Take something with you · Thesis 03
The search box is becoming a question.
Being findable used to mean being indexed. It's coming to mean being legible — and to whatever is now reading the web on a guest's behalf, most restaurants read as an empty page.
"Where should we eat tonight near the theater — something with a patio, one of us is vegetarian, and we need a table for six at 7:30."
Read that sentence again and notice that it cannot be typed into a search box. It isn't a query. It's a request with five constraints in it, and until quite recently there was no way to make it of a machine. You'd have decomposed it yourself: search "restaurants near theater," open eight tabs, check each for a patio, scan each menu for something a vegetarian would eat, then call three places to ask about a six-top at 7:30.
Now the whole sentence goes in at once, and three names come back.
Three. Not ten blue links, not a page of results with your listing at number six where a determined person might still find it. Three names, chosen and handed over, and you are either among them or you did not exist for that dinner.
The old contract: the machine indexed, the human read
For twenty years the arrangement was stable and everybody understood their role. Search engines built an index of doors. They ranked the doors. The human being did the reading — opened the tabs, judged the photos, made the call, decided.
Everything we learned about being findable was learned inside that contract. Ranking mattered because position determined which doors got opened. Keywords mattered because the index was matching strings. And critically, being poorly readable was survivable, because a motivated human could squint at a PDF menu, zoom into a photo of a chalkboard, or just phone you.
The human filled in every gap the website left. They were the interpreter, and they were free.
The new contract: the machine reads, and returns a judgment
Under the new arrangement, the machine does the reading and hands over a conclusion. The list collapsed. The interpreter went away.
And a machine reading on someone's behalf can only recommend what it can actually read and check. This is the whole thesis, and it is much less about technology than it sounds:
A PDF menu is not a menu. It's a picture of a menu. So is a photo of a chalkboard, a price inside a hero image, and a patio that exists only in a photograph.
To the thing answering that question about the theater and the patio and the vegetarian, a restaurant whose website is a beautiful full-screen video, a logo, an address, and a downloadable PDF has said nothing. It has no patio, because no text says patio. It has no vegetarian dishes, because the dishes are in a picture. It has no 7:30 availability, because nothing is bookable. It cannot be matched to any of the five constraints, so it is not returned.
Not ranked lower. Absent. There's no page six to be on.
This is a bigger change than SEO ever was
It's tempting to file this under search-engine optimization and hand it to whoever handled that before. It's a different kind of problem.
SEO was a competition for position among the readable. Everybody in the running had already cleared the bar of being a web page with words on it; the contest was over which page ranked above which. It was gradual, relative, and forgiving — being on page two still meant existing.
What's happening now is a qualification test, and it runs per question. For any given request — the six-top with the patio, the gluten-free birthday, the Tuesday walk-in at 9:40 — you either satisfy the constraints in a form the reader can verify, or you're not in the set of candidates at all. It's binary at the moment it matters, and it's invisible: nothing tells you that you weren't recommended. There's no rank to check. The traffic that never happened leaves no trace.
That invisibility is why this is underpriced by exactly the people it hurts most.
Legibility has a second half: currency
Readable isn't sufficient. The other thing any recommender is doing, whether it's a search engine, an assistant, or a friend, is estimating the risk of being wrong.
Recommending a restaurant that turns out to be closed on Mondays, or that dropped the dish someone was promised, or that hasn't had a patio since the renovation, is a bad outcome for whatever made the recommendation. So staleness gets treated as risk. Information that agrees across several places and was updated recently is safe to repeat; information that exists in one place and hasn't moved since 2019 is not.
Freshness stopped being a ranking trick. It became a trust signal — the cheapest evidence available that the thing being described still exists.
Which produces an uncomfortable inversion for anyone who built a beautiful site once. The site that was finished three years ago and hasn't needed to change is, by the new reading, the least trustworthy description of the restaurant available. The restaurant changed. The description didn't. Nothing in it can be relied on tonight.
The unit of discovery moved from the keyword to the constraint
Here's the shift stated as plainly as it can be.
A keyword is a topic: italian restaurant san jose. A question is a set of constraints: six people, 7:30, patio, vegetarian, walking distance from the theater, and — unspoken but always present — open tonight, worth the money, not too loud to talk.
Topics were things you could be broadly associated with. Constraints are things you either satisfy or don't, and each one is a place where you're kept in or dropped. The average independent restaurant's website answers, in machine-readable text, approximately none of them. Not out of carelessness — because the old contract never required it. The human interpreter used to handle the constraints, by calling.
The honest counterpoint
None of this means writing for robots, and if this turns into another round of keyword-stuffing with "vegetarian options" jammed into a footer, it will fail exactly as that always has. The saving grace is that the questions machines ask on a guest's behalf are precisely the questions guests have always asked on the phone. Answer the guest — plainly, in words, on the page — and the machine is satisfied as a side effect. The machine is a proxy for a person. It has never been the audience.
You used to be able to hear these questions
The strangest part of this shift is that none of the questions are new.
Every one of them has been arriving at the host stand for as long as there have been restaurants. Is there parking. Can you do a table of eight. Is it too loud for my mother. Do you have anything my daughter will eat. Is the patio heated. Do you still have the short rib. Somebody in the building answered them all day, well, in a warm voice, and never thought of it as marketing because it was just the phone.
What changed is not the questions. It's that they're now being asked somewhere the restaurant can't hear them — into a search bar, into an assistant, into a map — and answered by whoever left enough evidence lying around to be answered with.
The restaurants that get chosen for the six-top on the patio at 7:30 are not the ones with the best food. They're the ones whose answers were in reach at the moment the question was asked.
Take something with you
The twelve questions guests ask before booking
Open your own website on a phone, as a stranger would, at eleven at night. Give yourself sixty seconds per question. An answer only counts if it's in text on the page — not in a photo, not in a PDF, not "call us."
- Are you open tonight, and until when? Including the holiday, the private-event closure, and the kitchen closing an hour before the room does.
- Where do I park? The lot, the street after six, the garage across the way. The most-asked question in most cities and the least-answered on most sites.
- Can you seat six? Eight? Ten? The large table is the highest-value booking a restaurant takes and the one most often lost to silence.
- Do you take reservations, or is it walk-in? And if walk-in — how long is the wait at eight on a Saturday, roughly, honestly.
- Is there a patio, and is it heated? Any physical feature that only exists in a photograph does not exist to a reader.
- What can a vegetarian eat? Someone gluten-free? Named dishes, in words. "We can accommodate dietary restrictions" answers nothing and reassures no one.
- Will my kids be welcome? High chairs, a small plate, the hour before seven when nobody minds. Half of the family bookings in your neighborhood are decided on this line.
- How loud is it? The single most common reason a booking is quietly abandoned, and almost never addressed by anyone.
- What does dinner actually cost? Not a dollar-sign symbol — a range, from a real check. Price opacity reads as risk.
- Is it right for a birthday, or a first date, or my parents? Occasion fit is what people are really asking. Say who the room is good for.
- Do you still have the dish I came for? Anything seasonal, off-menu, or sold out by nine. This one is asked about you specifically, by people already leaning toward yes.
- Can I book, order, or ask right now — at 11pm? Whatever the answer is at eleven at night is the answer for a third of the people who ever consider you.
Score it out of twelve. Then do the same on the map listing, since that's what most people see first. Nearly every restaurant scores between two and four, and every unanswered line is a question being answered somewhere else, about somebody else.
Written by the team at Amplified AI, an Ide8 company. We build Amp, an AI Revenue Manager for restaurants, so we plainly have a view on this — which is why the argument above is made without any numbers we can't source, and why the twelve-question test is worth running whoever you run it with.
