What is hallucination liability?

The exposure exists whether you advertise or not

This is the part that surprises people. An assistant will describe your products, quote your prices and summarise your policies because users ask it to, and it will do so regardless of any commercial relationship with you.

When it gets that wrong, the consequences land on you. A customer told your warranty covers something it does not arrives expecting it. One told the wrong price arrives expecting that price. In some jurisdictions and some categories the exposure is not merely reputational.

Opting out of AI advertising does not opt you out of this. The only thing that changes is whether you have any visibility into it.

Where the errors come from

• Stale training data. Prices and policies from two years ago, stated with complete confidence.
• Contradictory sources. Where your site, a reseller and a review disagree, the model resolves the conflict by generating something plausible.
• Plausible interpolation. Asked about a product configuration that does not exist, a model will often describe one rather than say it does not know.
• Competitor confusion. Attributes of a similarly-positioned brand attached to yours.
• Ambiguity in your own material. Anything a careful human could misread, a model will.

What actually reduces it

The single highest-leverage action is consistency. A model reconciling agreeing sources reproduces them; one reconciling disagreeing sources invents. Making your price, policy and product facts identical everywhere they appear — your site, your resellers, your listings, your documentation — removes the ambiguity errors feed on.

After that: publish facts in structured, retrievable form so they are available at query time rather than only in training; state explicitly what your product does not do, which prevents plausible interpolation; and monitor, using the same fixed-prompt-set sampling as AI share of voice, extended to record accuracy rather than only presence.

The branded agent case

Where you operate your own branded agent, the liability is direct rather than second-hand. Anything it says is the brand saying it, with the contractual and regulatory weight that carries.

The mitigation is architectural: constrain the agent to retrieved facts for anything commercially binding — price, availability, terms, eligibility — rather than allowing it to generate them, and log what it said to whom. Generative fluency is appropriate for explanation and inappropriate for commitments.