Chatbot advertising grew a measurable spine: three questions before you fund it

What OpenAI shipped on 16 September, and what is still beta#

A fox reads a hedgerow by what has changed in it, not by what the birds are shouting, so start with the facts as OpenAI states them. On 16 September 2026 the company said it is testing Sponsored Agents, which let people start a conversation with a business-sponsored agent after clicking an ad in ChatGPT. The chat is clearly labelled, distinct from ChatGPT's independent answers, and separate from the conversation the user began. It is being tested with select advertisers in the United States.

Read the small print before you sprint. OpenAI's help centre calls Sponsored Agents a limited alpha test, open only to selected advertisers, with no early access requests accepted. An email to advertisers, published by Search Engine Roundtable, described a conversation that users may be able to start, a more tentative phrasing than the blog post. Chatbot advertising, in other words, has a new format that almost nobody can buy yet.

The performance half of the chatbot advertising news is more immediately useful, because you can buy it now. OpenAI's conversion-optimised campaigns page describes two billing models. oCPC bills on valid clicks. oCPM, optimised cost per thousand impressions, bills on impressions served while the system optimises toward a conversion event, counting actions that follow ad clicks and ad views rather than clicks alone.

Which of these is generally available? Here the trail forks. PPC Land reports, from an OpenAI email to advertisers, that conversion optimisation with impression billing is now generally available and that OpenAI recommends it for performance. We could not find a general-availability label for oCPC on any OpenAI page we read, and every help page we opened still titles the product Ads Manager Beta. Treat "generally available" as email wording until the documentation catches up, and treat oCPC's status as unconfirmed.

Three other changes round out the week. Ads Manager now targets five platforms separately, namely Android app, Android web, desktop web, iOS app and iOS web. An Ads Manager plugin lets you create, update and analyse campaigns by natural-language prompt inside ChatGPT, and HubSpot and Shopify integrations put the channel inside tools businesses already run, all per OpenAI's announcement. And an opt-in setting adapts headlines and descriptions to the conversation and translates them automatically. Hold that last switch in your mind, because it matters in the next section.

Chatbot advertising needs a script: what may the agent say?#

Chatbot advertising differs from every search ad you have approved, because the ad speaks back. A text ad is a fixed string a compliance officer can sign off in an afternoon. An agent produces sentences nobody has read yet, in a conversation nobody can preview. That does not make it forbidden. It makes the script, the boundaries and the logging your job.

Start with what OpenAI's ad policies already demand of chatbot advertising. The standards cover all ad assets, including copy, images or video and landing pages, and must hold end to end. Ads must be truthful, with no unfounded claims about capabilities, pricing or outcomes, no exaggerated results and no false endorsements. They must also be distinguishable from ChatGPT itself, and OpenAI may act on ads that imitate its appearance, functionality or voice.

Here is the gap. The policy text we read is written for ad assets, and it does not say in terms how it applies to a live agent conversation. PPC Land notes that OpenAI's post also does not say who writes the agent's instructions, which model runs it, or what the advertiser receives from the exchange. Until OpenAI says, assume the strictest reading, because a regulator will.

Now add your own regulator. In the US, the Federal Trade Commission's native advertising guide says an ad should not suggest or imply it is anything other than an ad, and that a disclosure, where one is needed, must be clear and prominent. A label helps. A label is not a licence to say anything beneath it.

The regulated categories are a narrow gate#

OpenAI says ads are primarily limited to consumer verticals during the test, and that financial services, healthcare and legal advertisers may be approved manually, case by case. In the US, the policy names ten financial product types it may allow, from credit cards to mortgages, and four it will not, including credit repair and debt settlement. For health it names nine categories it may allow and six kinds of product or claim it rules out, including unsupported treatment claims. Advertisers may be asked for proof of licensure, and gambling ads are disallowed.

So the first question for a regulated chatbot advertising buyer is not whether the agent is clever. It is whether the advertiser is eligible at all, in the market where the test runs. The same policy says ads for financial services outside the US are generally prohibited, and the same goes for health services. A European lender should check that before briefing anybody, not after.

Write the agent's rules like a promotion sign-off#

Both lines are folkfox's illustrative examples, not OpenAI's text, written to show the gap between a promise and a process. The rules behind the second line of chatbot advertising copy are dull and durable: an approved claims library, a list of outcomes the agent may never predict, no personalised advice, a handover to a human on request, and a full transcript log kept for the length of your regulator's retention period.

One switch deserves special suspicion. OpenAI's announcement says advertisers can opt into AI-powered text customisation, which adapts headlines and descriptions to the conversation and translates them automatically. For a regulated promotion, every adapted line is a new sentence and every translation is a new legal text. In chatbot advertising, leave the setting off until compliance has approved the method, not merely the original copy.

Then mind the moment in chatbot advertising. OpenAI's ads basics page says ads are not shown to Plus, Pro or Business users or to under-18 accounts, and its placement policy keeps ads away from mental and personal health conversations. A sponsored chat inverts that comfort, because the user now types straight to the advertiser's agent and may volunteer exactly the detail the placement rules exist to protect. Build the agent to decline it, and to say why.

