ChatGPT Ads: An Agency Playbook for a Controlled Test

A ChatGPT Ads pilot needs to end with a decision the business can defend. Did it reach useful demand? Did the resulting leads or customers justify the cost? What evidence supports another test, a larger budget or a stop?

Buying clicks is one part of answering those questions. The campaign also needs a clear offer, a destination that fulfils it, reliable conversion records and an evaluation method that can distinguish platform attribution from additional business. If any part fails, the dashboard may look decisive while the test remains inconclusive.

OpenAI’s 5 May 2026 announcement introduced beta self-serve buying and cost-per-click (CPC) bidding alongside pixel and server-side measurement. That was this article’s original starting point. The operational guidance below was revised against documentation checked on 24 October 2026. Access, formats, bidding and measurement have since expanded; confirm the settings available to the actual account before activation.

Define the decision before the campaign

Treat the pilot as a connected system:

Business question and approved test plan
    → eligible account, campaign and audience
    → relevant offer and truthful creative
    → working landing page and customer action
    → permitted, validated conversion events
    → platform reports and independent business records
    → evaluation against a stated decision rule

The parts need different kinds of evidence. An approved ad can still fail to deliver. A received conversion event may never be attributed. A correctly attributed lead can still be useless to sales. Even a valuable customer might have bought without the ad.

Begin with a question narrow enough to answer. For a customer relationship management (CRM) software provider, one useful test would ask whether a campaign addressing small sales teams produces qualified trial users at an acceptable acquisition cost. Follow that question from the offer to the trial, then check whether the trial becomes useful product adoption.

Agree on three levels of success:

• Operational: the intended campaign delivers and measurement behaves as designed.
• Commercial: resulting customers or qualified leads meet the business’s quality and cost criteria.
• Causal: an appropriate comparison supports an estimate of business that would not otherwise have happened.

A small pilot may answer the first two while remaining too small for the third. Record that limitation before spending. Calling a traffic test an incrementality experiment does not give it the necessary design or volume.

Choose a client with an eligible offer, a dependable destination, enough activity to investigate the question, and people who can reconcile media results with customer records. Consider the buying cycle as well as conversion volume. A trial can happen quickly; proving retained customer value takes longer. A thin-margin account that requires immediate proven efficiency may have little room for an exploratory test.

Set the outcome, maximum authorised spend, observation period, reporting owner, quality checks and stop conditions. Include fees and the work required to run the pilot. That is the test plan against which later campaign changes should be judged.

Confirm access and the terms of the test

There are still three useful access routes: direct Ads Manager, an agency relationship, or a technology partner. OpenAI’s May announcement named agency groups including Dentsu, Omnicom, Publicis and WPP, and technology partners including Adobe, Criteo, Kargo, Pacvue and StackAdapt. Those are documented routes to investigate, rather than a promise that each partner exposes every current feature.

Access now extends beyond the US. On 21 October, OpenAI announced a further country expansion. Its current Ads Manager availability list includes the UK and US among supported self-service countries and ties eligibility to the legal entity being billed. Creating an account is separate from gaining access. Confirm the advertiser’s country, domain, verification, billing and permitted campaign markets.

The campaign setup guidance also says some new self-serve accounts initially serve only their home country. Account restrictions are separate from the locations selected in a campaign. Saving an international target does not establish that it can deliver.

Check the offer against current ad policies. Category eligibility, geographic restrictions, truthful claims and the complete destination experience all matter. A research-heavy buying process does not by itself establish eligibility for financial, health or other regulated advertisers.

Ask partners for the evidence and control you need

A partner should reduce operational friction without making the pilot impossible to inspect or move.

Area	Establish before committing spend
Access and responsibility	Which client account is used, who can change it, who approves activation, and who handles support
Measurement	Access to the intended data source, event definitions, diagnostic tools, reporting exports and any account-specific restrictions
Commercial terms	Media cost, platform fees, markup, minimum charges and the total test commitment
Data and continuity	What history and business records the client can retain, export and use after the engagement
Experiment support	Which holdout or geographic design is actually possible, its prerequisites and who will analyse it
Exit and communication	Exclusivity, cancellation, transfer, and permission to share results or a case study

Keep commitments concrete. Access to an attribution partner is different from a supported lift study. A fee agreement does not establish data rights. Ask for the deliverables the test needs, then make those responsibilities part of its operating record.

Build creative around a customer task

Conversation can carry constraints that a short search leaves unstated. For the CRM test, the brief might concern a small team comparing follow-up automation, migration effort and pricing. The advertiser needs an offer that helps with those decisions and a destination that shows the relevant evidence.

This is a hypothesis about customer needs derived from research, sales or support. It is not a transcript supplied by the platform, nor knowledge of the exact conversation in which an ad will appear.

