Search · original research
60 restaurant searches, zero AI Overviews
What actually decides whether a brand's own website shows up in search, and what changes when the customer asks an answer engine instead.
· 7 min read
Every marketing newsletter I read has told me the same thing for a year now. AI is eating search. Overviews are taking the clicks. Optimise for the answer engine or disappear.
I wanted to know whether that was true for the kind of business I actually work with, so I measured it instead of reading about it.
I ran 60 Singapore restaurant searches through Google and 21 of them through Perplexity, then classified all 371 domains that came back by hand. Whose site is this? The restaurant's own, a food blog, a directory, a forum, a news masthead. Every query, every domain and every classification call is published, including the parts where the data contradicted what I expected.
Three things came out of it. The first one surprised me. The second one is the one worth acting on.
Google showed no AI Overview on a single one of the 60 queries
Not a low rate. None.
I checked this on live pages rather than raw HTML, because an AI Overview is injected by JavaScript and does not exist in the page source. Then, because a result of zero is worthless if your detector is simply broken, I pointed the same detector at two queries that should definitely trigger one. “How to boil an egg” and “why is the sky blue” both produced an AI Overview, in the same browser, in the same session, minutes apart. The detector works. The 60 negatives are real.
Restaurant discovery in this market is still ten blue links and a map pack.
But the specific fear, that a generated answer is currently sitting above the results and absorbing the clicks for these searches, did not show up anywhere in 60 attempts.
What decides whether your website appears is the intent, not the AI
This is the finding I would act on tomorrow.
I split the 60 queries into four kinds of intent, 15 each. Occasion searches, where somebody needs a venue for a birthday or an anniversary. Category searches, the “best chilli crab” type. Location searches, “restaurants in this neighbourhood”. And practical constraints, things like dog friendly or private room or open on a Monday.
Then I counted how much of each result page belonged to a restaurant's own website.
| Intent | Restaurant's own site | Publishers |
|---|---|---|
| Occasion | 35.6% | 34.1% |
| Practical constraint | 31.8% | 34.8% |
| Location | 21.8% | 39.1% |
| Category, “best X” | 11.1% | 51.9% |
And the version of that which makes it concrete. How many queries returned no restaurant website at all, anywhere on the page?
| Intent | Queries with zero brand sites |
|---|---|
| Occasion | 0 of 15 |
| Location | 1 of 15 |
| Practical constraint | 2 of 15 |
| Category, “best X” | 7 of 15 |
Nearly half of “best X” searches contain no restaurant's own website anywhere. Every single occasion search contains at least one.
That gap is enormous, and it is much bigger than any gap I measured between Google and Perplexity. Which intent your demand sits in matters more than which search product your customer is using.
If your traffic goal is “rank for best pizza in town”, you are not competing with other restaurants. You are competing with food bloggers for a slot inside their listicle, and the honest move is to go and get into the listicle rather than to rewrite your homepage for the fourth time.
If your customers arrive by occasion or by constraint, the ground is already yours to take, and it was before anyone said the word AEO.
On answer engines, you are read even when you are not shown
Perplexity looks generous to brands at first glance. Restaurant websites take 41% of what it cites, against 21% on Google. Twice as good.
That number is wrong, and the way it is wrong is instructive.
Perplexity simply cites more sources per answer, about 16 against Google's 9. Compare like with like by looking only at the first 10 sources on each side and the advantage evaporates completely. 20.0% on Perplexity, 21.5% on Google. The top of an answer engine's source list is about as dominated by publishers as Google's first page is.
The difference is not that answer engines prefer brands. It is where the brands sit:
| Position in the citation list | Restaurant's own site |
|---|---|
| 1 to 5 | 20.0% |
| 6 to 10 | 20.0% |
| 11 and beyond | 78.3% |
The visible top of the list looks like Google. Everything past position 10 is overwhelmingly restaurants' own websites, and that part is collapsed behind a tab that most people never open.
Here is why that is not just a losing position with extra steps. Those sources are not rankings that nobody scrolls to. They are the input the model read in order to write the answer at the top. It does not weight a source by whether a human would have clicked it. A restaurant sitting at citation 14 still helped produce the paragraph the customer actually reads.
Being retrieved is a much lower bar than ranking, and it is worth real effort. Across all 21 matched queries, every single one pulled at least one restaurant's own website into the answer. On Google, three of them contained none at all.
What gets you retrieved is unglamorous. Say plainly what you serve, where you are, what it costs, how many people you can seat, whether you take walk-ins, what the room is actually like. Pages that answer a question in words a machine can lift are pages that get lifted. That is a very different objective from winning a listicle, and a far more achievable one for a single venue.
Two things I got wrong
Worth writing down, because both read well and both were false.
I started this because two test searches came back with no restaurant websites at all, and I assumed brands were shut out of Google entirely. Across 60 queries that is simply not true. Occasion searches run at 35% brand sites and not one of them was empty. I had generalised from two data points.
Then I assumed answer engines favour brands over publishers. True in the raw numbers, gone the moment I controlled for how many sources each engine returns. I nearly published the 41% figure. Checking it took ten minutes and changed the conclusion.
What this does not tell you
It counts citations, not covers. A citation at position 14 might be worth nothing commercially and this data cannot tell you otherwise. It is one answer engine, one market, one vertical, one day, 60 queries. It says nothing about ChatGPT or Gemini or Google's AI Mode.
The classification is judgment, so I published the judgment. Whether a hotel's website counts as the restaurant's own site is arguable. I argued it one way, wrote down why, and left the file where anyone can disagree with it.
The short version
Check what your customers are actually typing before you spend a cent on AEO. If they search by occasion or by constraint, the results already have room for you. If they search “best X”, your competitor is a food blog, and no amount of schema markup changes that.
None of this needed a new strategy. It needed counting.