Did you know agents have journeys too? We know — because we measure them.
Every day, hundreds of real buying queries across ChatGPT, Perplexity, Gemini, and Claude — for banks, insurers, retailers, manufacturers. Eighteen months of data.
What we found changed how we think about how brands win and lose in the agent internet. Here are eight convictions from making brands legible to machines — their claims machine-readable, their evidence current, their tools accessible.
Agents find you. They just don't cite you.
The finding that stops every client conversation: an agent can read your pages, use your information, and still name a competitor in the final answer.
It finds your page. It reads your content. And it recommends someone else.
You can be loved by customers and ignored by their agents.
The evidence exists. The agent just never gets to it. It's four layers deep, behind JavaScript, or locked in a tool.
Your content isn't yours anymore.
Every AI model already "knows" things about your brand. That knowledge lives in what we call the fossil layer — training data, cached retrievals, old web content. And most of it isn't yours. Journalists wrote it. Comparison portals wrote it. Forum posts wrote it. Their version of your story got scraped and frozen. Your positioning changed. Your pricing changed. Your products evolved. The old version of your brand can persist long after the business has moved on.
Harvard Business Review recently highlighted how AI models can misclassify established brands — in one case, Pernod Ricard's mass-market scotch Ballantine's was being positioned as a prestige product by a major model [S1]. Nobody at Pernod Ricard wrote that. Nobody approved it. It's in the fossil layer now.
If the internet already contains a version of what you're about to publish, you're not overriding the fossil layer. You're reinforcing it. Before anything ships, one test:
The only content worth producing expands the web's knowledge — and overrides the fossil layer. Today's rates. This customer's eligibility. Performance data from your own testing. Proprietary calculations. Verified specs no one else has published.
Stop producing volume. Start producing knowledge that's missing from the internet.
Agents weren't built to recommend. They were built to act.
A brand gets recommended by ChatGPT. The marketing team screenshots it, shares it in Slack, puts it in the board deck. "We're winning in AI." That's where most measurement stops.
But the agent's journey is just beginning. It now tries to act on the recommendation: get a rate, check eligibility, compare terms, start an application. And for most brands, that's where everything breaks.
Citation is not the finish line. It's where the agent's real work begins.
The agent doesn't complain. It doesn't file a ticket. It quietly moves to the next brand on the list — the one whose data was structured, whose tools were accessible, whose answer came back before yours timed out.
Being recommended is the starting gun. Most brands aren't ready for the race that follows.
Your best tools are behind glass.
In our financial-services pre-sales audits, the pattern repeats: brands ship 20–30 interactive tools per site — calculators, configurators, rate comparisons, eligibility checks. This is exactly what agents need to act: proprietary logic no competitor can replicate.
On average across these audits, fewer than 1 in 3 of those tools is even reachable by an agent. Zero are end-to-end usable without human interaction [S2].
This is where the agent's journey ends.
And most brands make it worse. Bot protection, CAPTCHAs, rate limiting — security that can't tell a scraper from a customer's agent trying to buy something.
The proof only you can offer is trapped in tools only humans can use.
"Don't Make Me Compute."
"Don't Make Me Think" changed how we design for humans. Every unnecessary click, every confusing label, every moment of friction was a cost — paid in attention. Users left.
The same principle now applies to agents. But the currency isn't attention. It's tokens. Time. Compute. The agent has a task: extract proof, verify claims, pull structured data, complete an action. Every unnecessary token between arrival and task completion is a cost. And efficiency is the selection mechanism [S3].
Every unnecessary token is a cost the agent passes on — or avoids entirely.
SEO trained brands to optimize for one system that laid out the rules: Google. Faster pages ranked higher. Core Web Vitals, time to first render, mobile performance — the playbook was clear because the judge was singular.
Agent experience follows the same logic, but the rules are no longer set by one company. They're set by physics. Cost. Energy. Every model has different token consumption rates. Every agent makes different cost-benefit calculations. And the users themselves configure their agents for efficiency — because they're the ones paying. Everything points in the same direction, but the world is orders of magnitude more complex because there's no single rulebook to follow.
