How Often AI Overviews Appeared for Cafe Searches
Short answer: I do not have a trustworthy rate to give you yet, and I would rather tell you that plainly than hand you a number I made up to sound useful.
A lot of owners ask me some version of the same question. They have read that AI Overviews are taking over Google search, and they want to know how often one actually shows up when a real customer types something like best coffee shop near me. It is a fair question. It also turns out to be much harder to answer honestly than most of the articles claiming to answer it let on.
I went looking for a real trigger rate before writing this. What I found instead was a pile of secondhand statistics from marketing blogs with no dated method behind them, and exactly one dated record from Hillcane's own work, a single Birmingham search recorded on September 9, 2026. One search is not a rate. It is one data point. So this post explains what a trigger rate actually is, what one dated search does and does not show, and gives you a method to run your own small test this week.
What I Actually Have on File
Before I get into method, I want to be specific about what exists in Hillcane's own research right now, because that is the whole point of being honest about a gap instead of papering over it.
Across the regional audit files I keep for the nationwide census, there is a section reserved for AI answer snapshots tied to each city. When I went through those files, the pattern was consistent: city after city recorded with no live AI panel, most often because Google was gating or blocking automated style access to the search results page at the moment the search was attempted. The one exception on record is a Birmingham search dated September 9, 2026, where an AI Overview panel was confirmed present.
One confirmed appearance out of a set of attempted captures across many cities is not a frequency. It tells you that AI Overviews can appear for a coffee related search, which nobody seriously doubts, but it cannot tell you how often they do, because the sample is effectively one. Anyone who publishes a specific percentage, forty eight percent of searches or forty percent of local queries, based on a study you cannot see the method for, is asking you to trust a number with no visible receipts. I would rather show you my one data point and admit its limits than dress it up as more than it is.
This is also why I am not citing any of the third party statistics sites that show up when you search for AI Overview trigger rates yourself. Several exist, several disagree with each other by a wide margin, and none of them are Hillcane's own dated, repeatable record. Under the rule I hold every Hillcane post to, a number only belongs in a post if it traces back to Hillcane's own data or to Google's own documentation. A blog post claiming a national percentage, with no visible query set, no visible date, and no way to check the work, does not meet that bar, no matter how confident the headline sounds.
It is worth being clear about what those recorded no live panel notes mean too, since it would be easy to misread them. They do not mean Google never showed an AI Overview for a coffee search in those cities. They mean the automated search attempt itself was blocked or gated at the moment it ran, which is a separate fact from whether the feature would have appeared to an ordinary person searching normally at that same moment. Treating a blocked search as proof of absence would be just as dishonest as treating one confirmed hit as proof of a high frequency, so I am not doing either.
Why a Trigger Rate Is Harder to Pin Down Than It Sounds
It would be convenient if AI Overviews behaved like a coin flip, showing up some fixed share of the time no matter what you searched. They do not work that way, and understanding why helps explain why a single confident number is usually a red flag rather than a useful fact.
Google's own developer documentation on AI features describes them as generated dynamically based on the query, the available information, and Google's own systems at that moment, not as a feature applied on a fixed schedule or percentage. That means the exact same search phrase, run in two different cities, at two different times of day, or on two different devices, is not guaranteed to produce the same result. A rate measured today, with today's query wording, on today's device and location, does not necessarily hold up next month, or even next week.
Query wording matters more than people expect too. Best coffee shop in a specific city, coffee shop near me open now, and where to get specialty coffee near a named neighborhood are three different shapes of question, and there is no reason to assume Google's systems treat them identically. A study that only tests one phrasing and reports it as the rate for cafe searches is quietly making a much narrower claim sound much broader than it is.
None of this is an excuse to skip measuring it. It is a reason to measure it carefully, with a fixed method, a fixed date, and a stated sample size, rather than trusting a single vague percentage floating around without a visible source.
There is also a practical reason automated capture attempts kept getting blocked across the regional files I checked. Google has an obvious interest in making sure the results a person sees are the results a normal browsing session produces, not results shaped by a script pretending to be a browser. That means any serious attempt to measure this has to look and behave like a real person searching, on a real device, in a real private window, one search at a time, rather than a fast automated sweep. That is part of why the method in the next section matters as much as the eventual number does.
A Method You Can Actually Run Yourself
Here is the protocol I would use, and the one I am planning to run properly and publish with real numbers once it is complete. You do not need any special tools to do a small version of this yourself, just a normal browser and about twenty minutes.
Pick one device and stick with it for the whole test, since results can differ between a phone and a desktop. Use a fresh private or incognito window so you are logged out and not carrying personal search history into the result. Clear location settings or confirm the location you want the search to run from, since local results respond to where the search appears to originate.
Run a small, fixed set of searches, the same day, back to back. A reasonable starting set is three phrasings: best coffee shop in your city, coffee shop near me open now, and where to get specialty coffee in your neighborhood or city. For each search, record four things: whether an AI Overview panel appeared at all, whether it named any specific shops, whether it linked to any outside sources, and whether you hit a captcha or blocked page instead of a normal result.
Do this across enough searches to say something real, even if that number is small. If you run nine searches and hit a captcha on three of them, report six completed searches and three excluded, rather than quietly dropping the blocked ones and reporting a rate as if it came from nine. The excluded count is part of the honest number, not a footnote to hide.
Take a screenshot of every result, whether an AI Overview showed up or not, and keep a simple log with the date, the device, and the exact search phrase typed. This matters for two reasons. First, if you ever want to compare your result to a later test, you need to know exactly what you searched and when, since a different phrasing or a different week can produce a different outcome. Second, a plain screenshot is the only real proof behind any number you eventually report, to yourself or to anyone else, and it is the same discipline Hillcane holds its own screenshots to before a figure goes into a published post.
