How to appear in AI answers as an advertiser

Search ads sit beside links. An answer engine composes a response from sources instead, so there is no equivalent open slot inside the answer that a brand can simply purchase.

What determines whether a brand gets named?

The source material the engine retrieves is one input, alongside model knowledge, query handling, ranking and safety systems. A brand absent from relevant source material has less evidence to be named, but presence in a source does not guarantee retrieval or a mention.

Why not simply buy ads on each engine?

You can, and increasingly you will be able to. But each engine sells only its own surface, and each will monetize the head of the market first. That leaves the long tail of specific questions, where purchase intent is highest, unaddressed by any of them.

What can be bought today?

In the CrawlClick pilot, an approved advertiser can fund disclosed sponsored placement on eligible publisher pages. The billable event is the recorded insertion into source material. Whether an answer engine later retrieves or repeats it is measured separately and is not guaranteed.

Does disclosure reduce the effect?

That is the open empirical question, and it deserves an honest answer: we are measuring it rather than assuming. Disclosure may reduce raw surfacing, or a well-attributed sponsored claim may outperform an unsourced organic one. We would rather find out than quietly ship the covert version.

How do I know whether it worked?

Each placement carries a traceable marker. The experiment harness can query engines independently and score mentions, disclosure carry-through and citations rather than relying on a publisher’s self-report. That efficacy experiment has not yet produced a commercial go verdict.

Can I test before committing budget?

Approved participants can replay a proposed bid and targeting policy against recorded eligible traffic before enabling delivery. Replay estimates auction opportunities; it does not predict how often a resulting source passage would appear in an answer.

What should I start with?

Start with one falsifiable hypothesis: a specific destination, a narrow topic set and the answer behavior you want to measure. The invite-only pilot starts in Stripe test mode with a $100 minimum when real funding is enabled.

What does the retrieval step actually pull?

Passages, not pages, and not brands. When a question is asked, the retrieval layer pulls the fragments that best match it and the model writes with those in front of it. Retrieval is one input among several — model knowledge, query handling, ranking and safety systems all act on the result — but it is the only one a publisher or advertiser can put anything into.

That makes the passage rather than the domain the unit an advertiser can work on. A well-known brand with no passage addressing the specific question has less evidence available to be named than an obscure one that has a paragraph answering it directly. Less evidence is not the same as a worse outcome, which is exactly why the effect has to be measured rather than assumed.

Why does the long tail matter more here?

Because that is where the questions are specific enough to carry intent and numerous enough that no engine will sell them individually. “Which tool does X for a team of five with Y constraint” has a buyer attached to it and will never be a line item on a media plan. The head of the market — the category terms — will be monetised by the engines themselves, and probably first.

What cannot be bought?

The model’s judgement. A placement can put a claim into the source material; it cannot instruct the model to prefer it, and any product that offers to is describing prompt injection rather than advertising. It also cannot make a claim survive contradiction: if three other sources say something different, a disclosed sponsored passage does not outvote them, and it should not.

What are the honest limits of the measurement?

Presence is observable in a sample; reach is not observable at all. A traceable marker lets the experiment harness record that a placement appeared in an answer it asked for, and no engine publishes how often any answer was shown or to how many people. Anyone quoting impressions for an AI answer is inferring them. That is a deliberately narrower claim than the advertising industry is used to, and it is the reason the efficacy question is still an open experiment rather than a stated result.