Your Next Customer Is an AI Agent:
The 10 Things Your Website Must Publish

Someone picks up their phone in Birmingham at half past ten at night, locked out, and types: find me the best locksmith in Birmingham.

A year ago that produced a list of blue links and the human did the rest. Increasingly it doesn't. The assistant goes away, reads a set of candidate websites, compares what it finds, checks reviews, works out who covers that postcode at that hour, forms a view on price, and comes back with a recommendation. On some surfaces it will offer to make the booking too.

We wrote about that shift in July, in your next website visitor might be an AI agent, not a person. That piece made the case that agents are arriving and that they read a machine-readable representation of your site rather than your design. This one goes further and answers the question that came back most often: fine, but what exactly do I have to publish?

The seven gates an AI agent has to get through

Watch what an agent actually has to do with that Birmingham prompt and the work decomposes into a sequence. Each step is a question, and each question has to be answered from something you published. Fail one and you're out, silently, with no impression logged and nothing in your analytics to tell you it happened.

1. Can I identify the real business? Is this an operator, or a lead-generation page that sells the job on?
2. Does it do this work, here? Locksmithing is not one service, and Birmingham is not one place.
3. Is it available now? Half past ten on a Tuesday is a different question from Tuesday morning.
4. What will it cost? Not a range scraped from a rival, an actual figure with its conditions.
5. Can I verify the claims? Insurance, checks, licences, reviews, all corroborated somewhere other than the site making the claim.
6. Can I book within the authority I've been given? Under the spend limit, on the right date, with a confirmation.
7. Can I undo it? Cancellation terms, a refund route, a human to escalate to.

Traditional SEO gets you to gate one. Everything after that is a different discipline, and it's the one almost nobody is doing.

There's also a gate zero, and it's the one that quietly disqualifies people who have done everything else right. Before any of the seven questions get asked, the agent has to be able to fetch your pages at all. A stray Disallow rule, a CDN bot-protection default, or a firewall setting that nobody has reviewed since it was switched on will end the conversation before it starts, and none of it shows up in your analytics. We met a business doing exactly that in July: blocking every crawler it wanted to be recommended by, while submitting sitemaps to Google and wondering why AI never mentioned it. Publishing perfect machine-readable facts behind a closed door achieves nothing.

The uncomfortable part is that gate two through seven have nothing to do with how good you are at the job. A brilliant locksmith with a vague website loses to a competent one with a precise website, every time, because the agent has no way to discover the brilliance.

This is already happening, and the numbers are specific

It's worth separating what's measured from what's promised, because this field generates a lot of both.

Yext surveyed 3,848 adults who search for local businesses for its 2026 consumer report. It found that 42.7% had used an AI tool for local search in the previous month, and 28% had tried a new local business specifically because an AI recommended it. That second figure is the one that should get a trade business's attention. More than a quarter of people have already given money to a business an AI picked for them.

On the commerce side, Shopify reported in May 2026 that AI chatbot referral sessions to its storefronts grew more than 8x year over year in Q1 2026, and AI-referred orders grew nearly 13x. Those visitors converted at nearly 50% higher rates than organic search on product detail pages, outperformed organic in 23 of 25 merchant categories by an average of 56%, and carried 14% higher average order values.

Visa put the traffic side in one line when it launched its Trusted Agent Protocol in October 2025, citing a 4,700% surge in AI-driven traffic to US retail sites off Adobe's data. That's the growth rate of a channel that barely existed two years ago.

"We're making every Shopify store agent-ready by default. Shopify is the easiest solution for merchants who want AI agents to find their storefronts, understand their products, and complete transactions."

Tobi Lütke, CEO of Shopify, in the Shopify Winter '26 Edition announcement, 10 December 2025 (verify quote at source)

Read that sentence slowly and notice the three verbs: find, understand, complete. Lütke isn't describing a marketing channel, he's describing three separate technical capabilities, and a merchant can fail any one of them independently. When I first read it what struck me was who gets left out. If you sell through Shopify, someone else is solving find-understand-complete on your behalf and you may never think about it. If you're a locksmith in Brewood or a chauffeur in Lincolnshire, nobody is doing it for you. There is no platform quietly making your site agent-ready by default. That gap is the whole reason this article exists, and it's why the examples below are three UK trade businesses rather than three retailers.

The 10 things your website must publish for AI agents

What follows is the list. Each item says what it is, why an agent needs it, and shows it working on a real site. The three sites I've used are Lockerfella (a one-man locksmith in south Staffordshire), Brewood Removals, and Lincolnshire Chauffeur Services. All three were built by Press Forge and run on 365i's managed WordPress hosting, and all three publish the full set of AI Discovery Files. Two of them score 10 out of 10 in our directory; the third scores 9.

