1 · What we asked
One question in each test
The same three questions, word for word, in the same 10 suburbs on both days: Surry Hills, Newtown, Bondi, Parramatta, Manly, Chatswood, Marrickville, Cronulla, Penrith and Balmain. On 10 Nov we used ChatGPT. On 14 Nov we used ChatGPT and Google AI Mode.
“What's the best café in your suburb?”
The cafés test. On 10 Nov ChatGPT named 41 cafés. On 14 Nov it gave 50 names and Google AI Mode gave 61.
Tift field tests, 10 and 14 Nov 2026
“Where can I get my e-bike serviced in your suburb?”
The e-bike service test. On 10 Nov ChatGPT named 40 shops and mechanics. On 14 Nov it gave 38 names and Google AI Mode gave 56.
Tift field tests, 10 and 14 Nov 2026
“Who's a good emergency electrician in your suburb?”
The emergency electrician test. On 10 Nov ChatGPT named 38 electricians. On 14 Nov it gave 34 names and Google AI Mode gave 50.
Tift field tests, 10 and 14 Nov 2026
2 · The numbers
What we found over two days
82 of 119
10 Nov names that ChatGPT named again on 14 Nov. Google AI Mode, asked for the first time on 14 Nov, named 66 of them. One tool or the other named 93.
All three tests, 14 Nov 2026
71
Names both tools gave on 14 Nov, out of 122 from ChatGPT and 167 from Google AI Mode
All three tests, 14 Nov 2026
33 of 41
Cafés ChatGPT named again, the most of any trade. Electricians were the fewest, 20 of 38.
Cafés and emergency electrician tests, 14 Nov 2026
8 of 10, then 5 of 10
E-bike answers in ChatGPT that gave a price, 10 Nov then 14 Nov. Google AI Mode, 14 Nov only: 4 of 10 answers, not counting ads.
E-bike service test, 10 and 14 Nov 2026
0 of 10, then 3 of 7 we could check
Café answers in ChatGPT that listed what looks like a café's own website among their sources, 10 Nov then 14 Nov. Google AI Mode, 14 Nov only: 2 of 10 answers linked a café's site, Instagram or Facebook.
Cafés test, 10 and 14 Nov 2026
The three 10 Nov pages name 119 businesses: 41 cafés, 40 e-bike shops and 38 electricians. The e-bike page counted 42 on the day, but its Parramatta answer ended with “and 2 more”, so 40 are recorded by name and every e-bike comparison here uses 40. Google AI Mode has no 10 Nov run, so its repeat figures show how much of the 10 Nov list it also named, not how much changed. A business is counted once for each answer it appears in.
3 · How we tested it
The method and its limits
Pick 10 suburbs
A spread across Sydney: inner city, east, west, north and south. The same 10 in all three tests, on both days.
10 Nov: ask ChatGPT one question per suburb
The same question in each suburb, in ChatGPT on 10 Nov 2026, from a signed-in free account in a temporary chat.
14 Nov: ask the same questions again, in two tools
Word for word, one run per question per tool. ChatGPT in a temporary chat, not signed in. Google AI Mode signed in, with personalisation off. It was run for the first time that day.
Record who was named and what was listed as sources
Every business named in the reply, and the sources each answer listed. We matched each 10 Nov name by reading the 14 Nov answers. Screenshots were kept on 10 Nov. The 14 Nov answers were saved as text, not screenshots, and are available on request.
Check what held
For each 10 Nov finding, we counted whether it held in the 14 Nov ChatGPT answers. Google AI Mode has no 10 Nov run, so we report its 14 Nov figures only. Source types were read from source titles and link hostnames, not from opening the pages.
The limits
Two days, two tools, one run per question per day, run by us. ChatGPT was signed in on 10 Nov and not signed in on 14 Nov, so some day-two change may come from that, not from the answers moving. Google AI Mode was first run on 14 Nov, so it has no 10 Nov baseline. One run cannot separate a real change from ordinary variation, and AI answers change. For the cafés test on 10 Nov we also tried Perplexity logged out, which stopped after 2 questions, so it is not in these results.
The three tests, 10 Nov and 14 Nov 2026
Cafés, Hospitality
We asked: “What's the best café in your suburb?” ChatGPT named 33 of the 41 cafés from 10 Nov again. Held in ChatGPT: a guide, list or review site among its sources (10 of 10 answers on 10 Nov, then 7 of 7 answers whose sources were saved). Did not hold in ChatGPT: no café's own website among the sources (0 of 10 on 10 Nov, then 3 of 7 answers we could check listed what looks like one). Google AI Mode, 14 Nov only: it named 19 of the 41, and 2 of 10 answers linked a café's site, Instagram or Facebook.
The finding: cafés were the trade ChatGPT repeated most.
E-bike service, E-bikes
We asked: “Where can I get my e-bike serviced in your suburb?” ChatGPT named 29 of the 40 e-bike shops again. Held in ChatGPT: every answer asked the bike's brand (10 of 10, then 10 of 10). Did not hold in ChatGPT: on 10 Nov, 8 of 10 answers quoted a price from a shop's own website. On 14 Nov, 5 of 10 quoted a dollar price for a named shop. We did not check where each 14 Nov price came from. Google AI Mode, 14 Nov only: it named 29 of the 40, and 4 of 10 answers quoted a price, not counting ads.
