Author: Ben Rogers

Google’s AI can now phone a local business directly and ask for a price, on behalf of a person who never picks up the phone themselves. The capability is live. It’s also a narrower, more concrete version of a bigger trend: the next lead through the door might not be a person at all.

Handing a decision to someone speaking on your behalf is not new. Roman law had a name for it, a procurator, empowered to strike a deal on a principal’s behalf when the principal couldn’t be there. What’s new is who that agent has become: not a person weighing tone, but a script with no patience, moving on the moment the first answer isn’t good enough. A business that spent years earning a human agent’s trust now has to earn something colder, on the first ring.

What actually answers the phone

It’s 2pm on a Tuesday, and the phone rings at the office. What answers?

For a lot of operators, an IVR menu that takes four selections to reach a human, or a front-desk person who’s covering three jobs at once and quotes a rough number from memory because there’s no time to check the file. Sometimes it’s voicemail.

A person wouldn’t count any of those as a failure. A caller who reaches voicemail usually just calls back, or waits for a callback, or tries the next name on their list. An AI doesn’t do any of that. It takes whatever it gets in that one call, whether that’s a wrong number, an outdated rate, or no answer at all. Then it reports that back as the business’s actual offer. No human on the other end thinks, “They just missed my call, I’ll try again.”

The real shift isn’t that AI is calling businesses. It’s that the call now has exactly one attempt to produce a usable answer. Most operators’ phone process was built for a caller with patience, not a system with none.

What the call itself asks for

The pattern is close to what a busy adult child might do calling on behalf of a parent: state the job in plain terms (a three-bedroom house, a date range, an origin and destination) and ask for a rough number. A good human salesperson handles that fine, because they can read hesitation, offer a callback, or say “let me get you an exact quote in five minutes” and mean it.

An AI caller skips the social reading and won’t wait around for the callback either. It asks its question once, in a fairly literal way, and it either gets a workable answer or it doesn’t. If the person who answers says “it depends, let me have someone call you back,” that’s a dead end from the AI’s side of the conversation, not a follow-up opportunity. It moves to the next business on its list, because that’s the entire job it was given: gather a workable number, not build a relationship with whichever business happens to answer first.

What Google Duplex showed in 2018, and what ships now

Back in 2018, Google Duplex called a hair salon and a restaurant live on stage at Google I/O, complete with human-sounding pauses and “umms.” That demo was a stunt: one hand-picked call, rehearsed categories, and enough backlash over undisclosed AI callers that Google had to commit publicly to having the system identify itself. The underlying idea hasn’t changed since 2018. What shipped as a stage trick now ships as a search feature: Google describes “Ask for me,” which calls local businesses on a searcher’s behalf to check pricing and availability. By mid-2026 Google’s own Search blog described the same capability for “select categories like home repair, beauty or pet care”, rolling out to US users that summer.

That expansion path matters more than any single category on the list. A capability that requires a customer to seek out and configure a specific tool spreads slowly. A capability that Google folds directly into Search spreads at whatever pace Google decides. Google’s own blog says AI Mode passed 1 billion monthly active users within about a year of its US launch. Moving and relocation aren’t on the published category list yet, and Google hasn’t said what share of “Ask for me” calls land on any specific industry. But the model doing the asking doesn’t know or care what industry it’s calling. Only sequencing stands between a nail salon and a moving company.

This isn’t a technology problem

Fixing this requires knowing what the phone answers with, right now, without anyone standing next to it. Does the person who picks up have current pricing on hand, or are they estimating from memory? Does the IVR menu bury a real quote request four layers deep? Is there a version of the pricing that’s accurate enough to give out cold, on the first call, to a caller who won’t wait for a follow-up?

Most operators have never tested this from the outside. The business’s own phone process was built and refined for human callers, who tolerate friction, voicemail, and “let me get back to you.” A system gathering quotes on someone else’s behalf tolerates none of it. It moves on to whichever competitor’s phone happened to produce a real number on the first try.

