The Recommendation

The short version

For twenty-five years the job was simple: get a human to your website, then convert them. That job is ending. Not because search is dying, but because the searcher is changing. The buyer stays. Their job as the researcher goes to a personal AI agent that compares, diagnoses, checks availability and, increasingly, books. Your website stops being a place a person reads and becomes a source an agent either trusts or skips. So the question changes. It’s no longer “how do I rank?” It’s “how do I become the recommendation?” You earn that with five things an agent can actually read: transparent pricing, real online scheduling, a page for every service, content built from the questions real customers ask, and one consistent, verified body of truth about your business everywhere it appears. Install those and you’re legible to the thing now making the decision. Skip them and you’re invisible to it.

First, the plain-English version

This guide uses a few terms that get thrown around without ever being explained. If they’re already second nature, skip to the next bit. If not, here’s the lot, plain, because the jargon is mostly hiding how simple the shift really is.

AI agent (and “agentic”). A chatbot answers a question. An agent does the job: it browses, compares, fills in forms, even books and buys, working in the background for someone, with their say-so. “Agentic” just means a machine did it, not the person. The whole guide hangs on this one word: the agent takes action.

AI search and “zero-click.” Google used to hand back a list of links. Now it reads the web and writes the answer straight on the page. “Zero-click” is the punchline: the searcher gets their answer and never lands on anyone’s website.

Structured data (you’ll hear “JSON-LD” or “schema”). A bit of hidden text on your site that labels your details so a machine reads them cleanly: this is the price, these are the hours. It’s the difference between a machine guessing what your page says and being told.

Knowledge catalog. One tidy, verified store of everything your business knows: services, prices, answers, policies, built so a machine can read it straight off, instead of crawling your whole site to work it out.

Agent-to-agent. A shared language that lets your buyer’s agent talk straight to your business’s systems and get an answer back, app to app, on everyone’s behalf.

Hold the first one, an agent takes action, and the rest follows.

Why this matters now

People stopped clicking.

In the 2024 clickstream study, 58.5% of US Google searches ended with zero clicks: the searcher got the answer on the page and never visited a website. By early 2026 that was past 68%. More than half, heading toward two-thirds, of all the “traffic” everyone optimises for never lands anywhere.

Then the search box itself changed. At Google I/O in May 2026, Google called its redesigned box the biggest change to search in over twenty-five years: a Gemini-powered box that holds a conversation, takes images and files, and runs background agents that go and do the work for you. Around the same launch, Google introduced Gemini Spark: a 24/7 personal AI agent that, in Google’s own words, helps you “navigate your digital life, taking action on your behalf and under your direction.” The demo slides said it plainly: browse it, book it, buy it. It does the research, compares the options, and helps complete the booking.

Sit with what that means for a trade. A homeowner used to type “gutter cleaning near me,” read a few sites, scan reviews, and call someone. Now they say to their agent: my gutters overflow every time it rains, find someone good, check what they charge, see when they can come, and book it. The agent reads your Google Business Profile, looks across the web, weighs the options, and acts. The homeowner never visits your site. They may never call. They just get a recommendation and a booked appointment.

This isn’t a someday story. Google’s CEO Sundar Pichai has named 2027 as the inflection point where these agentic workflows shift “pretty profoundly,” and reckons only about 0.1% of the world is living this future today. The plumbing is already shipping: agentic booking is live in Google’s AI Mode for Ultra subscribers in the US, scanning across OpenTable, Resy, Booksy, Fresha and others to actually make reservations. And agents are learning to talk to each other: Google’s Agent2Agent (A2A) protocol, launched at Cloud Next in April 2025, is already routing real tasks between agents built on different platforms at 150-plus organisations. Your agent talks to the supplier’s agent talks to the logistics agent, and an answer comes back. No search, no website, no phone tag.

The yellow-pages-to-Google shift took a decade and ended careers that didn’t move. Google-to-agents is the same size of shift, moving faster.

If this still sounds like futurist hype, fair, most “AI changes everything” talk is. So don’t take the claim; take the receipts. Every shift above is a shipped product with a date and a source at the foot of this page. None of it is a prediction. It’s already live; it’s just unevenly distributed.

The real story: the buyer outsources the work

Here’s the part to get straight, because everything else follows from it.

The customer isn’t going away. The researcher is. Everything a buyer used to grind through, comparing websites, reading reviews, calling around, waiting for callbacks, chasing quotes, checking availability, they hated all of it. Now an agent removes that friction and does it in the background. The buyer just receives the shortlist and the booking.

So you are no longer optimising for a human skimming your site at 9pm. You’re optimising to be the business an agent can find, trust, and recommend. Make information easy and structured, and you become easy to recommend. Create confusion, gaps, or contradictions, and you become easy to skip. The agent doesn’t argue with a missing pricing page; it just moves to the competitor who has one.

Old model: search, links, research, decision. New model: question, agent, recommendation, action.

And notice where the weight moves: to the question. People won’t ask agents for “HVAC repair near me.” They’ll describe a problem: my AC stopped, repair or replace? … my roof was damaged in a storm, what now? They’ll let the agent diagnose. Which means the businesses that have published real answers to those real questions are the ones the agent has something to pull from. The rest are invisible, not because they’re bad at the work, but because the agent has no evidence they do it.

