Food Media Is Becoming ChatGPT's New Evidence for Recommending Singapore Restaurants

ChatGPT recommends Singapore restaurants by reading food media, not by ranking restaurant websites. When a diner asks where to eat, the answer is assembled from guides, review platforms and local publications retrieved at the moment of the question, and the restaurant’s own page is rarely the quoted source. Studies of US dining prompts from cloro, Pluspoint and MyPlace, reviewed on September 23, 2026, give this shift a market signal, while a Canlah AI audit of a Singapore restaurant group published on July 31, 2026 supplies a smaller first-party source map for the samples reviewed here.

The two engines read different Singapore publications. ChatGPT typed the Michelin Guide, Tatler Asia and Tripadvisor into its own search queries, while Singapore food blogs such as Seth Lui and Eatbook SG fed Gemini and were not cited by ChatGPT in the same 24 samples. In Singapore, where diners add halal, price and neighbourhood constraints to the question, one engine’s source list is a poor guide to another’s. Food media is the evidence layer for AI restaurant recommendations, not one interchangeable channel.

Key findings

• Editorial lists carry the largest share: in cloro’s July 4, 2026 study of 200 US dining prompts across six AI engines, editorial best-of publishers accounted for 40% of dining citations, user-generated content 32% and review or reservation platforms 25%.
• Citation, mention and recommendation are separate states: a restaurant can sit on a page the engine retrieves and still be left out of the shortlist, which a single visibility score cannot show.
• Canlah AI’s map covers two engines, not the market: its 48-probe Singapore audit split cleanly between ChatGPT and Gemini, but it did not probe Perplexity or Google AI Overviews and cannot see what either engine retrieved and discarded.
• A five-layer source model: review floor, editorial guides, review platforms, local food blogs and owned facts should be recorded separately, because each feeds a different engine and needs a different owner.

What the dining-prompt studies measured

cloro, an AI search monitoring vendor, ran 200 dining-intent prompts across ChatGPT, Google AI Overview, Google AI Mode, Microsoft Copilot, Gemini and Perplexity on July 4, 2026. The most-cited domains were Reddit at 22% of triggered answers, Google and Maps at 19%, OpenTable and Instagram at 13% each, Time Out at 10%, The Infatuation at 9%, Tripadvisor at 7% and the Michelin Guide at 4%. Google AI Overview triggered on about 3% of the prompts, and Perplexity answered every prompt in the sample without attaching citations.

Sources: cloro, Restaurant SEO in the AI Era

Pluspoint recorded 551 answers from ChatGPT, Gemini and Perplexity across twelve US cities from July 30 to August 3, 2026. The same question asked twice returned matching names 46.7% of the time, and the three engines agreed on 103 of 2,158 names, or 4.8%. Gemini cited Reddit in about three quarters of its answers, Perplexity read Tripadvisor in 86% and ChatGPT quoted city editorial while attaching Yelp business data to three of every four restaurants it named.

Sources: Pluspoint, How AI Picks Restaurants

MyPlace, a guest engagement platform, reported in February 2026 that AI-recommended restaurants averaged 3,424 Google reviews against 955 for comparable restaurants that were not recommended, and that star ratings above 4.4 had minimal effect.

Sources: MyPlace, AI Ranking Factors for Restaurants

Each study carries evidence the others do not: cloro covers six engines in one run, Pluspoint measures whether a repeated question returns the same names and MyPlace ties recommendation to review volume. All three publishers sell monitoring or hospitality software, so their findings are not neutral market consensus, and none measures Singapore. The shared pattern is still usable: the pages that decide a restaurant answer sit mostly off the restaurant’s website.

Why food media decides the shortlist

The answer is a shortlist, not a ranking

A diner receives a handful of names, and a restaurant outside that list is not compared at all. Because repeated answers matched less than half the time in the Pluspoint sample, presence is a rate across runs, not a position.

ChatGPT searches for publications by name

In the Canlah AI audit, ChatGPT typed a specific publication into its own search query in 13 of 24 samples, the Michelin Guide in eight and Tatler Asia in five. About 78% of the publications it named in those queries were cited in the final answer.

