Advertising is moving inside AI assistants — and it quietly breaks the keyword, the click, and the attribution model all at once.

For twenty-five years, paid search rested on one clean assumption: a person types what they want, advertisers bid to appear next to that intent, and the click hands the visitor over. Everything downstream — bid strategy, landing pages, quality scores, last-touch reporting — was built on top of that single mechanic.

That mechanic is now optional.

Ads are appearing inside AI assistants, and the assistant doesn’t work from a query. It works from a conversation. It has already asked follow-up questions, already compared three options, already narrowed the field before any brand gets a look in. By the time an ad surfaces, the buyer isn’t at the top of the funnel. They’re most of the way down it — and you weren’t in the room for the part that mattered.

This is not a new placement. It’s a new position in the journey.

The keyword stops being the unit of intent

Traditional search ads are matched to a string. AI ads are matched to a situation.

OpenAI has described its ChatGPT ad selection as drawing on the context and intent of the live conversation, the advertiser’s landing page and copy, hints the advertiser supplies, and — where personalisation is switched on — signals from the user’s broader history with the assistant. Read that list again from a media-buying perspective. Almost none of it is a keyword.

The practical consequence: your targeting surface is no longer a list of terms you can pull from a planner. It’s the shape of the problem your buyer is describing in their own words, three turns into a conversation, often without ever naming your category.

Marketers who have spent a decade optimising match types are about to discover the skill doesn’t transfer cleanly. The new equivalent of keyword research is problem research — mapping the questions, workarounds, and half-formed use cases that lead someone toward your product before they know the product exists.

Creative got cheap. Judgment didn’t.

The same technology is compressing the production side. Generative tools now run across the whole campaign lifecycle — concepting, variant generation, cross-channel testing. Google Ads will already spin up headline and description permutations and let the system find the winning combination.

When producing an asset costs almost nothing, the asset stops being the scarce thing. Deciding which ideas deserve to exist becomes the scarce thing.

That’s a genuine reallocation of where marketing talent earns its keep. Less time assembling the forty-third variant. More time on the questions no model can answer for you: what does this brand actually stand for, which of these ideas is worth scaling, and does this output sound like us or like everyone else.

Cheap production is only an advantage if your competitors are worse at choosing. Most of them will not be.

The click becomes a receipt, not a signal

Here is where the finance conversation gets uncomfortable.

Ads inside assistants are already being sold on both CPM and CPC. But the click is no longer where the influence happens — it’s where the influence gets recorded, if it gets recorded at all.

Picture the real sequence. A buyer asks the assistant to explain a category. Asks which approach fits a mid-market team. Asks about integration risk. Asks for a shortlist. Four moments of genuine persuasion, none of them visible in your analytics. Then one referral click lands, and your attribution model gives it full credit like it’s a paid social ad.

Sometimes there won’t even be that. As agent-led transactions mature, the purchase can complete inside the assistant. No click, no session, no landing page — a conversion with no observable path.

So the question the industry now has to answer: does the AI platform get paid for delivering the click, or for shaping the decision that produced the purchase? Those are very different pricing models, and whichever wins will reset acquisition economics for everyone.

Practical read: if you are still grading channels on last-click, you are about to systematically underfund the thing that’s actually working.

Your product feed is now a brand asset

This is the least glamorous section and probably the most actionable.

When an assistant recommends products, it reasons over structured data — pricing, specs, availability, attributes. Amazon’s shopping assistant, built on its own catalogue plus web information, answers product questions, compares options, and makes recommendations inside the normal shopping flow. Startup Gravity is reportedly working with Best Buy and Target on agent-to-agent advertising, where AI agents trade recommendations using live advertiser catalogues.

In that world, a thin, stale, or inconsistent feed doesn’t just cost you a shopping placement. It costs you eligibility to be considered at all.

The feed has quietly stopped being an e-commerce compliance task and become an input to AI visibility. For B2B, the same logic applies to anything machine-readable about you: specification pages, pricing tiers, integration lists, comparison content, documentation. If a model can’t find a clear answer about what you do and who you’re for, it will confidently use someone else’s version.

You no longer control your own description

Brand safety used to mean controlling adjacency — which page, which content, which context. That definition is now too narrow.

An assistant can summarise your product inaccurately in the same breath it shows your ad. A May 2026 analysis of more than 55,000 Google searches examined over 98,000 claims in AI-generated overviews and found roughly one in nine unsupported by the sources cited. The study also found that the quality of a source and the fidelity of the claim drawn from it were largely independent — a reputable citation is no guarantee the summary is right.

Eleven percent is not a rounding error when the output is the first and possibly only description of your product a buyer ever reads.

The new brand-safety practice is monitoring how assistants describe you, across the questions your buyers actually ask, with the same seriousness you’d apply to a review site or an analyst report. Most teams have no process for this at all.

When agents negotiate with agents

The furthest edge of this is agents transacting on behalf of businesses — and here the governance question stops being theoretical.

A 2026 paper on auditable agents makes a sharp distinction: accountability is impossible without auditability. Its argument identifies what a system needs before anyone can meaningfully be held responsible — the ability to reconstruct actions, coverage across the full lifecycle, checkable policies, clear attribution of responsibility, and evidence that hasn’t been tampered with.

Apply that to advertising. If an agent recommended a competitor over you, can anyone reconstruct why? If an agent misdescribed your product and a customer acted on it, who carries the liability — the platform, the advertiser, or the model provider?

Nobody has a settled answer. Which means AI governance is now a commercial function, not just a technical one. The organisations with real oversight frameworks will be able to scale into these channels. The ones without will either move too slowly or move recklessly, and both are expensive.

What to do about it this quarter

1. Audit your machine-readable surface. Feeds, spec pages, pricing, comparison content. Assume a model is your most important reader.
2. Ask the assistants about yourself. Run the twenty questions your buyers actually ask. Log how you’re described, what gets recommended instead, and what’s simply wrong.
3. Rebuild targeting around problems, not terms. Interview customers about how they described the problem before they knew the category name. That language is your new targeting input.
4. Fix attribution before you scale spend. Introduce incrementality testing and self-reported attribution now, while the volume is small enough to learn from.
5. Move creative effort upstream. Let the tools handle variants. Spend the recovered hours on positioning, brand voice, and deciding what’s worth making.
6. Give AI governance an owner. Not a policy document — a named person who monitors representation, approves agent-led placements, and can explain a decision when someone asks.

Search advertising rewarded marketers who understood what people typed. The next channel rewards those who understand what people are trying to work out — and who make sure the systems doing the explaining have accurate, well-structured, unambiguous information to work from.

The click isn’t disappearing. It’s just stopped being the interesting part.

Working out how this changes your demand gen motion? That’s the conversation we’re having with clients right now.