A five-minute audit worth running

This is concrete enough to test today, without waiting for an AI to call first. An operator can ask someone outside the business (a friend, a family member, anyone who doesn’t already know the internal pricing sheet) to call the main line cold and ask for a rough price on a standard job. The test measures how long it takes to reach a human, and whether the caller gets a real number or a “let me find out and call you back.” The same call placed after hours usually goes to voicemail or routes somewhere unstaffed, which is its own answer.

That test reveals what an AI caller would experience, except a human tester can also report on tone, on whether the wait felt reasonable, and on whether the eventual answer felt trustworthy. An AI reports a price, or no price, to whoever it’s working for, and moves on either way.

The businesses that are already ahead of this

Some operators aren’t waiting to find out how their phone performs, because they’ve made the phone a smaller part of the story to begin with. A business with an accurate, structured instant-quote tool on its own website has already solved most of this problem before the phone ever rings, because a caller, human or otherwise, who can get a real number online never needs to place the call at all. When a business treats its pricing as structured, always-current data, rather than something that lives in a salesperson’s head, it holds up under both forms of pressure, a person asking online and an AI asking by phone.

A phone audit like the one above tends to embarrass the businesses that haven’t done that work. Not because their service is worse. Nobody who wasn’t already patient with them has ever tested the gap between what they actually charge and what their phone process can produce on demand.

What changes if this becomes the normal way people shop

The interesting version of this trend arrives once AI-mediated calling stops being a novelty and becomes one of several normal ways a household starts a moving search. In that world, every AI system running this kind of search samples the businesses that show up as usable answers: an accurate price on first contact, no dead-end voicemail. It skips the rest, the same way a directory listing with no phone number never gets called.

That’s a different kind of competitive exposure than losing a customer to a competitor with a better sales pitch. A better sales pitch stops being the advantage it used to be once the system deciding who gets sampled never hears it. The exposure is closer to being systematically excluded from a channel, because the infrastructure behind the phone was never built to be interrogated by something that won’t wait. The operators fixing this now are making their business legible to a form of demand that’s still early enough to fix cheaply, before it’s the default way a meaningful share of moving searches happen.

What “good” looks like here

The fix isn’t complicated, even if it takes real work to build. It starts with a short, clearly-stated price range for the most common job types, available to whoever answers the phone without needing to check a spreadsheet. It continues with an IVR path, if one exists at all, that reaches a real quote option in one or two selections, not four. And it includes a documented after-hours answer, even a simple one, rather than silence. None of this requires new technology. It requires treating the phone the way a website already gets treated: as a piece of infrastructure worth auditing and maintaining.

“This only matters for a tiny fraction of calls right now”

The objection that this only matters for a tiny fraction of calls right now is a fair one. It’s also the kind of objection that stops looking reasonable in retrospect, once the underlying trend has run its course for another year or two. Nobody has precise numbers yet on what share of enquiry calls in this specific industry currently come from an AI acting on someone’s behalf, and any specific percentage offered here would be invented, not sourced. The mechanism is measurable, even if the volume isn’t: the capability exists, it’s shipping inside a product Google has already put in front of a mainstream search audience, and the direction of travel only points one way. Waiting for the volume to become undeniable before fixing a phone process that was probably worth fixing anyway means fixing it under pressure. By then, the competitor who moved first has already captured whatever share of AI-mediated calls existed during the gap.

The question worth sitting with

If an AI called your office right now and asked for a price on a three-bedroom interstate move, what would it be told? Not what your best salesperson would say with time to think. What would genuinely answer the phone.

If you don’t know the answer, that’s the finding.

The phone call is the narrow case. The wider one shows up when an AI never calls at all and just reads what your website already tells it: see what happens when an AI chooses the mover.

Google’s AI Mode passed a billion monthly users in May 2026. It can now call a local business directly on a customer’s behalf and ask for a price.

It changes something more specific than “AI is coming for the industry.” It changes who gets to compete for a job in the first place, and it does that before a human customer ever types a business’s name into a search bar. Call it AI visibility: whether the system doing the choosing can actually read what a business has to offer.

Every era hands the job of choosing a tradesperson to a different gatekeeper. Getting chosen has always meant being legible to whichever one is currently in charge. A medieval market required a trader to register with the local authority before a single customer could legally buy from their stall. A moving company later needed to be legible to a phone book, then to a search engine’s crawler. The gatekeeper keeps changing shape. What it demands of a business trying to get chosen does not: information a stranger or a system can actually read, in whatever format that stranger currently reads.