What the oCPM optimiser learns to hunt#

In chatbot advertising, an optimiser is a hound with one scent. Point it at an event and it hunts the cheapest route to that event, whether or not that route leads to revenue. That is a property of optimisation in general, not a fault OpenAI has confessed to, but the design of conversion-optimised campaigns makes the choice unusually weighty. Each campaign uses one standard event, fixed at creation, and custom events are not supported. Existing CPM, CPC, oCPC or oCPM campaigns cannot be switched to another objective, so a wrong guess means a new campaign.

OpenAI's Conversions API documentation lists the standard events: appointment scheduled, checkout started, contents viewed, items added, lead created, order created, page viewed, registration completed, subscription created and trial started, with app installed and app opened for native apps. Line them up by depth and the trap becomes plain.

If the chosen event is a page view, the quarry is anyone who loads a page. If it is a registration and registration is free, the quarry is anyone willing to hand over an email address for nothing. Neither is wrong for every chatbot advertising business. Both are wrong for a lender whose margin lives three steps later, at the funded loan. OpenAI's own tips say to pick an event with enough signal, because events that happen too rarely make performance hard to judge. The two pressures pull apart, and the resolution is a judgement, not a setting.

That forecast is tidy, and it is testable. If a sponsored chat screens out the idly curious, expect fewer clicks, a higher conversion rate and about the same cost per genuine conversion. If it merely adds a step, expect fewer clicks and the same rate. Only a comparison against a matched campaign says which fox is in the hedge, and we found no published performance data for Sponsored Agents, so the thread is a forecast, not evidence.

View signals widen the net, and the net can over-credit#

The other change beneath the bonnet is that oCPM lets the system learn from conversions after views as well as clicks. OpenAI's stated rationale is that someone may see an ad while exploring an idea, return later and convert without a direct click. That is plausible. Chatbot advertising inherits the classic view-through problem: a person who was always going to buy can be credited to an ad they merely scrolled past.

PPC Land recalls that when view-through conversions first arrived on 19 August, they were kept out of bidding, billing and cost-per-acquisition arithmetic, and that the oCPM model reverses that posture. Treat the reversal as a reason to test incrementality, not as proof of it. The optimiser will happily learn from a signal that flatters it.

Reading view through conversions inside every window#

Measuring chatbot advertising starts with a menu. OpenAI's Measure Results page lets you choose a 7, 14 or 30 day click-through reporting window and either a 0 day or a 1 day view-through window, in any supported combination. A view-through conversion counts when someone converts within one day of seeing an eligible impression and no qualifying click takes the credit. If both qualify, the click wins. Set the view window to 1 day and the Conversions total includes view through conversions, set it to 0 day and it counts clicks only, and the two can be shown in separate columns.

Two clauses deserve a highlighter. First, changing a reporting window affects reporting only: OpenAI says it does not change optimisation, bidding or billing. The attribution window is a lens, not a lever. Second, the logic is last-touch, and OpenAI says outright that it is not multi-touch attribution. The window you pick changes the story you tell the finance director, not the hunt the optimiser runs.

Data hygiene sets the ceiling on all chatbot advertising measurement. OpenAI's conversion measurement guide asks for the pixel, the Conversions API or both, says the click reference called oppref must survive redirects and travel with server events, and asks you to reuse one event ID so a conversion sent both ways is counted once. It also says reported totals may include modelled conversions where available, that Ads Manager and your analytics will differ because of windows, time zones, consent and deduplication, and that a difference is not necessarily an error. Allow 24 to 48 hours for conversions to arrive.

Choose the window from the decision, not the dashboard#

In chatbot advertising, a window should mirror how long your customer takes to decide. A subscription app may live inside seven days. A mortgage, an insurance policy or a private clinic procedure rarely does, and 30 days is the honest choice. Report the widest click window with the 1 day view window, then re-cut every result with view-through switched off. If the answer only survives with the view credit on, your campaign may be taking credit for a decision already made.

Add the consent test. OpenAI says conversion data should be shared only after clear information to users and any consents the law requires. The pixel's automatic advanced matching detects supported customer details in forms and hashes them in the browser before sending, and OpenAI says raw details are not sent that way. For a health or finance form, decide deliberately whether that is a signal you are allowed, and willing, to share, and ask your data-protection lead before flipping the setting that OpenAI's optimisation tips recommend.

A four-week test plan for AI ads in a regulated category#

AI ads that talk back reward a patient prowl, and chatbot advertising rewards it twice over. Here is the plan we would run, in an order that protects the licence first and the budget second.

The budget floor is modest: OpenAI's campaign guide lists a minimum daily budget of 25 USD or 15 GBP, so the expense that hurts is legal and analyst time, not media. That is why the pre-flight matters more than the flight.

The same instinct, letting an algorithm find the quarry while you guard the rules, runs through our pieces on Google's AI Max migration and on bid control handed to the algorithm across Microsoft and ChatGPT. For the wider ChatGPT picture, read how ChatGPT search ads gained a self-serve front door and our measurement guide for the OpenAI pixel. Snapchat is stepping into the same undergrowth with Sponsored Snaps inside Chat, and on the Meta side our companion piece covers Meta One business tiers and competitor tracking.

If you run a regulated brand and want a chatbot advertising pilot with the approvals, the event choice and the reporting built before the first pound is spent, that is the work of folkfox PPC and paid search, informed by our experience in FinTech marketing and healthcare marketing, with the answer-engine side handled by SEO and GEO services.

Frequently asked questions#