Insight / signal
Your website now has two customers: humans and agents
Your next customer might not arrive with a mouse. It might arrive as an agent building a shortlist.
Most websites are still built for a human with a screen.
A person searches. A person clicks. A person reads the page, compares options, fills in the form, books the demo, buys the product, downloads the guide, or gives up because the site is a mess.
That model is not disappearing. Humans still matter. Taste still matters. Trust still matters. Nobody sensible is saying your website should become an API manual with a logo stuck on top.
But there is a second customer arriving.
The agent.
Not in the sci-fi sense. Not a little robot wandering round the internet with a briefcase. The boring, practical kind: software that searches, compares, summarises, recommends, fills forms, books meetings, raises support tickets, renews subscriptions, pulls receipts, prepares buying options, and hands the result back to a person.
That changes what a commercial website needs to be.
For years, the marketing conversation has been about visibility. Can Google find you? Can humans understand you? Can you convert the visit?
There is a new question now.
Can software understand your business well enough to do something useful with it?
That is not the same as “write content for AI.” I think that framing is already dragging people into bad work. It encourages gimmicks. Secret files. Fake mentions. AI-written pages for every possible query variation. Another wave of SEO theatre with a shinier badge.
The more useful version is simpler and less glamorous.
Make your business legible.
Make the offer clear. Make the product data reliable. Make the proof findable. Make the page structure sane. Make the forms usable. Make the buying steps obvious. Make receipts and confirmation paths clean. Make public information crawlable. Use schema where it actually describes something real. Add machine-readable docs or APIs where the workflow needs them. If an agent is allowed to act, give it a safe route and a record of what happened.
That is not marketing garnish. It is commercial plumbing.
And it is becoming harder to ignore.
The strongest signal I saw this week was not another model benchmark. It was the agent commerce discussion from the Startup Ideas feed: agents will need identity, tools, inboxes, memory, wallets and receipts. Strip out the big-market language and the point is sound. If agents become part of the buying journey, the internet needs infrastructure for agents to find, evaluate, transact and report back.
That matters for every business that expects the web to bring it customers.
At the same time, OpenAI is pushing Codex beyond software teams. Their June 2 updates say Codex now has more than 5 million weekly users, that non-developers make up about 20 percent of users, and those non-developer users are growing more than three times as fast as developers. The examples are not just coding work. They cover marketers, analysts, operators, sales teams, designers, internal apps, dashboards, creative briefs, customer context, follow-ups, deal reviews, research, reports and workflow automation.
In other words: agents are not staying inside engineering.
They are moving into normal business work.
That is where this gets commercially interesting. Once agents sit inside everyday business workflows, they will start touching other companies’ websites and systems on behalf of humans.
A sales agent may research vendors before a human sees the shortlist. A procurement agent may compare suppliers. A marketing agent may check whether a partner has enough proof to cite. A customer service agent may look for documentation before opening a ticket. A founder’s assistant may compile options for a new tool, agency, venue, recruiter, consultant or software subscription.
If your business is hard for those systems to understand, you have a problem.
Not an existential one by Thursday. We can all calm down.
But a real one.
The current AI-search panic is too narrow. People keep asking how to rank in ChatGPT, or how to get into AI Overviews. Those are fair questions but they are not the whole problem.
Google’s own AI optimisation guidance is useful here because it kills some of the nonsense. Google says AI Overviews and AI Mode are still grounded in core Search systems: crawlable pages, indexing, quality, retrieval, useful content, and pages that actually help people. It also talks about browser agents inspecting visual renderings, DOM structure and accessibility trees.
That is the line worth paying attention to.
The future is not just “can a model quote your page?” It is “can an agent use your site without guessing?”
A lot of websites fail that test now.
They hide the useful information in PDFs nobody can parse cleanly. They have vague service pages with no real detail. They use forms with weird validation and no clear next step. They bury prices, eligibility, delivery terms, product specs, locations, stock, support routes and proof. They rely on JavaScript widgets that look fine to a person and confusing to a crawler or browser agent. They use generic content that says nothing an AI system could trust as distinct evidence. They have no structured data, or worse, structured data that describes fantasy rather than the page.
Then they will wonder why AI systems ignore them.
This is where the agency opportunity sits, but only if agencies do not turn it into grift.
The weak offer: “We do AI SEO and get you ranked in ChatGPT.”
That already smells bad.
The stronger offer: “We make your business easier for humans, Google, AI systems and browser agents to understand and use.”
That is less sexy. It is also much harder to fake.