The click is already disappearing

For twenty years, discovery worked through search results a person read, compared and clicked. That’s breaking down. Pew’s research on AI-summarized search found that click-through to traditional results drops when an AI summary appears above them: from 15% of visits to just 8%. Nearly half the click-through disappeared, not because rankings moved, but because the reader never scrolled past the summary.

This is the shift for anyone who still does their own clicking. A narrower group of buyers goes further: they hand the whole search to an agent and never read a summary at all. For that group, a business’s homepage isn’t competing for attention. It’s competing to be parsed correctly by a system that never sees the page the way a person does.

What “machine-legible” actually means

Schema.org defines a structured format for local business listings: a JSON-LD block that states a service area, a price and a pricing unit in a shape a program can parse without guessing. Two pricing pages can look identical to a person. One is a paragraph of marketing copy ending in “contact us for a custom quote.” The other has the same information wrapped in that structured block. Only one of them hands an AI agent a usable number.

Named crawlers cover part of this, though not the part that matters most, and one of them isn’t a crawler at all. Google’s own documentation says Google-Extended “doesn’t have a separate HTTP request user agent string”: it’s not a bot reading a page, but a permission flag that controls whether content Googlebot already fetched can later train Gemini. The flag has no effect on search indexing or ranking. GPTBot and ClaudeBot are real, separate crawlers, but they serve the same after-the-fact purpose. OpenAI says GPTBot “is used to crawl content that may be used in training” its models. Anthropic describes ClaudeBot’s job the same way. Neither one reads a page at the moment an agent is deciding who to call. That work runs through a different set of bots: OAI-SearchBot and PerplexityBot exist specifically to “surface websites in search results,” and fetchers like ChatGPT-User and Perplexity-User read a page the moment someone asks a real question. When a site gates its pricing behind a form, a phone number, or a quote-request flow with no static, crawlable answer underneath it, that site is legible to a person willing to fill out a form and wait. It isn’t legible to a system reading in bulk, on someone’s behalf, in the seconds before it reports back a shortlist.

A middle case is even more common than having no pricing information at all. A rate sheet might exist only as a PDF behind a “Download Our Rates” link, or as pricing baked into a photographed infographic. Or the price might only appear after a customer clicks through an interactive quote calculator built entirely in JavaScript, with nothing rendered on the page for a program that doesn’t execute scripts. To a person, all three look like a business that publishes its pricing. To an agent reading the raw page, a PDF is a file it may or may not open, an image has no extractable text, and a client-side calculator can be functionally blank until a human clicks a button the agent never will. Being half-legible produces the same outcome as not being legible. The agent moves on to the next name on its list.

An agent doesn’t browse. It reads.

An AI agent gathering quotes on someone’s behalf doesn’t scroll past a hero image, read a testimonial, or form an impression from a logo. It reads whatever structured, machine-legible information a business makes available, and it reads several competitors in the same pass. When an operator’s pricing, service area or availability aren’t legible to that kind of read, the operator doesn’t lose the job. It never enters the set of options the agent reports back.

That’s a different failure mode than losing to a lower bid. It’s not competing at all. A homepage built to impress a human visitor can be functionally invisible to the system now shortlisting on that visitor’s behalf. The two failures look identical from the outside: fewer inbound quote requests, no obvious cause, nothing on an analytics dashboard flagged red.

Two operators, one quote request

Here’s a hypothetical two-bedroom apartment move: 18 miles, second-floor walk-up on both ends, no stairs at the truck. A customer’s AI agent is gathering quotes for it, querying five local movers.

Operator A’s site has a clean, modern homepage: photos of a branded truck and three testimonials. A single button reading “Get a Custom Quote” opens a contact form asking for name, email and phone number. No price appears anywhere on the page in text a program can parse. The agent has nothing to extract. Operator A doesn’t appear in the comparison the agent hands back to the customer, not because the price was too high, but because there was no price to read.

