sales, search, support, and research language.

2. Create three context hypotheses. Define the circumstances in which the offer becomes useful. Use those hypotheses to organise ad groups and creative rather than copying a paid-search keyword structure.

3. Build one coherent experience. Make the ad, landing page, product data, and conversion event express the same promise. Use AI to accelerate production, then review every claim, price, image, and action.

4. Instrument before launch. Establish UTMs, Pixel or CAPI events, CRM matching, margin fields, and a pre-test baseline. Decide what success means beyond the click.

5. Compare paid and earned presence. Run the same customer tasks through controlled ChatGPT tests. Record organic appearance, representation, paid delivery, post-click quality, and eventual outcome.

6. Keep an experiment memory. Store the task, creative, audience hypothesis, platform configuration, result, caveat, and next decision. Platform reporting shows performance. Memory turns it into organisational knowledge.

For Ad-Tech and Commerce Engineers

1. Build an account-scoped API client. Keep each advertiser’s key isolated. Model campaign, ad-group, ad, file, and insights resources explicitly.

2. Add safe campaign operations. Use idempotency where your orchestration layer needs it, asynchronous jobs for bulk work, retries for transient failures, and approval gates before activation or material budget changes.

3. Separate generation from execution. Let a model draft titles, descriptions, context hints, and asset specifications. Validate schema, policy, evidence, destination, and spend before the Ads API receives them.

4. Implement both measurement paths. Use the browser pixel where appropriate and CAPI for reliable server-side events. Deduplicate shared events. Hash permitted identifiers correctly and retain consent records.

5. Connect product truth. Version catalogue data, validate required fields, monitor freshness, and reconcile advertised price and availability against the transaction system.

6. Build an independent report. Join OpenAI insights with analytics, CRM, orders, returns, and support outcomes. Preserve the raw inputs needed to challenge a platform recommendation.

For Agencies and Marketing Leaders

1. Inventory the fragile value. List services that exist because campaign tools are difficult, production is slow, or platform access is restricted. Assume those barriers will fall.

2. Identify proprietary knowledge. Determine which benchmarks, causal findings, taxonomies, creative principles, and customer truths would remain if every platform login disappeared tomorrow.

3. Build reusable control planes. Standardise approvals, evidence checks, budget limits, naming, measurement, audit history, and rollback across publishers.

4. Form smaller mixed teams. Put strategy, creative judgment, data, and engineering around one commercial problem. More hand-offs will not protect work from automation.

5. Sell better decisions. Measure the quality of allocation, experiments, learning, and customer outcomes. Hours and asset counts will become increasingly weak proxies for value.

Trust Sets the Limit

The value of an agent grows when people delegate consequential work. Delegation depends on confidence that the system is helping them rather than steering the answer towards the highest bidder.

OpenAI’s current principles draw boundaries around that risk. Ads are labelled and placed separately from answers. Advertisers do not receive conversations, memories, or personal details. Shopping results are described as organic and independent from ads. Users can control how their data is used for advertising.

Those rules are part of the commercial architecture. A user who distrusts the recommendation delegates less. A platform that weakens answer independence may create more inventory while making the surrounding task less valuable.

Concentration still raises practical questions:

• How can an advertiser audit exclusion without seeing private conversational context?
• How can independent measurement verify an auction run entirely by the publisher?
• How should conflicts be handled when an ad, an organic product, and an app can all serve the same task?
• Which conversion signals improve reporting, and which also influence personalisation or optimisation?
• Who carries responsibility when the recommendation is sensible but the product data is stale?
• How can a brand move its experiment history when the platform changes?

Trust will set the ceiling for the channel. Independent evidence will determine how much of that trust advertisers should borrow.

The Next Reinvention

The first advertising agents survived the decline of space brokerage by learning the advertiser’s business. They added research, planning, writing, design, production, and strategy. Their value moved because the market around them moved.

ChatGPT applies pressure to those services together. It lowers the cost of making assets, exposes campaign operations through software, owns the media surface, interprets intent, records outcomes, and moves closer to commerce. It also creates a new field of work around product truth, agent visibility, experimental memory, governance, and independent measurement.

For the person opening the coffee shop, this may feel simple. Describe the goal. Compare equipment. Create a site. Find a payment provider. Build a campaign. Ask for the next step.

Behind that simplicity, the advertising stack has changed shape. The advertiser can do more. The platform can see more. The agency must know more than how to operate the platform.

The next advertising company may employ fewer people who know where to place an ad, and more people who know which decisions should never be left to the platform selling it.

Reference Reading

OpenAI Ads Platform

• Advertise in ChatGPT - advertiser entry point, campaign workflow, reporting, and trust principles.
• New ways to buy ChatGPT ads - self-service rollout, agency and technology partners, CPC buying, privacy, and measurement.
• Ads in ChatGPT: The Basics - format, context signals, auction, pricing, reporting, and brand safety.
• OpenAI Ads developer hub - measurement and Advertiser API documentation.
• Advertiser API overview - account-scoped authentication, resources, bulk operations, and insights.
• Advertiser API quickstart - end-to-end campaign, creative, and reporting flow.
• Product-feed advertising - catalogue-driven campaign creation and product-level measurement.
• Conversion-optimised campaigns - optimisation against tracked conversion events.

Measurement and Privacy

• JavaScript Measurement Pixel - browser events, identifiers, first-party referral storage, and deduplication.
• Conversions API - server-side conversion events.
• Conversion measurement - Ads Manager measurement guidance.
• ChatGPT privacy controls - answer independence, advertising privacy, and user controls.

Apps, Discovery, and Commerce

• ChatGPT release notes - Chat, Work, Codex, and product consolidation.
• Developers can submit apps to ChatGPT - app directory, MCP foundation, and contextual discovery.
• Powering product discovery in ChatGPT - merchant feeds, ACP, and product discovery.
• Shopping with ChatGPT Search - organic product selection and shopping context.

Advertising History and Market Structure

• J. Walter Thompson archive - the evolution from space selling to research, production, and full service.
• A Model of the Evolution of Full-Service Advertising Agencies - economic and historical analysis of agency development.
• Retail media networks: definition, development, and differentiation - first-party data, retailer-owned inventory, closed-loop reporting, and trust.
• Google digital advertising antitrust ruling coverage - context on vertically integrated publisher and advertising technology.

Our Related Research

• ChatGPT Ads Go Self-Serve - campaign activation and measurement mechanics.
• The Two-Layer Discovery Problem - inference-based and protocol-based brand discovery.
• The Agentic Commerce Readiness Stack - simulation, validation, memory, and governed execution.