Reading Your Own Result Without Overreacting
Once you have run your small sample, resist the urge to treat it as a verdict on the whole internet. A result from three searches, or even thirty, run on one day, in one place, tells you about that day, that place, and that exact wording. It is a real data point about your situation, not a national statistic you can quote as fact elsewhere.
If an AI Overview showed up and named shops, take note of which kind of shops got named and which kind of sources the panel appeared to be drawing from, in general terms. If nothing appeared at all, that is also a real and useful result, not a failure of the test. Zero is a valid outcome. It tells you that, at least for that query and that day, the traditional search results are still what a customer sees first, which has its own implications for what you should be spending time on.
The most useful thing a small test like this does is turn a vague anxiety, AI is taking over search, into a specific, checkable fact about your own market, on a specific date. That is worth more to you than a secondhand percentage from a source you cannot verify, because it is yours, it is current, and you know exactly how it was produced.
It also gives you a fair way to check back in later. Run the same three phrasings again in a month, on the same device, and compare the results side by side. If the outcome shifted, you have a genuine, dated observation about how the feature behaves for your specific market over time, which is exactly the kind of small, honest, repeatable evidence this whole approach is built around, rather than a single snapshot treated as a permanent truth.
What I Am Doing With This on Hillcane's Side
I am not going to publish a trigger rate in this post because I do not have one I trust yet. What I can tell you is the plan: a fixed thirty city sample, three query templates per city, run on one day, logged out, on desktop, with every completed search and every blocked search recorded openly, sample size and date sitting right next to any figure that comes out of it.
When that sample is complete, I will publish the real count, with its denominator, its date, and its limits stated plainly, the same way every other Hillcane data study works. Until then, the honest answer to how often AI Overviews appear for cafe searches is that nobody, including me, currently has a verified number worth repeating, and the useful thing you can do this week is not read one more blog post that claims otherwise, it is run six searches of your own and see what happens in your market.
Related Reading
More from Hillcane on this topic, plus the pages on the site that sit next to the work.
Frequently Asked Questions
How often does an AI Overview appear for a coffee shop search?
I do not have a verified rate to give you. Hillcane's own research has one confirmed AI Overview appearance, a Birmingham search dated September 9, 2026, against a set of other city attempts recorded with no live panel, often because Google was blocking automated access. One data point is not a rate, so I am not publishing a percentage until a proper dated sample exists.
Why can't you just quote one of the percentages other sites publish?
Because I cannot verify their method, their query set, or their date, and several of them disagree with each other by a wide margin. Hillcane's rule is that every number in these posts traces back to our own dated data or to Google's own documentation. A percentage with no visible receipts does not meet that bar, no matter how confidently it is stated.
Does Google say how often AI Overviews appear?
Google's developer documentation on AI features (the page titled AI Features and Your Website, accessed September 21, 2026) describes the feature as generated dynamically per query rather than applied on a fixed schedule or share of searches. It does not publish a trigger rate, so any specific percentage you see elsewhere is not sourced to Google directly.
What is the best way to test this myself?
Use one device, a fresh private window, and a fixed set of search phrasings run the same day. Record whether an AI Overview appeared, whether it named shops, whether it linked sources, and whether you hit a captcha instead. Run enough searches to report a real completed count, and note any excluded, blocked searches openly rather than dropping them.
Does query wording change whether an AI Overview shows up?
It can. Best coffee shop in a city, coffee shop near me open now, and where to get specialty coffee near a neighborhood are meaningfully different questions, and there is no reason to assume Google's systems answer them the same way. A study that only tests one phrasing and reports it as the general rate is making a narrower claim sound broader than it is.
Should I report zero if no AI Overview shows up in my test?
Yes. Zero is a valid, useful result. It tells you that for that query, on that device and that day, the traditional results are still what a customer sees first, which matters for where you spend your time. Record it the same way as a hit: the date, the device, the exact phrase, and a screenshot. Run the same three phrasings again a month later so you can compare two dated snapshots side by side.
Is a small sample of a few searches worth anything?
It is worth something specific to you, not a national claim. Three phrasings run carefully on one device, logged out, on one day, tell you something real about your own market at that moment. Put the sample size and date next to the result. For comparison, Hillcane's planned study uses a fixed 30 city sample with three query templates per city, and it still reports its denominator and date alongside any figure.
What happens if I keep hitting a captcha or blocked page during testing?
Count it and report it. If you run nine searches and three hit a captcha, your honest result is six completed and three excluded, not nine with the blocked ones quietly dropped. Hillcane's regional audit files ran into the same thing repeatedly: across a record of 793 prompts in 264 cities, most attempts were gated or blocked, which is why only one live overview, in Birmingham on September 9, 2026, is on file.
Will you publish an actual trigger rate eventually?
Yes. The plan on file is a fixed thirty city sample across three query templates, run on one day, logged out, desktop, with every completed and excluded search recorded. Once that is done, I will publish the count with its denominator, date, and limits, the same way every other data study in these posts is built.
Does an AI Overview appearing mean a shop was named or cited?
Not necessarily. A panel can appear without naming any business, or it can name several, and the only confirmed appearance on file here, a Birmingham search on September 9, 2026, is one data point. That is why the method in this post records four things per search: whether a panel appeared, whether it named shops, whether it linked outside sources, and whether a captcha blocked the run. Keep appearance and naming as separate columns in your log.
Work with Hillcane
I looked for a real number on how often AI Overviews show up for coffee shop searches and found one confirmed data point, not a rate. Here is the honest gap and how to test it yourself.
If you want help reading your own AI Overview or search results correctly instead of guessing, reach Hillcane at hillcane.co/contact or call (256) 384-2449.
Comments