AI Search Runs on Questions, Not Keywords — What 126 Million Prompts Tell Us

Short answer: people don't type keywords into AI assistants. They type long, specific questions — often with four or five conditions bundled into one sentence.…

Short answer: people don't type keywords into AI assistants. They type long, specific questions — often with four or five conditions bundled into one sentence. If your website only answers one of those conditions, the AI will recommend a competitor who answers more of them. The fix is to write down plain answers to the questions you actually get asked.

Semrush has expanded its AI Visibility Index for 2026 to cover 126 million AI search prompts — the real things people type into ChatGPT, Perplexity, and Google's AI answers when they're looking for a business like yours.

The headline number isn't the interesting part. The shape of those prompts is.

Keywords stayed short. Prompts got long.

In the old search world, someone looking for a plumber in Leeds typed three words: plumber Leeds. That's what everyone optimised for.

Today, the same person types something closer to:

> "Who's a reliable plumber in Leeds who could come out on a Sunday for a burst pipe, and roughly what would that cost?"

Count what the AI now has to match: the trade, the city, weekend availability, an emergency, and a price expectation. Five pieces of information instead of one.

Now imagine your homepage says: "Leeds plumber — 20 years' experience, fully insured." That's one piece the AI can grab. Experience and insurance are lovely, but nobody asked about them. So the assistant recommends the plumber whose site happens to mention Sunday call-outs and a typical call-out fee.

That's the whole game. AI assistants don't rank pages the way search engines used to — they assemble an answer from details they can find. They can only use what you've actually written down.

Why "we cover all your needs" is now a liability

Most small-business websites are written to sound reassuring rather than to be specific. Phrases like competitive rates, flexible availability, and serving the surrounding areas were never great, but they used to be harmless. They aren't anymore.

An AI assistant reading "competitive rates" learns nothing it can repeat. An AI assistant reading "most kitchen taps replaced for £90–£140" has a sentence it can use in an answer — and a reason to name you.

Vagueness doesn't just fail to help. It actively removes you from the shortlist.

What to do this week: write your ten questions

Here's the exercise. It takes an hour and it works for almost any local or service business.

Step 1 — Collect the questions. List the ten things you genuinely get asked most often. Not keywords. Questions, in the customer's own words. Pull them from your phone calls, your inbox, your quote conversations, your DMs, your voicemail. Don't invent them — real phrasing matters, because it's closer to how people talk to AI assistants.

Step 2 — Answer each one in 40–80 words. Plain language, no marketing padding. One page is fine. Call it "Questions we get asked" and link it from your main navigation.

Step 3 — Be specific exactly where you'd normally hedge.

Price: a range beats a promise. "Most kitchen taps replaced for £90–£140."
Availability: "We answer emergency calls until 9pm, seven days a week."
Area: name the towns and postcodes you cover, not "Yorkshire and surrounding areas."
Turnaround: "Same-week appointments in most cases; emergencies usually within four hours."
What you don't do: "We don't fit boilers."

That last one feels like giving up business. It isn't. It stops you being recommended for jobs you'd only turn down, and — this is the part people miss — it makes every other answer on the page more believable. Specific limits are a credibility signal, to humans and to AI models alike.

Ten honest answers will do more for your AI visibility this month than another round of keyword tweaks.

For agencies: this is a repeatable deliverable

If you manage sites for clients, the question-mining exercise is one of the easiest wins to package. You already have access to the raw material — call logs, contact form submissions, live chat transcripts, the sales team's most-asked list.

Two things worth building into your process:

1. Interview the person who answers the phone, not the marketing contact. They know the actual questions. 2. Refresh the list quarterly. Questions shift with seasons, pricing changes, and new services. A stale FAQ page ages badly.

It's a clear, explainable deliverable for clients who've grown tired of hearing about rankings.

Then check whether it worked

The useful question is no longer "are we on page one?" It's "when someone asks an AI assistant for a business like ours, does our name come up?"

That's what a SimpleAIO audit measures. It runs the kinds of questions your customers actually ask across the major AI assistants and tells you — in plain English — which ones you're already being recommended for and which ones you're invisible for. It takes a couple of minutes, and it's usually the fastest way to see whether your new answers landed.