The finding: in ChatGPT the price finding and the no-guides finding did not hold, as 4 of 10 answers listed a guide or local outlet. Google AI Mode has no 10 Nov run, so its 29 shows overlap with ChatGPT's 10 Nov list, not stability.
Emergency electrician, Trades
We asked: “Who's a good emergency electrician in your suburb?” ChatGPT named 20 of the 38 electricians again. Held in ChatGPT: every top pick advertised 24/7 (10 of 10, then 10 of 10), a 4.9 or 5.0 rating showed for 9 of 10 top picks on 10 Nov and 10 of 10 on 14 Nov, and every answer asked what the emergency was (10 of 10, then 10 of 10). Google AI Mode, 14 Nov only: it named 18 of the 38. It said 24/7 in the opening line of 10 of 10 answers and in the top pick's own entry in 6 of 10, and showed a rating for 7 of 10 top picks.
The finding: electricians were the trade ChatGPT repeated least. No ChatGPT answer listed a guide or magazine among its sources, 0 of 10 on both days.
4 · What held
What held in ChatGPT, what did not, and what Google AI Mode showed on 14 Nov
Held in ChatGPT (10 Nov, then 14 Nov)
E-bikes: every answer asked the bike's brand, 10 of 10, then 10 of 10
Electricians: every answer asked what the emergency was, 10 of 10, then 10 of 10
Electricians: no guide or magazine among the sources, 0 of 10, then 0 of 10
Electricians: every top pick advertised 24/7, 10 of 10, then 10 of 10
Electricians: top picks showed a 4.9 or 5.0 rating, 9 of 10, then 10 of 10
Cafés: a guide, list or review site among the sources, 10 of 10, then 7 of 7 answers we could check
Google AI Mode, 14 Nov only
Google AI Mode gave 167 names on 14 Nov, 71 of them also given by ChatGPT, which gave 122
E-bikes: 10 of 10 answers asked the bike's brand or motor system. 0 of 10 listed a guide or local outlet among the sources
Electricians: 10 of 10 answers asked what the emergency was. 0 of 10 listed a guide or magazine
Electricians: 24/7 in the opening line of 10 of 10 answers, and in the top pick's own entry in 6 of 10. A rating showed for 7 of 10 top picks
Cafés: a guide, list or review site among the links in 1 of 10 answers. 8 of 10 answers had no outside links
Did not hold in ChatGPT (10 Nov, then 14 Nov)
Cafés: “no café's own website among the sources”. 0 of 10 on 10 Nov, then 3 of 7 answers we could check listed what looks like the café's own site. Google AI Mode, 14 Nov only: 2 of 10 answers linked a café's site, Instagram or Facebook.
E-bikes: on 10 Nov, 8 of 10 ChatGPT answers quoted a price from a shop's own website. On 14 Nov, 5 of 10 quoted a dollar price for a named shop. We did not check where each 14 Nov price came from. Google AI Mode, 14 Nov only: 4 of 10, not counting ads.
E-bikes: “no guides or magazines among the sources”. 0 of 10 on 10 Nov, then 4 of 10 ChatGPT answers listed a guide or local outlet. Google AI Mode, 14 Nov only: 0 of 10.
5 · From our work
The short version
Challenge
Which businesses do AI tools recommend in Sydney, and why?
Action
One question per suburb in each of three tests, 10 suburbs. ChatGPT on 10 Nov, then the same questions in ChatGPT and Google AI Mode on 14 Nov. One run per question per day, every answer recorded.
Result
ChatGPT named 82 of the 119 names again, cafés most and electricians least. Google AI Mode, asked for the first time on 14 Nov, named 66 of them. One tool or the other named 93. The two tools shared 71 names on 14 Nov. In ChatGPT some source patterns held, and three did not.
6 · What an owner can do
What the tests suggest, and your questions
What can a café do?
In ChatGPT, guides, lists and review sites were listed among the sources in all 10 answers on 10 Nov and in all 7 checkable answers on 14 Nov. Google AI Mode, 14 Nov only: a guide in 1 of 10 answers and no outside links in 8 of 10. Keep your Google profile accurate as well.
What can an e-bike shop do?
Say which brands and motor systems you service, with prices. In ChatGPT every answer asked the bike's brand on both days. ChatGPT quoted a price from a shop's own website in 8 of 10 answers on 10 Nov, and a dollar price for a named shop in 5 of 10 on 14 Nov.
What can an emergency electrician do?
Say 24/7, your suburbs, your licence and your callout terms clearly. In ChatGPT every top pick advertised 24/7 on both days, and every answer asked what the emergency was before choosing.
Can I see the raw data?
Yes. The 10 Nov names and sources for every suburb are on each full test page, and we'll share the 10 Nov screenshots on a call. The 14 Nov answers were saved as text, not screenshots. The raw answers are available on request.
Why did so many names change between days?
We cannot say. We ran each question once per tool per day, and ChatGPT was signed in on 10 Nov and not on 14 Nov. Ordinary variation, the sign-in change and real change all fit the numbers.
Will you run this for my business?
Yes. Tell us your business and suburb and we'll check what AI tools say about you.
Will you run it again?
Yes. Gemini and Perplexity are next, then we repeat every few months, because AI answers change.