OpenAI’s targeting documentation distinguishes context hints from explicit geography, platform and audience settings. Hints describe relevant use cases; they are neither exact-match keywords nor a guarantee of delivery against a phrase. The current system also documents audience inclusions and exclusions. A hint mentioning a location does not replace geographic targeting.

Translate familiar search habits carefully. Keyword research can reveal needs, but a keyword cluster is not an exact delivery contract here. Landing-page relevance helps explain an offer; it is not a substitute for an exclusion control. Customer audiences, where supported and permitted, are different from buying access to private chat histories.

For the CRM brief, develop a small set of distinct propositions, such as reducing manual follow-up or understanding migration before starting a trial. Keep the offer honest about what the product does. Copy directions might include “Compare follow-up workflows for a small sales team” and “See migration steps before starting a CRM trial,” provided the destination actually supplies those resources.

Within a copy test, keep the destination and other material settings stable. If testing a landing page, hold the creative stable where the campaign structure permits. Changing the proposition, page, bid and audience together can produce movement without showing which decision caused it. Where delivery is algorithmic rather than randomly split, an apparent winning variant is a useful lead for further testing, not a clean causal estimate of creative superiority.

Match the available format without implying endorsement

The advertiser basics now describe image assets alongside the advertiser name, icon, headline, copy and destination. Plan for the formats available to the account, use their supported specifications and preview the result.

Two further developments have different scopes. The September update describes a US test of Sponsored Agents and opt-in AI text customisation. The October update announced a visual-format test during image generation and Negative Phrases for qualifying advertisers. These are not interchangeable options available to every buyer. Negative Phrases address particular placement requirements; do not assume familiar search-negative behaviour.

Record whether text customisation is enabled, who reviews its effects, and which formats or placements the pilot includes. A test of approved copy becomes harder to interpret if its language is also being automatically adapted.

OpenAI’s user guidance says ads are labelled, separate from answers, and do not give advertisers access to private chats or memories. Eligible Free and Go experiences may carry ads; the listed paid ad-free plans do not. Platform personalisation settings affect selection, but do not provide the advertiser with the underlying private context.

The creative should offer a useful next step and accurately describe the click. It should remain recognisable as advertising. “ChatGPT recommends us” is not a conclusion a paid placement supports. The Answer Independence essay examines that trust boundary in depth.

Build the measurement path before activation

The measurement system has three jobs: record what happened, send only permitted events, and connect the appropriate definitions to the campaign. Instrumentation is part of the test’s design, rather than a repair after the first week of buying.

Define actions in business terms

For the CRM provider, a page view, trial start and qualified account are different outcomes. Decide when each action has really occurred, which system establishes it, and who owns the definition. A button click should not become a trial merely because the button says “Start trial.”

The supported-events reference retains the web event vocabulary below:

Events	Meaning and data shape
page_viewed, contents_viewed	Page or specific content activity; contents
items_added, checkout_started, order_created	Cart, checkout and completed purchase; contents
lead_created, registration_completed, appointment_scheduled	A recorded customer action; customer_action
subscription_created, trial_started	A subscription or trial begins; plan_enrollment
custom	A distinct action requiring its own name; custom

Event names and data shapes are separate fields. If sending value, use the documented currency and integer minor units: USD 49.00 is 4900. Define whether that value represents a payment, estimated lead value or another business measure. A trial is not subscription revenue.

The conversion-tracking guide separates the data source that receives events, the event setting that defines a conversion, and the campaign to which that setting is attached. Reuse the intended source for browser and server measurement. Receiving events alone does not configure a campaign goal.

Preserve references and distinguish their jobs

The Measurement Pixel sends browser events and captures an available oppref from the landing URL into the first-party __oppref cookie. Preserve the original reference through redirects. For server events, explicitly carry the available oppref into the backend record and outgoing payload; the server API does not capture it automatically.

Keep three identifiers distinct:

• Pixel ID: identifies the measurement source.
• oppref: links eligible activity to an ad click; it does not reveal the conversation.
• Event ID: identifies one business action, so retries and duplicate delivery can be recognised.

The browser reference, obref, is another supported matching signal. When permitted and available, the Conversions API accepts it as user.obref, copied unchanged from the Pixel’s __obref cookie. It belongs inside user; oppref belongs on the event itself. Neither should be invented when missing.

Use the browser and server deliberately

The browser can observe website interaction. The server can send an action confirmed in the business’s own system, such as a successfully created trial account. Combining them reduces dependence on one delivery route; it does not eliminate missing data or override a person’s measurement choices.

For the same action, use the same Pixel ID and event name, with the same identifier under the browser’s event_id and the server’s id. Custom events also need the same custom name. Reuse that identity for retries; create a new one for a genuinely new action. OpenAI documents first-received-event deduplication, so make the two versions consistent rather than expecting the second to repair missing value in the first.