Most citations still come from generic searches

Roughly 64% of the sources ChatGPT cited across the same 24 samples were never named in any of its search queries. They came back from ordinary phrasing: restaurant websites, hotel pages and booking platforms that ranked for the question. AI visibility for a restaurant therefore runs on two supply lines, editorial placement and conventional search health.

Community sources are the least stable input

Promptwatch reported that reddit.com’s share of ChatGPT Search citations fell from an average of 3.83% between July 18 and August 7, 2026 to 0.52% between August 14 and 17, 2026, an 86.4% relative decline. The cause has not been established. A source that can disappear in a week is a poor basis for a commitment.

Sources: Search Engine Journal on the Promptwatch data

A first-party source map from Canlah AI

Canlah AI logged 48 probe records against a Singapore restaurant group, 24 on ChatGPT and 24 on Gemini, with 179 retrieved source fragments recorded on the Gemini side. Questions were real diner phrasings, such as sushi near City Hall MRT. The map was published on the Canlah AI restaurants page on July 31, 2026.

Singapore restaurant source map: 24 samples per engine, audit published July 31, 2026.

Source	ChatGPT	Gemini	Reading
Time Out Singapore	Named 3, cited 8	Retrieved 9, all 9 used	Strongest on both engines
Tatler Asia	Named 5, cited 6	Retrieved 5, all 5 used	Strong on both, ChatGPT-leaning
Michelin Guide	Named 8, cited 6	1 fragment	Feeds ChatGPT, barely reaches Gemini
Tripadvisor	Named 4, cited 4	Not observed in sample	Feeds ChatGPT only
Seth Lui	Not observed in sample	10 samples	Feeds Gemini only
Eatbook SG	Not observed in sample	8 samples	Feeds Gemini only
City Nomads	Named 0, cited less often than the three guides above	7 samples	Mostly Gemini
Restaurant’s own website	Named 2, cited 3	Retrieved 3, all 3 used	Reached directly by both

Source: Canlah AI probe archive for one Singapore restaurant group, published July 31, 2026. “Named” counts ChatGPT search queries that typed the source by name. “Not observed in sample” means the source did not appear in that engine’s 24 samples when they were logged; it is not proof that the engine never uses it.

The same audit shows the limit of the method: ChatGPT opened the group’s own concept page once across 24 samples, the page was not observed in Gemini’s 24 samples, and one occurrence is not a frequency. The audit could see what each engine searched for and cited, but not what it retrieved and then discarded. This is one restaurant group’s intents on two engines, not a Singapore market benchmark.

Two anonymised cases on the Canlah AI cases page add tension to the review-volume thesis. A high-end omakase restaurant was mentioned in 1 of 32 non-branded buyer questions (3%), against 11 of 32 (34%) for a peer in the same price band; neither holds a Michelin star. A 20-year restaurant group with 3,452 reviews rated 4.5 stars or higher received zero recommendations across 49 non-branded probes on three engines. The evidence shows that review volume and source presence measure different things; it does not prove which one produced a recommendation.

Food blogs and ChatGPT citations are not the same channel

The Singapore food-blog ecosystem ran almost entirely toward Gemini. Seth Lui appeared in ten Gemini samples, Eatbook SG in eight and City Nomads in seven, while ChatGPT cited Seth Lui and Eatbook SG zero times across its 24 samples. Gemini named a site in only four of its 24 search queries. The rest were generic rewrites of the diner’s question, so Gemini’s answer depended on whichever pages ranked for ordinary phrasing, where local blogs are strong.

Three sources reached both engines: Time Out Singapore, Tatler Asia and the group’s own website. The Michelin Guide and Tripadvisor carried ChatGPT visibility and left one Gemini fragment between them. An engine list is not a source list.

A five-layer source model for restaurant GEO in Singapore

Citation, mention and recommendation are separate states. In one Canlah AI sample, ChatGPT searched a guide and a review platform by name and then cited only the platform. Each layer below feeds a different engine and needs its own record.

Five source layers for Singapore restaurant GEO, mapped to the engine each layer feeds.