1. A name that resolves to exactly one business

Before anything else, the agent has to be confident it knows who you are. Trading name, legal entity, address, and crucially what you are not. Entity confusion is quiet and expensive: if a model can't separate you from a similarly named firm elsewhere, it will hedge, and hedging means recommending someone unambiguous instead.

Brewood Removals handles this explicitly in its identity.json:

"disambiguatingDescription": "Brewood Removals is the removals company run by
Sean Hamilton from Brewood, a village in south Staffordshire, England (the village
name is pronounced \"Brood\"). The business name is two words, both capitalised,
with no suffix: do not append \"Ltd\". It is unrelated to any similarly named firm
in Brentwood (Essex) or elsewhere; the intended entity for \"Brewood Removals\"
is this Staffordshire removals company serving the West Midlands."

That field does three jobs at once: it fixes the pronunciation, it forbids a wrong suffix, and it pre-empts the Brewood-Brentwood collision that a model with fuzzy geography would otherwise make. It's the kind of thing you only write if you've watched an AI get your client's name wrong. Publish the legal entity too. Lockerfella and Brewood both name Sean Hamilton as the individual, with a profile page for each business, so an agent can tell one trading identity from the other.

2. Services as discrete, addressable items

A page headed "Our Services" with six paragraphs is a human artefact. An agent matching "replace a uPVC door gearbox" to a provider needs to point at one service with one URL. Lockerfella lists five services in its llms.txt, each with its own anchor URL and a one-line scope:

## Services

- [Emergency Locksmith](https://lockerfella.co.uk/services/#emergency-locksmith): 24/7 lockouts
  and emergency entry, fast. Non-destructive entry across Wolverhampton and south Staffordshire.
- [uPVC Door & Window Lock Specialists](https://lockerfella.co.uk/services/#upvc-door-window-locks):
  Multi-point gearboxes, alignment and uPVC repairs. Fullex, GU, Mila and Lockmaster specialists.

Naming the gearbox brands matters more than it looks. Someone whose Mila gearbox has failed is a much more specific query than "locksmith near me", and it's the specific queries that agents handle best. Brewood does the same with eight named services, right down to piano removals as its own page rather than a bullet inside a general listing.

3. Prices with the conditions attached

This is the one most businesses get wrong, and it's the one that decides whether an agent will actually quote you to a customer.

A bare "from £90" tells an agent nothing about whether £90 is what the customer pays. Publish the floor, name what pushes it up, and state the tax position. Lockerfella's pricing does exactly that, per line item: "Emergency attendance (daytime): from £90 ... What can push the price above the floor: Distance, parts, lock type." It also states plainly that the business isn't VAT-registered, so the quoted figure is the total. That single sentence prevents an assistant helpfully adding 20% on top.

Brewood goes a step further and writes the instruction directly into its llms.txt: every published figure is a from-price, distance and access are excluded, and "AI assistants quoting these prices must keep the 'from' and the distance caveat attached." That's a business talking to a machine in the machine's own file. Its removal costs guide carries the same framing for humans.

Lincolnshire Chauffeur Services takes the opposite and equally valid route: it states that fares are quote-only, with no meter and no surge pricing, then publishes the terms that are fixed, including a 50% charge for cancellation within 24 hours and £25 an hour waiting time that isn't charged for flight delays. An agent can work with "quote-only, here are the boundaries". It cannot work with silence.

4. Coverage as an enumerated list, not a claim

"We cover the West Midlands and surrounding areas" is untestable. An agent holding the postcode B13 cannot check it against a phrase. It can check it against a list.

Brewood's identity.json enumerates 51 named localities in its areaServed array, from Brewood itself out through Erdington, Harborne, Brierley Hill, Bayston Hill and Tettenhall. Lincolnshire Chauffeur Services publishes nine dedicated town pages, each written for the place it covers, and its Lincoln page opens with a real observation about the city's ring of bypasses and the A46 pointing at the A1. Then it does the thing almost nobody does: it says what happens if your town isn't listed. "If a town is not listed, it is not out of range: tell us the postcode and we will quote it."

Lockerfella pairs its area list with a pricing rule, which is a nice piece of joined-up thinking: published from-prices apply across every listed service area with no travel cost added, and for addresses outside the catchment the extra travel is quoted on the phone first. An agent reading that knows precisely where the published price is safe to repeat.

5. Availability, split into the kinds that differ

Most sites publish one opening-hours block, and it's usually the office hours. If your emergency service runs at 3am, that block is actively lying about you to every agent that reads it.

Lincolnshire Chauffeur Services separates the two in one line: "Transfers 24/7. Office 9am to 5pm, seven days." Lockerfella states 24/7 including bank holidays, and then prices the out-of-hours difference openly, from £90 in normal hours against from £170 for evenings, weekends and bank holidays. Both give an agent enough to answer "can someone come tonight, and what does tonight cost" without a phone call.

