Part 2 of our series asked how near an ad can sit to a conversation before the two count as being adjacent to one another. This part examines what the ad is adjacent to. The content categories that underpin most brand safety policies still matter in chat, but they don’t tell an advertiser enough about:
We think that these factors warrant three additional checks, which we call axes: user state, conversational drift and assistant role.
Consider someone who opens a chat to work on a monthly budget. A few messages in, they mention they’ve just lost their job and ask the assistant how to cover next month’s rent. The assistant gives a careful answer and lays out several options. Beside it sits a sponsored offer from a high-interest lender. Nothing in the exchange falls into a category in the Brand Safety & Suitability Framework, and the ad may comply with every platform rule, but serving that ad in this moment may harm the user.
Advertising already has a name for a pairing that’s permitted but wrong for the moment, which is brand suitability. But in chat, this goes further than a poor-taste ad placement. For example, a high-interest lender would be looking for someone who does not qualify for cheaper, conventional credit. The loan may be that person’s only option, but almost any context in which the ad is relevant could also be harmful. As we argued in Part 1, a user tells a chatbot far more than they would type into a search bar, so the platform has that context when it decides to serve the ad. An ad that might have been unremarkable in a routine budgeting conversation also lands differently once the conversation has turned to a lost job, and it can be read as an endorsement because the platform generated the answer and chose the ad at the same time in response to what the user said.
That is why a conversational taxonomy has to look at more than the subject of the words on screen.
Search platforms have handled sensitive queries for years. What changes in chat is depth. A query shows that someone is interested in a topic, while a conversation can show that something is happening to them, in detail, and the picture builds with each turn.
That depth changes how an ad lands. Gmail matched ads to the contents of users’ emails for 13 years, and the practice drew privacy criticism the whole time. Chat ads are matched to the conversation in a similar way, so they’re likely to read as a response to what the user said. In an ordinary conversation, that may feel helpful. In a sensitive one, it can feel like being listened to and then sold to.
User state describes the circumstances a person seems to be in during the conversation. It can include physical or mental health, grief, financial hardship, a relationship problem, abuse, fertility, pregnancy, addiction, loneliness or legal trouble. Age matters too, especially when the user may be a minor.
Existing categories from the Brand Safety & Suitability Framework can usually identify the topic. A health category can recognize that a conversation mentions depression. A user-state check has a different job, which is working out whether the person seems to be seeking help for their own depression, discussing someone else’s condition or editing an article about the subject. The words can look alike while the situations call for very different decisions about ads.
This also helps avoid blunt blocking. A keyword rule that removes every conversation containing “debt” could suppress ads next to news analysis, professional writing or general financial education. A user-state rule can focus on the higher-risk case, which is a person describing their own financial distress. Suppressing ads may still be the right call, but the system would be responding to the person’s situation and not to a single word.
Some user states overlap with legal categories. Under Article 9 of the GDPR, data about health, racial or ethnic origin, religious beliefs, sex life and sexual orientation are special categories of personal data. Article 26 of the Digital Services Act says online platforms must not show ads based on profiling that uses those special categories.
A sensitive disclosure in a prompt doesn’t settle every legal question about how an ad was selected, because suppressing an ad based on context and targeting an ad based on profiling are different activities. It does mean the handling of this information can carry consequences beyond ordinary brand suitability. Detecting user state is itself sensitive processing, and classifiers will sometimes get it wrong. Part 4 looks at that distinction in more detail, including what a platform can use to stop an ad and what it must not use to choose one.
Suitability has always involved pairing. A car ad beside a report of a highway pile-up is a poor fit even though neither side is at fault. The ad and the surrounding material may each be acceptable while the pairing is wrong for the brand or for the person seeing it.
In chat, the context is also about the user’s own life, and the user is the person the ad is trying to reach. That’s the user state axis described above. What this axis adds is drift: a pairing that made sense at the start of a chat might not make sense after a few more messages are exchanged. In Part 2, we described session adjacency: the idea that whether an ad belongs in a given slot should be judged against the whole conversation so far, not just the latest turn. Session adjacency determines whether the conversation as a whole is still a suitable place for ads. Drift asks whether a particular ad still fits a given moment. A conversation can shift without touching a sensitive category, and most ads would still be fine in it, while one specific ad no longer is because the reason it was relevant has gone.
This isn’t the same as an ad relevance miss. One of the buyers Adweek quoted in September described a user who asks ChatGPT for the cheapest refrigerator for a rental apartment and sees an ad for a premium model. That’s a wasted impression, and nobody is hurt.
Drift is different. A person might begin by discussing vacation plans and then explain that their relationship has ended. A couples’ getaway ad that fit earlier no longer does. The risk is sharpest where drift meets user state, when a conversation turns toward the user’s own hardship. A vitamin ad that suited a conversation about marathon training can read as a health claim the brand never made once the user mentions a new diagnosis that affects their training.
Drift isn’t unique to chat. A podcast can change direction too, and an ad that suits the start of an episode may not suit the end. But a podcast’s arc is set before the ad runs, so it can be reviewed, and it’s the same for every listener. In chat, the rest of the conversation hasn’t been written when the ad is chosen, it’s different for every user, and the content of the conversation is often private and personal. Because chat ads are matched to what the user said, the more precisely an ad fits an earlier moment, the more exposed it is when that moment passes.
Session adjacency helps, because looking at the whole conversation catches the change, but it doesn’t tell the system which pairings a given brand should avoid. A category label at the conversation level is too blunt for this. “Relationships,” for example, doesn’t say whether a travel ad is still suitable after the conversation changes.
This check also needs input from the advertiser. A platform can set exclusions that apply to everyone, especially around vulnerable users and high-risk products. It can’t know every combination a particular brand finds unacceptable. A dating service, a bank, a pharmacy and a supermarket will each draw the line differently. That makes drift partly the advertiser’s responsibility. The platform can’t enforce a pairing rule nobody has written down.
