OBENAN RESEARCH · EXPLORATORY FIRST EDITION
Search visibility is not AI visibility
On one day we compared the Google Maps comparison set for a restaurant question in Amsterdam, Berlin and Barcelona with what ChatGPT and Claude recommended for the same question.
Of the first three restaurants in each Google Maps comparison set, how many also appeared in the first three of both ChatGPT and Claude
0 of 3
Amsterdam
0 of 3
Berlin
1 of 3
Barcelona
Your customers now look in two places, and each one answers from a different list. Being easy to find is not the same as being recommended.
We asked Google Maps and two AI assistants, ChatGPT and Claude, for the best restaurants in three European cities on the same day, and compared every name they gave back.
The finding
The restaurants Google Maps put first were mostly not the ones the assistants recommended
Same question, same city, same day. We compared the first three restaurants Google Maps returned with the first three each assistant named.
Amsterdam
0 of the 3 matched
Neither ChatGPT nor Claude put any of the first three Google Maps results in its own first three.
Berlin
0 of the 3 matched
Both assistants named the same three restaurants, in the same order, and none of them was in the first three Google Maps returned.
Barcelona
1 of the 3 matched
One restaurant appeared in the first three Google Maps returned and in the first three of both assistants. It was the same restaurant in both.
One search per city, per assistant, on a single day. Enough to show what we found. Not enough to call it a rule. This is not a test of whether the answers stay stable over repeated runs.
How a comparison like that is counted
One search, two different lists
Illustration, not study resultsRestaurants shown on Google Maps
Restaurants recommended by AI (ChatGPT and Claude)
- Purple line = the same restaurant appears in both lists
- No line = the restaurant appears in only one list
Restaurant B appeared on both Google Maps and the AI list.
What this changes
AI visibility is not just search visibility with a new name
The two systems did not return the same restaurants for the same question. A business that a map search puts in front of people is not automatically a business an assistant recommends. Those are two different results, and this study measured both.
This is not a claim that the work behind search results stopped mattering. We did not test that. What we can say is that the two results came apart, so treating them as one thing to measure would have hidden the gap entirely.
The evidence they showed
The two assistants pointed at completely different kinds of evidence
ChatGPT linked the restaurants’ own websites in 46 of its 60 recommendations. Claude did it in none of its 60, in any city.
What each assistant displayed
Recommendation positions out of 60 per assistant.
Linked the restaurant's own website
Linked a guide or an award
Linked independent editorial coverage
Linked two or more separate third-party sites
Two more comparisons, on the same scale
Linked the Michelin guide itself
Linked Michelin or World's 50 Best
The counterexample
In Berlin they agreed completely, and still showed different evidence
In Berlin the two assistants named the same first three restaurants, in the same order. Not one of those three was in the first three Google Maps returned.
Claude showed no guide or award source anywhere in Berlin, across all 20 of its positions. ChatGPT showed one on 13 of its 20.
So two assistants reached the same answer while showing different evidence. What an assistant displays is not the whole of what moved it.
So a listing in a guide is not the thing to chase here. Two assistants arrived at one answer along visibly different paths, which is exactly why the links they display cannot be read as the reason for the answer.
The two assistants know the same restaurants. They disagree about who comes first.
Comparing ChatGPT with Claude, and not either one with Google Maps, both named 12 of the same restaurants in Amsterdam, 13 in Berlin and 10 in Barcelona, out of roughly thirty named in each city.
On the podium they parted. Berlin: 3 of 3, in the same order. Barcelona: 2 of 3. Amsterdam: 0 of 3.
Being known is not the same as being chosen. Both assistants may already have your restaurant on the list. That is not the same as either one naming you first.
Do the two assistants agree with each other?
Berlin
Same top 3, in the same order
ChatGPT and Claude compared with each other, and in the same order
Both assistants opened with the same three restaurants, first, second and third.
Against the first three Google Maps returned, the same city matched 0 of 3.
See all 20 positions for Berlin
A straight line means both assistants placed the same restaurant in the same position. A crossing line means they agreed on the restaurant but ranked it differently.
ChatGPT
Claude
Barcelona
2 of the top 3 matched
ChatGPT and Claude compared with each other, same first choice
They agreed on the same first choice, then the order came apart underneath it.
Against the first three Google Maps returned, the same city matched 1 of 3.
See all 20 positions for Barcelona
A straight line means both assistants placed the same restaurant in the same position. A crossing line means they agreed on the restaurant but ranked it differently.
ChatGPT
Claude
Amsterdam
0 of the top 3 matched
ChatGPT and Claude compared with each other, nothing in common
Neither assistant put any of the other one's top three in its own top three.
Against the first three Google Maps returned, the same city matched 0 of 3.
See all 20 positions for Amsterdam
A straight line means both assistants placed the same restaurant in the same position. A crossing line means they agreed on the restaurant but ranked it differently.