Efficiency is not a technical detail. It's a selection mechanism. The path that works gets reinforced. The path that costs too much gets abandoned. Not because someone decided — because the economics made it inevitable. This is true in markets, in ecosystems, and now in the agent internet.
The attention economy competed for eyeballs. The agent economy competes on tokens. Different currency. Same law.
The richest customer conversation is happening without you.
Right now, your customer is telling their AI agent everything. Budget ceiling, timeline, required features, unacceptable trade-offs, prior brand exclusions. The full picture. The kind of briefing your best sales rep would kill for.
Then the agent arrives at your door carrying hand luggage. A thin request. "Mortgage quote, first-time buyer." You never see the context that would let you actually help.
And here's the trap: every time you feed more information into the hyperscalers — product feeds, commerce integrations, structured data for their platforms — you're transferring intelligence upstream. The model learns your category's logic. The agent gets smarter about mortgage rates, eligibility criteria, comparison frameworks. Your knowledge powers someone else's context window. And your brand becomes more interchangeable downstream — because the agent no longer needs to visit you to know what you know.
The richest customer relationship you ever had is now someone else's context window.
Don't become a dumb pipe.
The instinct after reading all of this is obvious: open up, expose your tools, make yourself accessible.
The instinct is right about access. But access without intelligence is commoditization. When an agent sends a request and your brand simply executes — returns a number, confirms a spec, processes an order — you've added nothing the agent couldn't get from any competitor with the same data.
The brands that win in the agent internet won't just receive requests. They'll meet the agent with intelligence of their own. When the agent asks for a mortgage quote, the brand's system doesn't just return a rate — it checks eligibility, factors in tomorrow's rate decision, and recommends waiting 48 hours. That's not a database lookup. That's reasoning. That's the thick server.
Agents are built to find the best possible answer. Which side do you think they come back to?
The dumb pipe looks up and returns. The thick server reasons, enriches, and responds with something the agent couldn't have assembled on its own. Same request in. Fundamentally different value out.
And here's the compounding effect: the thick server is cheaper for the agent to interact with — because the brand does the reasoning, the agent doesn't have to. And every interaction where the brand adds intelligence is context the brand now holds. The black box from conviction six starts leaking in the other direction.
The question every brand now faces is not whether to give agents access. It's whether you have any intelligence on the other side when they arrive.
The pipe gets replaced. The thick server gets chosen again.
The rails are live. Your brand isn't.
This is the elephant in the room. Agents are not just going to find, evaluate, and recommend. They are going to transact. Pay for information. Complete purchases. Book services. Settle fees. Not as a 2030 vision — on infrastructure that is already live.
All five rails are live and documented [S4] [S5] [S6] [S7]. The transaction layer is no longer theoretical. Parts of it are already executing — and the implications go far beyond checkout.
When agents can pay, every interaction becomes an economic decision. The agent calculates: do I spend thousands of tokens parsing a bloated page for free, or do I pay a fraction of a cent for a structured answer from an endpoint that already did the reasoning? That's conviction five's efficiency principle applied to money, not just compute.
The infrastructure exists. The brands that can meet it on the other end are almost nowhere.
Recommendation without transaction is a dead end with better lighting.
Your website must become an agent.
Your website today publishes and waits. The agent internet needs it to receive, reason, and respond. That means connecting your internal world — products, pricing, eligibility, market intelligence — to an AI that can meet the arriving agent as a participant, not a page.
On one side: the customer and their agent, carrying the full context of what they need. On the other: your brand and its agent, carrying the full depth of what you can offer. When both sides bring their intelligence to the table, the interaction creates value neither could produce alone. That's the agent-to-agent interface.
Your agent meets their agent. This is where business happens.
That is the thick server from conviction seven. That is the escape from the black box in conviction six. That is the endpoint the transaction rails in conviction eight are waiting for.
Your best sales rep handles a dozen conversations a day. Your brand's agent handles thousands. In every language. Around the clock. At marginal cost. And every interaction makes it smarter.
In banking, we call this agent-to-agent banking. In insurance, agent-to-agent insurance. In hospitality, in manufacturing, in retail — the pattern is the same. Your agent meets their agent. Both sides bring their full intelligence.
Not a better touchpoint.
Agent meets agent. That's the new interface.