You know this, if you’re honest. You’ve watched your own buying change. You don’t trawl ten websites anymore either; you ask, and you trust the answer.

This is the whole reframe. The question is no longer “how do I rank?” It’s “how do I become the recommendation?”

The five things an agent needs to recommend you

Run your business past each of these honestly. If an AI agent landed on you right now, on behalf of a buyer, could it answer yes?

1. Can it see your pricing? “How much does it cost?” is the most common thing buyers ask. If there’s nothing on your site about cost, the agent has nothing to reference and recommends someone who does. The fear here, my competitors will see my prices, is the wrong fear. You don’t publish your price book. You publish a pricing guide: the ballpark, the low end and the high end, what pushes a job up or down, what the add-ons are, why one operator charges $299 and another $999. You become the authority on what this should cost. Agents quote authorities.

2. Can it book you? Agentic booking is already real. If the agent finds your price, likes your reviews, and then hits a dead end because there’s no way to schedule, you lose the job at the last step. You don’t need full open-calendar booking if your route density won’t allow it: even online scheduling for a virtual estimate gives the agent a slot to take. And check that whatever you run, Jobber, ServiceM8, Tradify, ServiceTitan, Housecall Pro, is moving toward integrating with agentic booking. If it isn’t, it’s fine for today and a liability for next year.

3. Does it understand every service you do? If you’ve got one thin “Services” page, the agent only knows the headline. Build a real page for every service. The agent can’t recommend you for gutter installation if there’s no evidence on the open web that you do gutter installation.

4. Can it find answers to your customers’ real questions? This is the one almost everyone gets wrong. Generic blog posts, “7 ways to prep your home for summer”, are what Google now calls commodity content: fluff anyone, or any chatbot, could produce. It doesn’t build trust and increasingly it’s filtered out. What earns trust is hyper-specific, real content built from the actual questions your customers ask. Not what you think they ask, what they actually ask, which lives in your phone calls. Every call your business takes is future content. Document the questions, answer them publicly, and you train the agent to treat you as the source.

5. Is your information consistent and verified everywhere? Agents are obsessive about trust because they’re making a recommendation they’ll be judged on. Mismatched names, numbers, or addresses across directories read as risk, and risk doesn’t get recommended. One consistent identity, real reviews on Google and the platforms that matter (reviews are now trust signals for the agent, not just for humans), and a side-by-side that explains honestly why you differ from the cheaper option: that’s what makes you safe to pick.

Most websites can’t tick those boxes today. That’s not a problem. That’s the opening.

The deeper asset: your knowledge catalog

The five things above are the crawl era: the agent reading your site for pricing, services, and answers. But Google has already shown where this goes next, and it has a name.

Google’s been building the scaffolding for businesses to be directly legible to agents, not just crawled: a Knowledge Catalog (a structured, verified body of what your business knows that agents query for ground truth), an Agentic Resource Discovery spec where ownership of your own verified domain becomes the cryptographic basis for an agent trusting you, and the A2A protocol that lets your agent transact with theirs. The enterprise versions of these are live now. The local-business version is coming, and it favours whoever has their house in order first.

Think of it this way. Today your website is a rough knowledge catalog, but a passive one, waiting on Google to crawl it and decide what to surface. The shift is from waiting-to-be-crawled to handing the agent a verified dataset of everything your business knows, which you can then deploy as an agent on your brand’s behalf. When the world runs on agent-to-agent interaction, you want an agent in the room: a digital twin of your business that can answer for you 24/7.

A knowledge catalog is a structured database of everything your business knows: customer questions and answers, services and service areas, your expertise and credentials, FAQs, pricing guidance, policies and process, reviews and testimonials, and insights pulled from real call transcripts. Today’s website is the front end. Tomorrow the knowledge catalog is the foundation, and the competitive edge stops being the best-looking or highest-converting website. It becomes the most complete, most trusted data.

The reason to start now is simple: almost nobody knows this is happening. First mover gets to build the repository of trust before the field fills up.

How Deliverators productises this

This is the shift the Deliverators system was rebuilt around, so it’s worth being clear about how the pieces map, both as our internal model and as what we install for a client.

The hardest input to get is the real one. Every operator thinks they know their customers’ questions; the gold is in the transcripts. So the engine is calls-to-content: an agent that plugs into a client’s customer calls, listens, and turns every valuable conversation (the three-minute-plus calls where something real happens) into the assets that matter: answer-driven service content, FAQ entries from genuine questions, social and video hooks, and structured Q&A. Not chatbot slop posted for likes. Real questions, answered publicly, which is exactly what trains an agent to trust the business.

That same engine rolls the answers up into a knowledge catalog: the verified body of truth a client can eventually deploy as their own agent. Pricing guidance, services, areas, policies, expertise, and the real-question content, all in one consistent place. We do this because we want clients holding first-mover advantage while the field is still empty.