For example, a confirmed CRM trial could use trial_started and an internal action identifier such as trial_042 in both routes. A browser event should wait for confirmation that the trial exists. If it is sent when the form merely opens, deduplication may preserve an incorrect account of what happened.

Do not send server credentials through the page. The Conversions API belongs on the server, with protected credentials and controlled retries. Its reference currently limits batches to 1,000 events and says an invalid event rejects the whole batch. Validate before sending, retain rejected-event diagnostics and monitor delivery failures. A successful HTTP request is evidence of transport, not proof of attribution or customer quality.

Make permitted collection part of the implementation

The Pixel’s consent control defaults to enabled unless explicitly disabled or a stored denial is found. Where consent is required, set the intended state before initialisation and test denial and withdrawal as well as acceptance. Forwarding events through a server does not exempt them from the site’s measurement requirements.

Audit advanced matching as well as the visible event names. The Pixel documentation describes automatic collection and hashing of supported customer information when that feature is enabled; server-side matching also supports specified identifiers. Hashing changes the representation of an identifier. It does not establish permission to send it. Choose necessary, permitted fields and inspect the actual payloads.

Add static analytics tags such as utm_source=chatgpt and utm_medium=cpc, plus stable campaign and creative identifiers. Keep free-text customer needs and sensitive details out of URLs. UTMs support an independent view of arrivals; they cannot recover an unobserved conversation or turn a later sale into proof of ad influence.

Prove that the records behave as intended

Before buying media, trace a test action from the landing page to the business record and intended conversion source. Keep schema validation, received events and attributed outcomes separate.

Check	Evidence to retain
Completed action	The trial, lead or order exists in the system that owns it
Abandoned or failed action	No completion event is emitted merely because a form was opened or submitted unsuccessfully
Browser and server report the same action	Shared source, event name and identity; one action remains one conversion
Refresh, webhook replay or network retry	The original action identity survives; no new conversion is invented
Redirect or missing click reference	Available references survive; missing references stay missing
Denied or withdrawn measurement consent	Browser and server follow the intended collection decision
Value, currency and timing	Payload agrees with the underlying business record
Campaign configuration	The intended event setting is attached to the correct campaign

Use the documented validation-only mode to check an API request without saving events. It does not prove receipt in live monitoring. Then, through an approved test procedure, verify event delivery to the intended source. The conversion guide describes a recent-event sample of roughly the last 15 minutes; that diagnostic view is not historical attribution reporting. Record which test actions were sent so they are not mistaken for campaign performance.

Verify the destination too: mobile layout, form success and error paths, pricing, terms, confirmation and the next useful step. The CRM trial needs a working route into the product, not merely a conversion tag. Assign an owner for tracking changes and a way to detect when a website release breaks the event path.

Launch with bounded spend and explicit controls

Choose the buying method for the question. CPM means paying per thousand impressions; CPC means paying per valid click. A controlled CPC test can help inspect post-click demand, but a click objective does not promise qualified customers.

The current conversion-optimisation guide describes oCPC in open beta: delivery seeks a selected conversion while billing remains per valid click. For fixed bidding, its CPA bid is an optimisation input, not a price charged only when a conversion occurs. The account must support the feature, and the goal uses one active standard event; custom events are not optimisation goals.

The bidding and budget guide also documents fixed bids and Maximize Results strategies. The latter use daily budgets. Confirm the account’s available strategy, the campaign objective, the billing event and the budget together. Do not copy a bid value between objectives without checking what it means.

For the CRM pilot, optimising to trials can make delivery more efficient at finding trial starts. It cannot teach the platform that a trial became a valuable customer unless an appropriate signal is available. Begin with a goal you can measure reliably and evaluate at sufficient volume. Keep downstream activation and lead quality in the business’s own assessment.

OpenAI’s advertiser basics give USD 3–5 as starting maximum-CPC guidance. That is a dated platform suggestion, not a forecast of clearing prices or a recommendation to spend beyond the client’s economics. Set bids in the account currency, check current guidance and compare the likely cost with what the test can afford to learn.

Build campaigns paused, inspect the complete settings, then obtain the already-defined approval for activation. Record geography, platforms, audiences, exclusions, context hints, creative, destination, conversion goal, automated customisation, strategy, budget and schedule. Review delivery and measurement daily at first, while allowing for reporting delay before changing the plan.

If using a plugin, API or agent to operate the account, those controls still apply. Separate permission to draft from permission to activate or increase spend. Confirm the client account and approved limits at the point of change, inspect the saved result, and keep a way to pause. The Secure Marketing Agents guide develops that execution architecture; automation does not change this pilot’s authority or evidence requirements.