Operator B’s site has a plain-text rate posted on its pricing page: two movers and a truck at $109 an hour, a two-hour minimum, and a $75 flat travel fee inside 25 miles, all wrapped in the structured markup described above. The agent extracts a concrete figure: two hours at the hourly rate plus the travel fee, $293, flagged as an estimate rather than a binding quote. That number lands in the customer’s shortlist alongside four competitors. Operator B is now one of the businesses the customer considers, whether or not $293 turns out to be the lowest number on the list.

Nothing about Operator A’s service, price or reputation was worse than Operator B’s. The difference sat entirely in what the page said to something that can’t ask a follow-up question.

“We still get calls, so this isn’t urgent yet”

The “we still get calls” objection is fair, and deserves a real answer, not a dismissal.

Most quote requests still arrive through a person who called or clicked, and that will stay true for a while yet. But the same structured pricing data that makes a business legible to an AI agent also makes it eligible for the price-forward rich results ordinary search already favors. They show a price range directly in a result before anyone clicks anything. When businesses stop keeping their pricing exclusively in a salesperson’s head and put it into a structured, always-current format instead, they hold up better under both kinds of scrutiny: a person comparing search results and a system reading on someone’s behalf.

The real cost of waiting isn’t the AI traffic missed this month. Publishing accurate, structured pricing takes time. A competitor who does it now is already indexed and answerable by the time a meaningful share of quote requests start arriving this way. Once the volume becomes undeniable, fixing it under pressure means fixing it a step behind whoever moved first.

The fix is smaller than it sounds

The other objection worth taking seriously is practical rather than strategic: most moving companies don’t employ a developer, and “add structured data to your website” sounds like a project that needs one.

It doesn’t require a rebuild. A JSON-LD block is a small, separate piece of code that sits alongside a pricing page without changing how the page looks or reads to a person. How much of that work is manual depends on the platform. Rank Math is one of the SEO plugins most WordPress moving-company sites already run for their title tags and meta descriptions. It also has a Service schema type built for exactly this: enter a price and a currency in a field, and the plugin writes the structured markup itself. Yoast, the other common WordPress plugin, does the same automatically for a business’s address and hours. Wix and Squarespace both generate some structured data of their own, but neither gives an owner a field for a custom price the way the WordPress plugins do. Getting a price into either one’s markup means using a manual or code-injection option instead of an automatic one, a real difference in effort, not a rounding error. The actual work isn’t code. It’s deciding on a real, current number for each service and publishing it in plain text somewhere a program can read, the same discipline a business already needs so a phone quote doesn’t contradict a website quote.

What exclusion actually costs

A business that ranks eleventh on a results page for a competitive search term still gets found by a person willing to scroll, or willing to try a more specific search the following week. Ranking eleventh is a bad outcome, not an invisible one.

Being excluded from an AI agent’s shortlist doesn’t work that way. An agent gathering five quotes reports back five names. No eleventh position exists for a person to stumble across by scrolling further, because the agent isn’t a ranked list a customer pages through. It’s a finished answer, handed in this scenario to someone who never sees a results page at all. A business with no usable pricing data isn’t ranked low. It’s absent from the answer, the same way a business with no listed phone number isn’t ranked low in a directory. It’s left out of the directory entirely.

That distinction changes how large the risk is. A slow decline in organic rankings is visible and gradual, something a business can watch happen and respond to. Exclusion from an agent’s shortlist leaves no ranking to watch. It just produces fewer quote requests from a channel the business never knew it had lost, with nothing on a dashboard pointing back to the cause.

AI Visibility: What Gets Read Before Anyone Calls

None of this is speculative. The capability exists today, inside a product already used by more than a billion people a month. How fast operators adjust to it is still open. Most won’t, because “make sure a machine can read your pricing” doesn’t feel urgent next to a quote that needs sending this afternoon.

Reading the page is the wide version of this problem, the one this piece has been about. The narrower version happens when the agent doesn’t stop at reading and picks up the phone instead. Google calls a business’s office directly and asks the question out loud, the same way a customer would, minus a customer’s patience for a callback.

An AI agent finds the operators who treat machine-legible pricing as infrastructure built once and kept current, not as a future problem to get to eventually. Everyone else keeps competing for a shrinking set of jobs that still route through a person who read past the summary. That’s a smaller share of the searches happening at all.