ChatGPT
Claude
What it means for you
Our interpretation: two different discovery jobs
What follows is our interpretation, not a measurement. What we measured is that the two lists diverged and that the assistants displayed different kinds of evidence. Why they diverged is a reading of that evidence, and other readings are possible.
Your own website is one source. The rest of what an assistant can see about you sits somewhere else entirely.
Google Maps
Is this a relevant, nearby, well-known local business?
The returned lists sat heavily in one central district in each city. Inside those lists, nothing on the business profile explained the order.
Be the clearest local match. Accurate details, the right category, a real presence in the area you actually serve.
ChatGPT and Claude
Is this a notable recommendation I can back up with public evidence?
Behind their first choices the assistants showed guides, awards, press and city sources. Claude never showed a restaurant's own website, in any city.
Be easy to verify. A clear site, one consistent identity everywhere, and third-party coverage that says the same thing about you.
What the evidence ruled out
Do more photos, replies or reviews move you up?
Across eleven checks, none of them explained which restaurants came first, and several pointed one way in one city and the opposite way in another.
Keep doing them for your customers if they help your customers. Do not expect them to decide your position.
Being easy to find on one surface did not carry over to the other. They are two related jobs, not one.
Where the Google Maps lists sat
Amsterdam
14 of the 20
in Amsterdam-Centrum
Berlin
18 of the 20
in Mitte
Barcelona
12 of the 20
in Ciutat Vella
These are also the districts with the most restaurants, and we did not collect a comparison group of places that never appeared. So this shows where the returned lists sat, not that a district was chosen.
What the assistants showed behind their first choices
8 of 9
ChatGPT showed a guide or award source
of its first-three picks, and a Michelin source in 7 of them.
9 of 9
Claude showed two or more independent sources
for every one of its first-three picks.
13 of 15
Guide or award recognition, in the assistants' own explanations
available first-three explanations named guide or award recognition as a factor.
This is what each assistant displayed, not what it used. Berlin is the reminder: Claude showed no guide or award source anywhere in that city and still named the same three restaurants, in the same order, as ChatGPT.
What we ruled out
What did not explain the result
We checked the things everyone assumes. If a simple explanation existed, three cities should have shown it. They did not.
| What we checked | Amsterdam | Berlin | Barcelona | What it tells us |
|---|---|---|---|---|
| Reviews collected all timeTop 3 compared with the other 17, in reviews | 2,472 fewer reviewsthan the other 17 | 1,331 more reviewsthan the other 17 | 4,252 fewer reviewsthan the other 17 | No consistent direction |
| Rating across all reviewsTop 3 compared with the other 17, in stars | 0.04 stars higherthan the other 17 | 0.04 stars lowerthan the other 17 | 0.10 stars higherthan the other 17 | No consistent direction |
| Reviews in the last 30 daysTop 3 compared with the other 17, in reviews | 8 fewer reviewsthan the other 17 | 29 more reviewsthan the other 17 | 20 fewer reviewsthan the other 17 | No consistent direction |
Each of these pointed one way in one city and the opposite way in another. Price and category showed nothing at all. So none of them explains which restaurants came first.
We tested five more explanations. None held across three cities.
More photos?
The first three had fewer in Amsterdam and more in the other two.
Amsterdam: first three 1,221, other seventeen 2,128 · Berlin: first three 3,841, other seventeen 1,262 · Barcelona: first three 7,124, other seventeen 3,073
photos, typical business
More kinds of photos?
No consistent pattern.
Amsterdam: first three 11, other seventeen 14 · Berlin: first three 12, other seventeen 13 · Barcelona: first three 17, other seventeen 15
photo categories, typical business
More replies to reviews?
Strong in Amsterdam, absent in Barcelona.
Amsterdam: first three 97 of every 100, other seventeen 53 of every 100 · Berlin: first three 41 of every 100, other seventeen none · Barcelona: first three 1 of every 100, other seventeen 3 of every 100
reviews that received an owner reply
More Local Guide reviewers?
No meaningful separation anywhere.
Amsterdam: first three 49 of every 100, other seventeen 51 of every 100 · Berlin: first three 63 of every 100, other seventeen 58 of every 100 · Barcelona: first three 42 of every 100, other seventeen 44 of every 100
reviews written by Local Guides
Different things mentioned in reviews?
We grouped every review keyword into eleven themes. No theme separated the first three from the other seventeen in all three cities.
Compared across the first three and the other seventeen in every city.
eleven themes, 412 distinct review keywords
Adding photos or replying to reviews may still be worth doing for your customers. This study did not find that either one explained which restaurants were returned first.
Why we dropped one of the checks
We could read the latest 100 reviews for each restaurant. Every one of the 60 reached that limit, so counting reviews inside a fixed window mostly measured how fast those 100 piled up. We removed that check rather than leave a number on the page that looks like it means something else.