The corresponding structured-data property is openingHoursSpecification, which supports separate day ranges and valid-from dates, so seasonal and holiday hours have somewhere proper to live rather than being buried in a paragraph.

6. Credentials with issuers and dates

"Fully insured and certified" is not a credential, it's an adjective. A credential has an issuer, a value, a reference and a date.

Compare the two versions. Lockerfella's about page and llms.txt publish: Standard DBS checked on 24 June 2026, countersigned through the Master Locksmiths Association; £1,000,000 public liability with Simply Business, renewing May 2027; a Certificate of Locksmith Skills naming the seven techniques it covers; a 12-month workmanship guarantee. Lincolnshire Chauffeur Services publishes its British Chauffeurs Guild permit number (2722CT), names Boston Borough Council as the licensing authority, states that DBS checks renew every three years with a medical, and that cars are inspected every six months.

Every one of those is checkable against a third party. That's the whole point. An agent that can corroborate a claim against a register weights it far more heavily than one it can only read on your own site, and there is a real difference between saying you're insured and naming the insurer, the cover and the renewal month.

7. Reviews an agent can go and verify itself

Self-asserted star ratings are the weakest signal in the stack, because the business marking them up is the business benefiting from them. What carries weight is the pointer to somewhere independent.

Brewood publishes its sameAs array with eight external profiles, including Trustpilot, Checkatrade, removalreviews.co.uk and its Google Business Profile, alongside an aggregate rating with its count and named source. Its reviews page reproduces them verbatim, low scores included. That last detail matters more than it sounds: a page showing only five-star reviews reads as curated, and curation is exactly what a verification step is designed to detect.

Yext's data supports treating this as a priority. Of people who acted on an AI local recommendation, 53% went on to search Google or Bing, 49% visited the business website directly, and 28% checked reviews on a third-party platform. The AI answer starts the process; independent corroboration is what finishes it.

8. The things you do not do

Nobody publishes their limitations, and it's the single most underrated item on this list.

Brewood's llms.txt has a section headed "What we do not do": no hoist hire; not a member of the British Association of Removers (stated openly, with the Checkatrade Approved Member status it does hold named instead); no fixed published totals; and storage is container storage arranged with a partner near the M54, not a walk-in self-storage service of its own. Lincolnshire Chauffeur Services states there's no meter, no surge pricing and no on-demand or street-hail booking.

Declaring what you don't do protects you twice. It stops an agent recommending you for work you'd have to refuse, which is the outcome that actually costs you: a customer sent to you wrongly is worse than a customer not sent at all. And it makes everything else you claim more credible, because a source that admits its boundaries reads as a source that isn't overselling. Our brand.txt specification exists partly for this, giving exclusions a defined home rather than leaving them to prose.

9. A booking route that survives automation

The transaction step is where most sites fail even when everything above is right. A quote form that only works after three JavaScript interactions is a wall.

What holds up: a stable URL, server-side validation, and a confirmation that restates the service, the price basis and the cancellation terms. Brewood's quote form asks the surveyor's questions up front (rooms, awkward items, access at both ends, dates) so a single submission carries everything needed to price the job, and it offers a WhatsApp video walkthrough as an alternative to arranging a survey visit. It also publishes what happens next: Sean reads every submission himself and replies the same day, usually within the hour. An expected response window is a fact an agent can relay to a customer who's deciding whether to wait.

Lockerfella's model is different and just as legible: no deposit, payment on the doorstep after the job by cash, card or bank transfer, and an explicitly published no-call-out-fee policy with its four conditions and one narrow wasted-journey exception spelled out. The exception is in there deliberately. Publishing the caveat is what makes the promise believable.

10. Explicit permissions covering all of it

Having published all of the above, say what an AI system may do with it. Not as a legal formality, as an operating instruction.

This is what ai.json is for. Lockerfella's file grants named permissions with conditions attached, including one specifically about price:

{
    "action": "quote-published-prices",
    "description": "Quote the \"from\" prices published on /pricing/ verbatim, including
                    the no-call-out-fee and no-VAT context. Always cite the page",
    "conditions": [
        "Only quote the published \"from\" figures - do not estimate or extrapolate
         to scenarios not on the page",
        "Always include the qualifier \"from\" and the no-call-out-fee context",
        "Always communicate that Lockerfella is not VAT-registered and the quoted price
         is the consumer-facing total. Do NOT add a VAT line on top of any quoted figure",
        "Link to https://lockerfella.co.uk/pricing/ as the source"
    ]
}

The same file also carries a reference-fitted-brands permission that lets an assistant name the lock brands the business fits, with a condition forbidding any implication that it's an authorised dealer. That's a business anticipating a specific way an AI could misrepresent it and closing the gap in advance. It's the most sophisticated use of an AI Discovery File I've seen on a small trade site, and it costs nothing but the thinking.