Report attribution, customer quality and lift separately

Give the channel its own reporting line, with the test question, settings, costs and limitations visible. Report spend, impressions, clicks, click-through rate, average CPC or CPM, attributed conversions and their cost. Add purchase value and return on ad spend only when the event values justify them. Report fees consistently and state which costs a ratio includes.

For the CRM provider, follow trials into activation, sales qualification and eventual revenue. A cheap trial and an activated customer are different acquisition outcomes. Reconcile business records without pretending every customer can be linked to an ad.

Record the report’s definitions

The Reporting guide currently documents API defaults of 30-day click attribution, one-day view attribution and ad-event time, which assigns results to the date of the ad interaction. It separately exposes click-through and view-through counts. Null values indicate unavailable results, not zero. Recent conversions require processing time and may change.

There is a documentation distinction to resolve in the actual account: the Pixel guide describes Ads Manager’s Conversions as click-through only, while the general Reporting guide defines API conversions as click-through plus view-through under the selected windows. Do not assume identically named totals match. Preserve the export surface, windows, time basis and date range; confirm the definitions before comparing or combining them.

Keep click-attributed and view-attributed results separate. A conversion after an eligible impression is not proof that the impression caused it. Match time zones and allow the reporting window to mature before comparing a new campaign with a completed one. Do not add totals from different platforms as though they describe mutually exclusive customers.

Use independent records without promising complete visibility

Compare platform reporting with permitted web analytics and CRM or order records. Investigate discrepancies through event definitions, matching, consent, deduplication, time windows and processing delay. Agreement is useful; disagreement is a diagnostic question, not automatic proof that one system is wrong.

Third-party measurement is no longer simply a future item. OpenAI’s October announcement describes attribution and data integrations, early geo-based incrementality work, and controlled brand-suitability pilots. Those services solve different problems. An attribution integration does not automatically provide a lift experiment, and a suitability evaluation does not expose real users’ private conversations.

Ask the chosen partner which markets, accounts, events, exports and study designs it supports. Keep early vendor case results attributed to their source; they are not a benchmark for this client’s CRM campaign. Who Owns the Pixel? explains the wider asymmetry between platform context and the advertiser’s business evidence.

Reserve causal language for an appropriate comparison

A baseline of direct visits, branded search, trials and CRM activity helps reveal changes and tracking faults. It cannot isolate an ad’s effect from seasonality, other campaigns, pricing, press or changes in the product. Describe that movement as a signal to investigate; calling it “assisted lift” would claim more than the comparison establishes.

If feasible, design a holdout before launch: comparable units receive different advertising exposure, with the outcome and analysis specified in advance. Vaver and Koehler’s geo-experiment research describes random assignment of non-overlapping regions to treatment and control. Applying that method here requires supported targeting, enough usable regions and activity, and an analyst who can assess the design.

Check contamination between test and control, changes in other media, the detectable effect, conversion delay and uncertainty. A simple before-and-after comparison or one convenient pair of regions is not automatically a credible experiment. If the available budget or targeting cannot support the design, report operational performance and attributed customer quality, with incremental impact unresolved.

Learning has value too. A well-documented proposition can inform sales material or suggest a later landing-page test. Keep that separate from proven revenue or savings; a promising message from one channel still needs testing elsewhere.

Diagnose the failure, then decide what to change

Use the whole system to locate the problem before reaching for the bid control.

No delivery: inspect account eligibility and review, billing, campaign and ad-group status, schedule, budget, targeting and creative approval. OpenAI’s troubleshooting guide explicitly distinguishes an active resource from one able to serve. Only after those checks should narrow opportunity, context fit and bid competitiveness become the main explanation.

Clicks without analytics arrivals: check redirects, destination availability, analytics consent and collection. An analytics session and a billable click have different definitions. OpenAI’s advertiser basics also warn that ad-review crawlers can visit tagged destinations without generating billable clicks; a tagged visit alone is not a paid customer arrival.

Arrivals without recorded actions: inspect the offer, page and form, then the event path. A genuine lack of trials and a broken trial event require different remedies.

Many recorded actions but weak customers: check whether the conversion definition rewards an easy action instead of the desired result. Revisit proposition, eligibility and product onboarding before asking the platform to find more of the same.

Conflicting conversion totals: reconcile identities, duplicate handling, campaign goals, windows and reporting clocks. Increasing spend makes a broken measurement system more expensive.

Change one material hypothesis at a time where possible and record the version and date. Treat promising themes as candidates for replication before expanding to adjacent needs or raising budgets substantially. Stop when the agreed limits are reached or when the evidence cannot support the decision the pilot was designed to inform.

The lasting deliverable is a learning record: question, approved settings, creative and page versions, event definitions, quality checks, costs, results, uncertainty and next decision. Another operator should be able to see what happened and why the team acted on it. Memory & Agency develops the conditions for reusing that knowledge: preserve the result’s scope and support so the next operator can judge whether it applies to a different campaign.