The one that did not flip
One clue did hold up in all three cities
In all three cities, the top three had slightly better ratings across their latest 20 reviews. That is an interesting clue, not a proven rule.
Amsterdam
- Top 3
- 4.9 stars
- Other 17
- 4.7 stars
- Difference
- +0.2 stars
Berlin
- Top 3
- 4.7 stars
- Other 17
- 4.6 stars
- Difference
- +0.1 stars
Barcelona
- Top 3
- 4.8 stars
- Other 17
- 4.7 stars
- Difference
- +0.1 stars
Promising, but too small a sample to call a rule.
Each of those top groups is only three restaurants. It is the one thing that did not contradict itself across three cities, and it is what we are testing next.
What we still cannot say
- Each top group holds three restaurants, so none of the three differences stands on its own.
- We could read only the 100 most recent reviews for each restaurant. Where that limit was reached, older reviews are outside what we measured. That is also why the 90-day count is the weakest of the four checks above: almost half of these restaurants reached the limit inside 90 days.
- The order these restaurants came back in has not yet been independently confirmed.
Two places to check
Two checks, not one
Customers use both places, so it is worth checking both. Each check looks at one of them only, and neither one repeats this study.
AI assistants
See how AI describes your business
Check what AI assistants can find about you today. This check does not tell you where you appear on Google Maps.
Check AI presenceGoogle Maps
Check where you appear on Google Maps
Check where your business sits for one search, at one moment. It looks at your business only, and it does not repeat the three-city comparison on this page.
Check Google Maps visibilityHow we did it
How the comparison worked
Ask the same question
We asked Google Maps and the AI assistants the same restaurant question in each city.
Collect both lists
We saved every restaurant each one returned, exactly as it came back.
See which appear in both
We compared the two lists to find the restaurants named in both of them.
The technical detail, for anyone who wants it
- Isolated sessions
- Each assistant was asked once per city in a fresh, signed-out session.
- Repetition
- This is a one-day exploratory snapshot, not a test of whether the answers stay stable over repeated runs.
- Provenance
- Provenance means where the data came from. The exact query, settings, location context, and collection time are recorded for every list.
- Identity resolution
- Identity resolution means confirming that two names refer to the same business. Where that could not be confirmed, the pair stays unmerged and is reported as ambiguous.
- Uncertainty and unavailable evidence
- Anything that could not be verified is shown as unavailable rather than filled in with an assumption.
This method cannot show why a restaurant appeared, cannot establish a lasting pattern for either place, and cannot report a citywide Google Maps position.
The work behind it
What we actually analysed
This was not a one-prompt experiment. We asked two AI assistants the same question across three cities, captured every ranked recommendation rather than only the winners, asked each assistant to explain itself and kept those answers, matched the restaurants across every list by address and official domain, and read the public business profile behind each one.
In total
120,548
data values analysed
Every populated field across all 60 locations, drawn from 253 distinct data fields: 215 on the business profile and 38 on every single review. Roughly 2,009 values for each location.
What we asked
3
cities
Amsterdam, Berlin, Barcelona
2
AI assistants
asked the identical question
6
complete ranking lists
captured in full, not sampled
What we captured
120
ranked recommendations
every position recorded, not just the winners
5
follow-up explanations
we asked each assistant why, and kept the answers
60
public business profiles
rating, location, category, price, review history
What we read
6,000
individual reviews
read one by one across all 60 profiles
2,009
data values per location
on average, across every populated field
253
distinct data fields
215 on the business profile, 38 on every review
What we checked
238
source links
every link the assistants displayed
88
distinct source websites
guides, city sites, press, official pages
5
confirmed cross-list matches
and 3 uncertain ones we refused to merge
What we opened to check this
249,557
photos counted across the 60 businesses
178
distinct photo categories
412
distinct review keywords, grouped into 11 themes
60
business profiles read field by field
We did all of this for one reason. Not to find out which restaurants appeared, but to find out what the visible ones had in common, how AI recommendations differed from what a map search returned, and which explanations were still standing after three cities instead of one.
Before you use this
How to read this study
What this study does and does not prove
- This is a one-day, three-city exploratory snapshot.
- It reports observed patterns, not causation or a guaranteed ranking formula.
- Future editions will repeat the AI and Maps checks, and the conclusions may change.
How every statement is labelled
| Label | What it means |
|---|---|
| Measured | Counted directly from the checked evidence |
| Observed | Seen in one source, not a general rule |
| Model-stated | What an AI assistant said about its own answer |
| Inferred | Our reading, with other explanations named |
| Unavailable | We asked for it and could not verify it |
What comes next
We are running it again
One day is a snapshot. The next edition repeats the same question in the same cities, several times over, so we can tell a pattern apart from a coincidence. Restaurants are the first edition.