Question wording decides whether Google writes an AI answer

Google decides, question by question, whether to write an AI answer

A study published on July 5, 2026 ran about 1,900 real questions through six AI engines. Google wrote an AI answer for 97.6% of the product-style questions in that set and for 3.0% of the restaurant-style local ones. The only difference between those two numbers is how the question was worded.

Google calls that block of AI-written text an AI Overview, and it appears at the top of the results page, above the usual list of links. Google makes a fresh call on every search about whether to write one at all. The research team at cloro sorted its 1,900 questions into groups by wording. Then it counted how often the block turned up.

Between the top row and the bottom row, the only thing that moved was the phrasing. On restaurant-style questions Google skipped the AI answer nearly every time, and the page led with the map and its three business listings instead.

Spoken-style is cloro's own label for the group. It means a question asked the way somebody says it out loud, rather than the three or four words they would punch into a search box. Anyone talking to ChatGPT is asking that way already. So is anyone using voice on a phone. Those questions pulled an AI answer 97.8% of the time, the highest rate in the study.

That call happens before your website comes into it, because Google is looking at the wording of the question and nothing else about you. Your reviews, your years in business, none of it has entered the picture yet.

The average everyone repeats describes nobody

The figure in circulation is that Google shows an AI answer on about half of all searches. cloro's own tracking ran higher than that. It watched a fixed set of 80 questions every week for 18 weeks (March through July 2026), and the blended number held in the mid-70s: about 69% in early March, a peak near 78% in May, about 67% in the most recent reading. So there are two averages doing the rounds, twenty points apart, and both of them are blends.

Four months of a flat-looking line on a chart. The flatness comes from averaging a 98 and a 3 together.

As a guide to your own business, either one is close to useless. Neither tells you whether your customers' questions land near the 98 or near the 3, which is the only part that changes what you do on Monday. Product-style questions pulled an AI answer 97.6% of the time in cloro's set. Restaurant-style questions pulled one 3.0% of the time. The owner whose customers ask the first kind and the owner whose customers ask the second get handed the same blended average, and neither of them is anywhere near it.

The 3% came from restaurant questions, and only restaurant questions

That 3% has now been measured twice, and both times on restaurant questions. The July 5 study put 400 restaurant-style questions in a group of their own. A second cloro study, Restaurant SEO in the AI Era, ran 200 dining questions through the same six engines on July 4, 2026, and Google's AI answer showed up on roughly 3% of those as well. Two separate question sets, published days apart, landed on the same answer. Neither set had a dentist's customer or an HVAC customer anywhere in it.

Restaurants are the only local trade anyone has measured this closely. Ricardo Batista (he founded cloro) calls his own finding one measurement from one set of dining questions, in English. That caution is worth reading twice.

I went looking for a study on dental questions or HVAC questions and could not find one.

Dentists, HVAC contractors, medspas, veterinary practices, law firms and in-home care agencies have never been through a test like this one. Nobody has published what Google does when a customer types the question a cracked tooth puts in their head. Or a dead furnace on a Sunday night. Maybe those questions look like restaurant questions to Google, and the AI answer stays away the way it did for restaurants. It's just as possible that Google reads them the way it reads a product question and writes an answer almost every time. Both readings are guesses until somebody types the questions in.

In the dining study, five of the six engines answered every single question. The number of sources behind those answers ran from around 8 for ChatGPT, Gemini and Copilot up to 21.6 for Google AI Mode, a separate Google feature that replaces the whole results page with AI-written text. Perplexity named restaurants on all 200 questions and attached zero sources to any of them. Nothing to click. The customer gets a confident list of names with no source to open and no way to check a single one, and they still walk away with three restaurants in their head.

So the honest read of cloro's work is narrower than the headline: it proves that the wording of a question decides whether an AI answer appears at all. Your own trade is still untested ground. That is exactly why somebody else's average is a poor stand-in for a check you can run yourself.

One finding does carry across every group cloro tested. ChatGPT, Gemini and Copilot wrote an answer 100% of the time, whatever the wording. Whether they name your business is a separate matter, and it is one you can settle in about ten minutes.

Run your own five questions this weekend

Here is the check. It takes ten minutes and costs nothing.

1. Write down five real questions. Pull them from the last month of calls, texts and front-desk conversations. Use the customer's own words (the ones they actually said on the phone). Something like: can someone look at a tooth that cracked last night.
2. Type each one into ChatGPT, Gemini and Google. Use the same wording every time, one question at a time. Do it signed out if you can (your own history changes what comes back).
3. On Google, write down whether an AI answer appeared at all. Some of your five will pull one. The rest will land on the map with its three business listings and nothing AI-written above them. That split is your own number, and no study can give it to you.
4. On every engine, look for your name. Where you're missing, write down who got named in your place. Those names are your competitors in AI search, whether or not they are your competitors on the street.
5. Run it again in a month. The answers move around, so a second reading a month later tells you which way yours is heading.

Write the five questions down before you start, and resist the urge to clean them up. The mother who asked whether you could see her kid before school on Thursday never said the words "pediatric dentist early appointment." The messy version is the one worth testing, because the messy version is the one that gets asked.

That typing hands you a list of your customers' real questions where the answer already belongs to somebody else. That list is the thing worth having.

The half-of-all-searches figure has been repeated enough that plenty of owners have quietly filed the whole topic under someday. Fair enough. Five questions and a phone will tell you more about where you stand than either of the studies in this post.