AI search is becoming an agent layer. Your website is not ready.

The useful Google I/O story is not another model.

It is that search is starting to behave less like a list of links and more like an operator.

That sounds like a small distinction. It is not.

Google says AI Mode has passed a billion monthly active users globally. AI Mode queries have more than doubled every quarter since launch. In the U.S., the average AI Mode search is three times the length of a traditional query. More than one in six searches now use voice or images. Planning-related queries have grown faster than AI Mode overall.

Strip out the launch polish and the behaviour is clear: people are not just typing keywords anymore. They are asking for help deciding, comparing, planning, and doing.

Then look at what Google actually announced around it.

Search agents that monitor topics in the background. AI Mode connecting to Gmail, Photos, Calendar. Universal Cart, where shopping can start in Search, Gemini, YouTube or Gmail and carry across surfaces. A Universal Commerce Protocol for cleaner checkout. And Gemini Spark — described as a 24/7 personal AI agent that can work under direction, ask before major actions, and eventually handle sub-agents and authorised payments.

Some of that is early. Some will land slower than it looked on stage. Demos are, as a rule, lying little goblins.

But the direction matters.

Search is not just becoming more conversational. It is becoming more agentic.

For businesses, that changes the marketing problem.

The lazy version of this conversation is already everywhere. How do we rank in AI answers? How do we get ChatGPT to mention us? Do we need an llms.txt file? Should we publish 500 AI-generated question pages?

That is the wrong frame.

Google’s own guidance was blunt about this a few days ago. AI Overviews and AI Mode still depend on core Search systems: crawlability, indexing, quality, useful content, and retrieval. The system searches around the user’s real intent, pulls from the index, and builds answers from sources it trusts enough to retrieve.

It also points at browser agents reading the visual page, the DOM, and accessibility trees.

Translation: this is not a magic new SEO game. It is the same old trust and usefulness problem, with a more demanding reader.

That reader may be a human. It may be a model. It may be a browser agent trying to work out what your business does, whether your offer fits, how much things cost, whether you are credible, what the next step is, and whether it can complete part of the job for a user who asked it to help.

Most business websites are not built for that.

They have vague claims, hidden detail, weak proof, fluffy service pages, broken forms, thin case studies, unclear pricing, inaccessible components, messy product data, and blog posts that say the same thing as everyone else with slightly different headings.

A human can sometimes fight through that mess.

An AI system will summarise it badly, ignore it, or route to a competitor with clearer evidence.

This is why I keep saying the next SEO problem is operational, not editorial.

Content still matters. Technical SEO still matters. Fast pages, clean indexation, sensible structure, media, local data, product feeds — all of it still matters.

But the deeper job is making the business legible.

Can an AI system understand what you sell? Can it tell who it is for? Can it see proof that is specific rather than beige? Can it compare your offer without hallucinating half the details? Can it find the price, process, constraints, outcomes, location, or next step? Can a browser agent actually use the page without getting trapped in some JavaScript nonsense or an inaccessible form?

That is a different level of marketing work.

It is not “write more blogs”. It is not “sprinkle AI keywords around”. It is not “publish a thought leadership piece because the calendar says Thursday”.

It is building an evidence layer around the business.

For a local service company, that might mean clean service pages, proper local proof, before-and-after examples, reviews with context, clear service areas, pricing guidance, FAQs that answer real buyer friction, and forms that a human or agent can actually use.

For a B2B firm, it might mean sharper offer pages, named use cases, real implementation detail, comparison pages that do not read like a legal team held them hostage, founder notes, better case evidence, and a content library that proves expertise rather than performing it.

For ecommerce, it means product data, availability, compatibility, returns, delivery, payment options, feeds, reviews, comparison detail, and checkout paths that do not crumble when a non-human browser inspects them.

This is where the post-agency opportunity sits.

The old agency reflex was to sell output. Blogs, ads, landing pages, emails, redesigns. Some of that is still useful. A lot of it was always theatre.

The new work is closer to an operating layer.

Make the business findable. Make it understandable. Make it trustworthy. Make it quotable. Make it easy to act on. Keep it updated as the business changes.

That is not as glamorous as promising to “win AI search” in 30 days. Good. Those promises usually smell funny.

It is also more defensible.

Because if AI systems are sitting between buyers and businesses, the companies with clearer evidence and cleaner action paths have an advantage. Not because they hacked the model. Because they gave it less room to misunderstand them.

So the useful question for a business owner is not, “How do I get AI to mention me?”

It is:

“If an AI agent looked at our business today, would it understand us well enough to recommend us, compare us accurately, and help someone take the next step?”

Most businesses will not like the answer.

That is where the work starts.