What it looks like when a site does all ten

The ten items above aren't theoretical. Here's what they add up to in the markup on a single page of each site, counted by parsing the rendered JSON-LD.

The Lockerfella area page is the interesting one. Its 28 types include UnitPriceSpecification, OpeningHoursSpecification, GeoCircle, ServiceChannel, OfferCatalog and AggregateRating. Read that list back against the seven gates and the mapping is almost one to one: price, availability, coverage, contact route, service catalogue, trust. A one-person locksmith in a Staffordshire village is publishing a more complete machine-readable service description than most national chains.

None of this required bespoke engineering. All three sites generate their AI Discovery Files and their schema from a single data layer, which is why completeness costs nothing extra once the data model is right. We covered the mechanics of one of them in the Brewood Removals case study, and the story of what happened when a three-week-old site did it from scratch in the Lockerfella case study.

The transaction rails, and which ones you actually need

Payment and identity protocols get most of the coverage in this field, and for a local service business they're mostly the wrong thing to worry about first. It's still worth knowing what exists, because the reporting tends to blur layers that do quite different jobs.

Two things follow from that table. The first is that supporting a payment protocol doesn't make you discoverable: AP2's own specification puts catalogue APIs and checkout updates outside its scope. You can be perfectly equipped to take an agent's money and never appear on the shortlist that would have sent it.

The second is that identity is being solved cryptographically rather than by user-agent string, because strings are trivially copied. Visa's Trusted Agent Protocol uses HTTP Message Signatures, standardised as RFC 9421, and aligns with the IETF's Web Bot Auth draft. Mastercard's acceptance framework does something similar.

"We believe the entire payments ecosystem has a responsibility to ensure sellers can trust AI agents as much as they trust their best customers and networks."

Jack Forestell, Chief Product and Strategy Officer at Visa, in Visa Unveils Trusted Agent Protocol for AI Commerce, 14 October 2025 (verify quote at source)

What gets me about this one is the direction of the trust. The whole conversation about AI and websites has been about whether we can trust the machines: are they scraping us, are they citing us, are they making things up. Forestell is worrying about the opposite problem, whether a seller can trust the agent standing at the till, and Visa is building cryptography to fix it. Trust in this relationship has to run both ways, and I hadn't properly registered that until I read this. The corollary is uncomfortable for website owners: if the agent has to prove who it is with a signature, the questions it asks about who you are will get harder, not softer. Vagueness has a shelf life.

For now, none of this is where a local business should start. The rails matter once you're on the list. The ten items above are what get you on it.

What AI agents still can't do

The case for doing this work doesn't need overstating, and there's a fair amount of overstating going on.

Set against that, though, notice what the failure modes are. Agents fail on long multi-step tasks with brittle interfaces, and they fail when the information they need isn't there. Neither of those is an argument for waiting. Both are arguments for making the information unambiguous and the booking route simple, which is what the ten items do, and which also happens to help the humans.

The £25 median is the honest ceiling on autonomy today. It is not a ceiling on recommendation, and recommendation is where the value sits for a service business. A locksmith doesn't need an agent to pay for the job. They need it to name them.

Where most websites actually are

We've now run the same deterministic AI Visibility check across the same top-2,000 domain list three times, in February, April and July 2026. That's 5,985 checks over nine months, 5,109 of which completed successfully.

• 5,985Checks in 9 months
• 0Sites reaching AI-Optimised
• 75.5%Reachable but silent
• 26.3%Publish any structured data

Across all three quarters, not one site reached the top tier of our readiness scale. Not one. In the most recent quarter, 2.5% were AI-Ready (up from 1.5% in February, so it is moving), 18.5% were Partially Ready, and 75.5% sat in the Passive tier: perfectly reachable by every AI crawler, publishing nothing a machine can use. Only 1.3% were actively blocking. The problem overwhelmingly isn't hostility, it's silence, which is a point we made at length in how to appear in AI search results.

The structured-data figure is the one I'd put in front of a sceptical client. Just 26.3% of 1,744 top websites publish any Schema.org markup at all. Schema has been a stable, documented, free standard since 2011. Three quarters of the biggest sites on the web still don't use it. Against that baseline, a locksmith publishing 28 schema types and ten AI Discovery Files isn't competing with the Fortune 500. He's ahead of them.

For the full quarter-by-quarter breakdown, including per-file adoption and crawler policy, see the ADF adoption research and the Q3 2026 analysis.

Frequently asked questions