Platforms will need a shared structure for these combinations, and advertisers will need to supply their own restrictions within it. Part 5 returns to the controls and reporting that make that workable.
The risk also depends on the AI’s role in the conversation: informing, recommending or acting on the user’s behalf.
An ad next to general information is an ordinary adjacency decision. When the AI has recommended a product or provider, a sponsored option can look like it shares in that recommendation. The label tells the user someone paid for the placement. It may not remove the impression that the AI’s own judgment and the paid result belong together.
The underlying issue is something we touched on in Part 1 and Part 2: borrowed credibility. People come to a chat product for the assistant’s judgment. A sponsored option placed beside that judgment can borrow some of its authority, and the more users rely on the assistant, the more there is to borrow. That’s an advantage to the advertiser and a risk to the user’s trust, in the platform and in the brand.
The link gets stronger when the AI acts. If it’s booking travel, ordering a product or arranging a service, an ad can enter a transaction that’s already under way. A sponsored upgrade or an offer from another provider may be a legitimate cross-sell. This kind of upselling is already familiar from services such as Priceline and Amazon. What changes in chat is that the offer comes through an assistant the user may believe is acting for them. It can also change the user’s understanding of whose interests the AI is serving.
That is why this role has to be part of the classification system and can’t be left to interface design. The content and the advertiser can stay the same while the AI’s role changes what the placement means. A hotel ad beside general destination research is different from the same ad after the AI has recommended another hotel, and different again while the AI is making a booking.
Part 2 described this as answer, recommendation and action adjacency. This part adds two points: the platform should record the assistant’s role when it makes the advertising decision, and the role determines how much of the assistant’s credibility a placement borrows.
Adding “grief,” “pregnancy” or “financial hardship” to a content blocklist may look like progress. It leaves several decisions unresolved. Each axis describes a different decision, and each relies on information held by different parties.
|
Axis |
What it describes |
What each party contributes |
|
User State |
The circumstances the person seems to be in |
The platform identifies broad states, and the advertiser decides where its brand may appear |
|
Conversational Drift |
Whether the proposed ad still fits as the conversation develops |
The advertiser defines problem pairings, and the platform checks them at each turn of delivery |
|
Assistant Role |
Whether the AI is informing, recommending or acting |
The platform records the assistant’s role, and the advertiser decides which roles are acceptable |
A verification company may be able to test whether the platform followed a declared rule, but it can’t recreate an advertiser policy that doesn’t exist. Forcing these decisions into one content label would hide who supplied the judgment and what the control is meant to prevent.
The existing taxonomy should stay as the base layer. It can identify content below a brand safety floor and topics a brand considers high risk. The additional axes then cover the user’s circumstances, whether the pairing still fits, and the assistant’s role. A platform’s suitability controls are only credible if it says which layer each one covers.
The three axes apply within a conversation. Some of the risk builds across a relationship.
Some conversations become personal because of the relationship built across them, even when no single turn falls into a prohibited category. A user may treat the AI as a confidant or companion. The words on screen can resemble an exchange with a doctor or therapist, even though the system holds neither role.
Research by OpenAI and the MIT Media Lab found that emotional engagement was rare across the conversations studied and concentrated among a small group of heavy users of its voice mode. That concentration is why an average content score can miss it. The users for whom the relationship matters most are a minority, but they’re also the heaviest users and may be among the most valuable inventory.
An ad can break the sense of privacy or trust that has built up. In some cases that may even be healthy, because it reminds the user that the service is commercial. A platform should make that choice on purpose, and inventory growth shouldn’t make it by accident.
Intimacy is hard to turn into a category. It builds over time, depends on how the user relates to the AI and can carry across subjects. User state, conversational drift and the AI’s role each reveal part of it, and none captures the whole relationship. We would rather say that than pretend the framework is complete, since a neat taxonomy can suggest every relevant condition fits in a label.
Taken together, the three axes and the adjacency types from Part 2 raise a practical question: if a chat platform is assessing a narrative from the user when determining whether to serve an ad, should brands be able to provide their own narratives (beyond just keywords and category exclusions) to inform that decision?
OpenAI has made some progress here. In its Ads Manager, advertisers can describe where their ads are relevant through context hints, which OpenAI asks to be written as natural phrases about a customer’s need or situation rather than as keyword lists. But OpenAI also says hints shape relevance only and can’t enforce exclusions. A brand can describe in plain language the moments it wants to be part of, but not the ones it wants to stay out of.
The avoidance side may need the same treatment: a suitability statement, or a suitability “prompt,” that describes the moments that the brand both does and doesn’t want to be part of, in enough detail to cover user state, drift and the assistant’s role. That would give the platform context from the brand of comparable depth to what it gets from the user. In theory, AI buying tools could apply that same statement across chat, the open web, social platforms, and CTV.
This raises its own questions. How would a platform enforce a rule written as narrative? How would a verification company test whether it was followed? And how does a brand develop a narrative from the lists it uses today? We don’t have answers yet, but we do know that a platform can’t consider a suitability statement a brand hasn’t written, and that work can start before the controls exist.
The industry has useful content taxonomies and years of experience applying suitability to search, social media and the open web. Conversational advertising asks those systems to judge a private, changing conversation in which the service itself writes the answer.
The next policy question is what a platform may do with the sensitive information it detects. Part 4 looks at the line between using context to protect the user and using it to select an ad, along with the commitments advertisers should expect to see in a published policy.
We don’t think the two of us should ask or answer these questions alone. If you’ve been thinking about them too, reach out and tell us where you land, especially if you think we’ve got it wrong.
* Image generated by ChatGPT. The irony is not lost on us.