Explore — Week 1
Someone is shopping for a supplier, a venue, a piece of equipment, a service provider. They open ChatGPT or Perplexity or whatever agentic assistant they use, describe what they need, and ask it to find and compare options. The assistant goes and reads a handful of websites, pulls out prices and availability and specifications, and comes back with a shortlist.
At no point did a human read your homepage.
This is not a hypothetical future. It is already how a portion of research and comparison happens, and the portion is growing as agentic browsing gets built into the tools people already use for everyday tasks. The practical question for a business isn't whether this is happening — it's whether your website works when the visitor is software.
What changes when the visitor isn't human
Search engine optimisation was built around a specific behaviour: a person types a query, scans a results page, and clicks based on a headline, a snippet, maybe a star rating. Decades of practice have gone into winning that click — keywords, backlinks, meta descriptions, page speed, all tuned to influence a human's split-second judgement.
An agent doesn't do any of that. It doesn't scroll. It doesn't get swayed by a well-written headline. It fetches a page, tries to parse it, and extracts whatever facts it can find — price, availability, opening hours, specifications, how to book or buy. If your pricing lives in a PDF, or your availability is rendered by JavaScript after three seconds of loading, or your "get a quote" flow requires a human to read tone and infer intent, the agent either fails to extract what it needs or gives up and tries the next site on its list.
Ranking well for a search engine and being usable by an agent are different problems. A page can rank on Google and still be nearly unreadable to an LLM trying to extract structured facts from it. Conversely, a plain, well-structured page with clear data can be trivial for an agent to use even if it would never rank for competitive keywords.
Answer engine optimisation and llms.txt
The term people have started using for this is AEO — answer engine optimisation, as distinct from search engine optimisation. The mechanics overlap with good SEO practice (clean markup, structured data, fast pages) but the target is different: not "rank first" but "be the source the answer engine actually cites and can act on."
One concrete piece of this is llms.txt — a proposed convention, similar in spirit to robots.txt, where a site publishes a plain-language, machine-readable summary of what it offers at a predictable location. Instead of an agent scraping your marketing copy and guessing at what matters, you hand it a curated summary: who you are, what you sell, how pricing works, how to transact, where the canonical documentation lives. It's early — adoption is uneven and the standard is still settling — but the underlying idea is sound regardless of whether this exact file format wins out: give machines a clean, direct answer instead of making them infer one from a page designed to persuade a human.
The businesses that will do well here aren't necessarily the ones that rank highest today. They're the ones whose information is structured clearly enough that an agent can lift it without friction.
Being operable matters more than being found
Getting discovered by an agent is only half the problem. The other half is whether the agent can actually do anything once it's found you.
If a customer asks their assistant to compare three suppliers and book with whichever has availability next week, the agent needs to check availability and complete a booking — not just read a page that says "contact us for availability." If your booking flow depends on a phone call, or a form that emails a person who checks a calendar manually, the agent stalls. It might report back to the customer that you don't have a way to book online, which is functionally true even if you'd have happily taken the booking by phone.
This is the same discipline Pattern talks about with any AI-facing system: an agent needs an interface it can operate reliably, not one that assumes a human is on the other end to interpret ambiguity. Structured data, clear APIs, predictable forms, consistent pricing pages — these matter more for this channel than clever copy or brand voice.
Where this gets uncertain
We don't have reliable figures yet for how much commercial traffic already arrives this way, and anyone quoting a precise number for your industry is likely guessing. The standards for agent-readable data are still forming, and it's entirely possible the specific format of something like llms.txt gets superseded by something else in a year. There's a real risk of over-investing in a channel that's still finding its shape.
There's also a trust problem. If an agent is summarising your business to a customer, it's working from whatever data it can extract. If that data is stale, incomplete, or contradictory across your own pages, the agent will faithfully repeat the error. Being legible to machines raises the cost of inconsistency between your website, your pricing sheet, and your actual operations.
None of that is a reason to ignore it. It's a reason to treat this the way we'd treat any early channel — worth a proportionate investment, not a wholesale rebuild.
The other side of this
There's a symmetry here worth noticing, even if it's a topic for another post. If your customers are starting to send agents out to do research and transact on their behalf, there's nothing stopping your business from doing the same thing on the other side — using agents for the research, comparison, and first-pass qualification work that currently sits with a person. The shift isn't just about how customers find you. It's an early signal of a broader move from buying tools that support a department to running the department itself with a mix of people and agents. We'll come back to that properly in a later piece.
What to actually do
Start by checking whether your own site would survive an agent trying to use it. Ask an AI assistant to find your business, compare it to a competitor, and try to book or buy something. See where it gets stuck. Look at whether your pricing, availability, and core transactional flows exist as clean, parseable data anywhere, or whether they only exist as prose aimed at a human reader.
If you want a second set of eyes on whether your business is actually operable by the systems your customers are starting to use, get in touch — it's a fairly quick thing to assess properly.