For a real business, the work looks something like this. Rewrite the main service pages so they explain who the offer is for, what problem it solves, what it costs, what happens next, and what proof exists. Fix technical SEO basics so important pages can be crawled, indexed and cited. Add useful schema for products, services, organisation details, FAQs, reviews or local business data where it reflects reality. Create comparison pages that are honest and evidence-led, not thin keyword sludge. Turn buried product knowledge into clear documentation. Make quote, booking, support and checkout paths obvious to a browser agent as well as a person. Add public docs or tool endpoints for workflows where an agent genuinely needs to take action. Keep receipts, confirmations and state changes clear enough that an agent can report back without making things up.
None of this requires pretending every site needs a full agent API tomorrow.
Most do not.
A local business probably needs cleaner pages, better structured data, a sane booking flow and proof that can be cited. An ecommerce business needs product feeds, availability, delivery data, returns policy and checkout clarity. A B2B service business needs specific offers, case evidence, sector pages that are not copied rubbish, and forms that do not punish the buyer. A software company may need docs, sandboxes, MCP tools, OAuth, status pages and support workflows.
The point is not to bolt agent features onto everything.
The point is to ask what an agent would need to know or do if it were helping a buyer.
That question exposes a lot of weak marketing very quickly.
If your offer is vague to a human, it will be worse for an agent. If your proof is thin, the agent has nothing solid to cite. If your site structure is chaotic, retrieval gets noisy. If your forms are awkward, browser automation becomes brittle. If your product data is stale, recommendations become risky. If your content is generic, you look interchangeable.
This is why “agent-readable” is better than “AI-optimised.”
AI-optimised sounds like a trick.
Agent-readable sounds like a standard.
Readable means clear. Structured. Useful. Current. Actionable. Safe enough to use. It means the website is not just persuasive at first glance. It can survive being inspected by software that needs facts, routes and evidence.
That fits the way marketing is changing.
Post-agency marketing cannot just be content production. AI has made content cheap enough that more output is not a strategy. The useful work is designing the operating layer: research, positioning, proof, pages, workflows, follow-up, measurement, and the systems that keep all of it current.
Agent-readable websites sit right in the middle of that.
They are not just an SEO job. They are not just a web development job. They are not just a content job. They are the place where marketing, data, UX, technical structure and business operations meet.
Which is exactly why most companies will neglect it.
It is not glamorous. It does not make a dramatic demo. You cannot post a screenshot of “we fixed the accessibility tree and clarified the quote flow” and get the same cheap dopamine as a chatbot doing jazz hands.
But this is the stuff that decides whether AI systems can actually work with your business.
There is also a trust angle here.
If agents are going to act for customers, they need reliable information. If agents are going to act for businesses, they need safe permissions and clear audit trails. If a buying journey includes software making intermediate decisions, the business has to be explicit about what is true, what is available, what changed, and what happened.
Vague marketing copy will not cut it.
Neither will the old trick of hiding everything until the prospect books a call. That may still work in some markets, but it becomes a liability when software is trying to shortlist options before the human gets involved.
The buyer journey is not becoming fully automated overnight. People will still want conversations. They will still judge tone, taste, evidence and trust. In high-value services, humans will still make the final call.
But the research layer is already changing.
The shortlist may be built by an assistant. The first comparison may be generated. The obvious weak options may never reach the human. That means invisibility will not always look like losing a ranking. Sometimes it will look like never being included in the agent’s answer in the first place.
The practical move is not to panic.
It is to audit.
Pick one offer. Pick one buying journey. Ask: what would a human need to understand this properly? What would an AI system need to retrieve and cite it accurately? What would a browser agent need to complete the next step? What proof would make the recommendation safer? What data is missing, stale or trapped in the wrong format? Where could an agent misunderstand the business? Which actions should be available to software, and which should stay human-approved?
That is a better client conversation than “Do you have an AI strategy?”
More specific. More commercial. Less consultant cosplay.
For Foundry, this is an Agent-Readable Site Audit. Not a giant transformation deck. A focused diagnostic. Take one company. Run the site like a human buyer, a search engine, and a browser agent. Check the offer clarity, proof, crawlability, structured data, product and service information, action paths, forms, documentation, receipts, and obvious workflow endpoints. Give the client a 30-day fix plan.
For IT resellers, for example, the audit checks whether AI systems can understand their vendor categories, product evidence, buyer questions, comparison points, implementation services, support routes and quote process. That is much sharper than “AI marketing.” It ties into real buying behaviour, real channel complexity and real commercial leakage.
The businesses that win here will not be the ones chasing every AI visibility rumour.
They will be the ones making themselves easier to understand, trust and use.
By humans first.
By agents second.
And by the commercial systems that increasingly sit between the two.
Your website was built for people.
Fine.
Now make sure software can tell what the hell it is looking at.