AI Search Intelligence Brief
The developments worth paying attention to — and what they tell us about where digital discovery is going.
Hillary Bassett Ross — Founder, Kinetic Pencil
Hillary Bassett Ross is an AI search and web optimization strategist with 20 years of experience across content strategy, SEO, UX, analytics and digital optimization for companies including Microsoft, Cisco and Liberty Mutual. She is the founder of Kinetic Pencil, where she helps marketing leaders understand what changes when AI becomes a primary layer between their brand and their customer—and how content, websites and digital ecosystems need to change with it.
Something clicked for me this week.
We’ve spent a lot of time talking about AI as a new way people find things. But the more interesting shift is what happens after the machine finds it.
Because the machine isn’t just retrieving anymore. It’s recommending, buying, doing the work.
- retrieving
- recommending
- buying
- doing the work
And across a pretty random mix of announcements and research, the same question kept surfacing:
When do we let the machine take over, and what has to be true first?
01The big one
The next phase of AI search is delegation
For a couple of years, the big question has been: Will people use AI to find things?
That’s starting to feel like the wrong question.
The better one: What will people let AI do once it finds them?
This week showed that boundary from two directions.
Consumers are happy to let AI gather grocery prices and options. Letting it run the whole shop? Much less so. Until the benefit gets concrete, like saving 30 minutes. Then willingness jumps.
Enterprises flip the pattern. In network operations, companies are handing machines serious autonomy in high-stakes environments, because the volume and complexity have outgrown what humans can manage by hand.
So delegation isn’t really about trust.
It’s math.
What the machine saves you on one side.
What it costs you if it gets it wrong on the other.
Somewhere in between sits what I’m starting to call the Delegation Threshold: the point where we decide the machine can take the next step for us.
AI Search IntelligenceKinetic Pencil
And that threshold moves. That’s the Delegation Frontier.
Knowing where it sits, for your customers and inside your own operations, may turn out to be one of the most useful things an organization can measure.
02The signal
Organizations are quietly being rebuilt for machine decision-making
If machines are going to do more than retrieve, organizations have to become understandable to them at a much deeper level.
That’s what I kept seeing underneath this week’s news.
Mastercard is building infrastructure so agents can read merchant catalogs, pricing and fulfillment, then transact within rules the merchant and customer set.
Banks are working on a different layer: proving the human actually meant for the machine to act.
EY is telling grocery suppliers to organize around consumer needs instead of product categories.
And Teradata is making the case from inside the enterprise: give an agent structured context about the organization, and it may need far less computing work to get the same job done.
Separate announcements. I don’t think they’re separate developments.
They’re all building toward a machine being able to answer questions like:
- What can this organization actually do?
- What proves it?
- Am I allowed to act on it?
That’s a lot bigger than getting an LLM to mention your brand.
It’s making the organization itself legible enough for a machine to reason about. Which keeps strengthening my Digital BOM hypothesis: the organizations best set up for agents will be the ones whose capabilities, evidence and permissions a machine can reliably resolve.
Machine readability isn’t just a discovery advantage anymore.
It might be operational infrastructure.
03The data point
Under 1% → 26%
In February, Anthropic says Claude could “lead” less than 1% of its measured AI R&D work (meaning take most of a task end-to-end from a high-level prompt, with a human supervising).
By August: 26%.
Real caveats here. Anthropic is measuring its own models, inside its own company, with no standard cross-industry method yet.
But honestly, the number interests me less than the question behind it.
Most organizations are still asking:How many employees use AI?
Anthropic is asking:How much of the work does the machine actually own?
Usage tells you someone opened the tool.
Autonomy tells you the operating model changed.AI Search IntelligenceKinetic Pencil
04Connecting the dots
Here’s where it gets interesting.
If machines are moving from finding → deciding → acting, then “AI visibility” is only the first layer of something much bigger.
- finding
- deciding
- acting
Google’s new multimodal reporting in Search Console adds a twist: sometimes the journey doesn’t start with words at all. Someone shows the machine an object, and the machine works out what to ask. Then it still has to figure out whether anyone can actually meet that need, and whether it’s allowed to act.
Meanwhile, Gartner’s consumer research adds a tension I don’t want to smooth over. As synthetic content gets easier to produce, people say they want more evidence from real humans, not less.
So the answer isn’t more machine-readable content.
It’s machine-readable evidence. Original research. Documented outcomes. Proof.
What a skeptical human wants and what an evidence-seeking machine needs are turning out to be the same thing.
AI Search IntelligenceKinetic Pencil
The stack I’m watching looks like this:
- Semantic
- Structural
- Evidence
- Perceptual
- Operational
- Permission legibility
- Decision provenance
Each layer answers something the machine has to resolve before it can confidently go from finding an organization to acting on someone’s behalf.
That’s a much bigger discipline than AI SEO.
05On my radar
Things I’m not ready to call yet.
Sponsored agents. OpenAI is testing business-sponsored agents you can talk to after clicking an ad. That could turn advertising from ad → landing page into ad → conversation → recommendation. I’m most curious how anyone will measure it.
Autonomy metrics. Anthropic proposed standard ways to measure AI autonomy. Watching whether that escapes the frontier labs and becomes useful for enterprise transformation.
Permission architecture. Cloudflare now separates machine permissions for search, training and agents. Machine access is clearly not going to stay a simple allow/block switch.
The capability/cost curve. Individual model releases matter less to me than the pattern: more autonomous work, cheaper. That keeps shifting where delegation makes sense.
06The question I’m left with
For a long time, the job was getting found.
- Rank.
- Earn the click.
- Convert on the website.
But look how much of that journey can happen somewhere else now. A machine searches and compares. An agent recommends and buys.
The website might never be the center of it.
So the question changes. It’s not Can people find us? It’s not even Can AI find us?
It’s this:
When a machine has to decide on someone’s behalf, does it know enough about you, trust your evidence, and have permission to choose you?
That’s the shift I’m watching now.