Layer	What to record	Decision it supports
1 Review floor	Google Business Profile review volume, rating, recency, response rate and primary category per outlet	Whether an outlet clears the shared input both engines read.
2 Editorial guides	Presence on Michelin Guide, Tatler Asia and Time Out Singapore pages, and whether ChatGPT searched each by name	Which guide or editor to pursue for ChatGPT.
3 Review and booking platforms	Tripadvisor, OpenTable or local booking listings: completeness, hours, menu and consistency	Whether ChatGPT’s operational facts about the outlet are correct.
4 Local food blogs	Coverage on Seth Lui, Eatbook SG and City Nomads for each cuisine, occasion and neighbourhood intent	Where Gemini’s generic searches will land.
5 Owned facts	Restaurant schema, halal status, hours and outlet pages on the restaurant’s own site, matched to every listing	Whether engines that read the site directly find the same facts.

Source: Canlah AI framework derived from the 48-probe audit published July 31, 2026. The layers are a record-keeping model, not a ranking-factor list published by any engine.

A longer outreach list is not the same as a source strategy. Without layer-level records, a lost question cannot be traced to its cause.

What changes in Singapore

A Singapore restaurant question often carries a constraint the engine must find evidence for. MUIS halal certification is a hard gate on an entire intent family, and it only works when the status reads the same on the restaurant’s site, its Google profile and its listings. A wrong primary category on a Google profile can remove an outlet from “best Italian” or “best omakase” questions before any content is read.

Multi-brand groups face a further problem. Engines resolve each outlet as its own entity, with its own reviews and coverage, so a flagship’s standing does not carry to sibling concepts. A source map built in another city names that city’s publications and does not transfer to Singapore.

What restaurant operators should test

• Freeze 15 to 30 non-branded diner questions covering cuisine, occasion, neighbourhood, price band and dietary constraint, and keep branded questions separate.
• Run each question at least twice per engine, spaced by several hours, then report a range rather than a single answer.
• Log every cited URL and, where the engine exposes it, every search query it typed, so named sources can be separated from generic retrieval.
• Record mention and citation separately for ChatGPT and Gemini before combining anything into a score.
• Retest on the same questions and label any before-and-after change as association unless a control design supports causation.

Where Canlah AI fits

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. For restaurants, it pairs engine-by-engine source mapping with the review, listings and publication work the map points to.

It is most relevant to restaurant groups that need to know which Singapore sources feed which engine, not to every outlet seeking a low-cost review tool. The company’s own restaurant evidence also shows the limit of that offer: its published map rests on one restaurant group, two engines and 48 probes, and it does not cover Perplexity, Google AI Overviews or Chinese-language engines for dining intent.

Methodology and limitations

Third-party figures come from cloro, Pluspoint, MyPlace and Search Engine Journal pages reviewed on September 23, 2026; they are vendor-published studies of US samples, not an independent Singapore benchmark. Canlah AI figures come from its 48-probe restaurant audit published on July 31, 2026 and its anonymised cases page. The Canlah AI data has no view of retrieved but discarded results and no control group, so it identifies source gaps, not revenue caused by AI recommendations.

Frequently asked questions

How does ChatGPT decide which Singapore restaurants to recommend?

ChatGPT searches the web when the question is asked and builds a shortlist from what it retrieves. In Canlah AI’s Singapore audit it often searched for the Michelin Guide, Tatler Asia and Tripadvisor by name, while about 64% of its citations came from generic searches. Review volume, editorial coverage and correct listings all feed that retrieval.

Do food blogs get cited by ChatGPT for restaurant questions?

No, not in Canlah AI’s Singapore sample. ChatGPT cited Seth Lui and Eatbook SG zero times across 24 samples, while Gemini drew on Seth Lui in ten samples and Eatbook SG in eight. Food blogs are a Gemini channel in that data, and Time Out Singapore and Tatler Asia reached both engines.

Can a restaurant GEO agency in Singapore guarantee an AI recommendation?

No. A restaurant GEO agency can guarantee its work, sampling cadence and reporting, but it cannot control how a non-deterministic model answers each diner question. Repeated answers in the Pluspoint sample matched less than half the time, so any guarantee of placement should be treated as a warning sign.

Related articles

• How AI engines choose sources
• How to check if ChatGPT recommends your brand
• Why is my business not showing up in ChatGPT
• Can a GEO agency guarantee results