When customers ask AI instead of Google: what changes, what doesn't, what's next.

Calibration first: search isn't dying, and for urgent, nearby needs the search box and the map still take most of the traffic. But assistants are increasingly the first stop for the considered questions (comparisons, how-much, is-this-normal), which are exactly the questions that used to earn your articles their readers. Worth understanding before your competitors do; not worth panic.

Answers, not links

The structural change is simple: a search engine returns a ranked list and the customer does the reading; an assistant does the reading and returns a conclusion. For the businesses being concluded about, that moves the contest. On a results page, position was the prize: being third still got you seen. In an answer, you're either part of how it was assembled or you're absent; there's no third place to occupy. The nearest thing most businesses have seen before is the map pack, a shortlist assembled by machine from data you feed it (how the map decides), and, encouragingly, the inputs that win the pack overlap heavily with the inputs assistants trust.

How assistants choose sources

Nobody outside the model companies knows the ranking math, and anyone claiming to has mistaken a guess for a product. But the observable behaviour is consistent enough to act on. Assistants lean on material they can verify and attribute: pages that answer a question directly and completely, facts that agree with what the rest of the web says about you (consistency again: name, location, services), structured markup that makes the page legible to machines, and signs a real practitioner wrote it (specifics, trade-offs, the honest "when not to"). Thin listicles built to game the old contest make poor citation material for a system whose job is to be right. In other words: assistants are grading for substance, which is inconvenient for spam and convenient for anyone who actually knows their field.

What's oversold

The gold rush has arrived wearing an acronym (GEO, "generative engine optimisation") and some of it is the old snake oil rebottled. Nobody can guarantee your business appears in assistant answers; the systems change constantly, differ from each other, and personalise. "AI visibility scores" mostly measure a vendor's own sampling of prompts, not your market. And wholesale rewriting of your site "for AI" is redundant at best: the qualities assistants reward are the qualities good pages already have. Treat any pitch with a guarantee, a proprietary score, or a panic clock the way you'd treat the same pitch about rankings. The agency red flags transfer to this market unchanged.

The play that works either way

Here's the practical comfort: the work that earns citations is the work that was already right. Publish genuinely useful answers to the questions your customers ask: the guide so specific and honest that a machine summarising the topic can't responsibly skip it (the same logic as writing for buyers, not browsers). Keep your facts consistent everywhere machines look: profile, directories, site. Add the structured markup that labels what you are, where you serve, what you answer. Keep reviews real and current, because assistants read proof the same way humans do. Every item on that list pays in ordinary search today and in assistant answers as they grow. That's what a durable play looks like: no bet on which interface wins.

The agent-native lens: from cited to chosen

When the assistant is asked to choose. The answer is starting to turn into an action. Ask an assistant to "find someone to review the lease and book a call this week" and naming sources is only the first half; it then has to confirm what you offer, roughly what it costs and whether a time is free, and bring the person something to approve. A citation rewards a page that explains. An action needs facts a machine can check without guessing: prices or price ranges, terms, turnaround and how to begin, written as plain text, with the same facts in schema.org JSON-LD. If you take bookings or orders online, keep the flow short and clearly labelled. Some businesses may go further and expose those steps directly, through an API or an MCP server whose tools/list offers, say, check_availability and request_booking. Either way, nothing is committed until the person approves and you confirm.

When the monitoring runs on agents. What assistants say about you can be sampled, honestly. An agent can put the same fixed set of buyer questions to several assistants each week, keep the answers, and note which sources were cited and what was said. Read that log for errors rather than scores: an old price, a retired product, last year's hours. The agent drafts a fix to whichever page or profile the mistake most likely came from, and a person approves it before it goes live. Label any chart built from the log as a sample of your own questions, not a measure of the market.

Measure the observable, label the rest

Some of this is measurable now: assistant browsers show up in your analytics as referral traffic, and "how did you hear about us" catches the customers who say "I asked ChatGPT". Log both. The rest (how often you're cited, what answers say when you're not in the room) is genuinely dark beyond your own sampling, and honest reporting labels it that way instead of decorating it with invented precision. As assistants begin to book and buy for people, more of it becomes countable: tag the bookings and orders that arrive that way, where you can tell, and the dark share shrinks. The test for any spend here is the same as everywhere else in this library: does it produce customers you can count? If a vendor's answer relies on a score only they can see, you already know what the score is for.