The point isn’t that this only works one way. An operator can build pricing guides, switch on online scheduling, and write real content themselves, and should. The point is the shape of the work has changed, and any marketing that still starts with keyword-stuffed pages or AI-generated filler is preparing clients to be un-recommended. The work now is real content, structured data, and a verified identity, built for the agent that’s about to be doing the choosing.

The practical stack: what actually ships

Concept is nothing without the install. Here’s the stack that makes a trade legible to an agent. Each piece is either technical (build it once, build it right) or editorial (the raw truth only the business has). Get both and you’re machine-readable and trustworthy; get one without the other and you’re neither.

Structured data on every page (JSON-LD schema). This is the single technical lever that genuinely moves the needle. AI engines lean on structured data to understand, verify, and cite a business: pages with proper schema markup are materially more likely to be quoted in AI answers, and JSON-LD is the format Google, Bing, ChatGPT and Perplexity all read. For a trade that means an Organization entity up top and a LocalBusiness entity per location (name, address, phone, hours, geo), plus Service markup on each service page. One rule from Google’s March 2026 update: the schema must describe what’s actually visible on the page. Markup that lies about the content now costs you trust instead of building it.

A page, with schema, for every service. No agent recommends you for a job there’s no evidence you do.

A pricing guide per service. The ballpark, the range, what moves it. The authority play, not your price book.

Real online scheduling or lead capture, wired to a CRM with agentic-booking on its roadmap, at minimum a virtual-estimate slot an agent can take.

Answer content built from real calls. The questions your customers actually ask, answered as genuine pages. Note: Google deprecated FAQ rich snippets in May 2026, so don’t build these as a schema trick to win a snippet. Build them as real content that answers the decision-stage question. That’s what survives.

One consistent NAP plus a verified Google Business Profile. Same name, address, phone everywhere, with schema sameAs pointing to your GBP. Inconsistency between your site and your listings is the fastest way to lose an agent’s trust: it reads mismatched details as risk, and risk doesn’t get recommended.

An llms.txt file, honestly. This is a plain-text map of your best content for AI tools, served at your domain root. Here’s the truth most people selling it won’t tell you: Google has stated outright that llms.txt does nothing for Search rankings or AI Overviews, only about one site in ten runs one, and the search bots barely fetch it. So do not let anyone sell it to you as an SEO trick. But it is the first real business-to-agent standard, and the agentic tools (in-product assistants, MCP servers, coding agents) do read it. It costs an afternoon. Install it as a cheap, early bet on the agent web, with clear eyes about what it is and isn’t.

AI share-of-voice and mention-rate tracking. Stop tracking where you rank. Start tracking how often agents recommend you, and how often they mention you. Those are the new scoreboard.

Who owns what

The split is simple: the agency owns the machine; you own the truth that feeds it. Neither works alone: structure with no real input is slop, and the best raw material with no structure never gets read.

Asset	Agency / Deliverators owns	You own internally
Service pages + JSON-LD schema	Builds the pages and the markup	Supplies the real service detail
Pricing guides	Writes and publishes them	Gives the real numbers and what moves them
Online scheduling	Wires it to the CRM	Sets the availability rules
Answer content (calls-to-content)	Runs the engine, drafts, structures	Grants call access; approves what’s true
Knowledge catalog	Builds and maintains it	Owns it as your asset; keeps it current
NAP + GBP consistency	Audits and fixes the directories	Verifies the GBP, keeps details current
llms.txt + technical setup	Implements and monitors	(purely technical)
Reviews	Sets up the ask-and-respond flow	You and the crew actually ask and reply
AI share-of-voice monitoring	Tracks it and reports	Acts on the gaps it surfaces

The build: install this for one service first

Don’t boil the ocean. Pick the one service that brings your best work, and make the agent able to recommend you for it. Each step leaves you with something real.

Five steps, one service. Notice these are all the editorial half: the raw truth only you have. The technical half (the JSON-LD schema on each page, the sameAs link to your GBP, the llms.txt file, the monitoring) is the agency’s job, wired in behind the same five steps. You bring the truth; the machine makes it legible. Action creates evidence, and the agent can only recommend what you’ve given it proof of. Then repeat for the next service. You’re not building a website. You’re building the body of trust the recommendation will be drawn from.

Why now, not next year

There’s a window here, and it’s a narrow one. The agents are arriving whether you prepare or not, but right now almost no trade has a pricing guide, real-question content, or a verified body of truth an agent can read. That means the first operator in each area to get legible becomes the default answer before there’s any competition for the slot. And defaults are hard to dislodge; the agent that learned to trust you first keeps recommending you. Wait until your competitors have figured this out and you’re not installing an advantage anymore, you’re trying to claw back one. You don’t need to do all of it this quarter. You do need to start one service this week, while the field is still empty.

The one line to remember

Rankings were about being found. The next three years are about being recommended. Stop optimising for the human who’s about to stop searching, and start being legible, provably, consistently, in their own customers’ words, to the agent that’s about to start deciding.

Start with the Visibility read: it scores the five signals an agent reads before it recommends you, and names the first one to fix. No sales call, just an honest read of whether buyers and their AI agents can even find and recommend you yet, or whether you’re still invisible to the thing now making the decision.