Do Google Reviews Actually Affect AI Restaurant Recommendations?

TLDR: No AI company publishes a formula for how reviews feed into a generated restaurant recommendation, so treat any specific weighting you read online with real scepticism. The best current evidence says review volume is loosely associated with whether an AI engine recommends a restaurant at all, star rating is only associated with which recommended restaurant comes first, and other signals, such as having a working website and accurate listing details, measured at least as strongly in the one study that checked. Reviews belong in your operating mix. They are not the single lever worth betting a budget on.

A restaurant-group marketing lead asks a reasonable question before spending more on review generation: does any of this actually move whether ChatGPT, Gemini or Perplexity names your restaurant, or are we just chasing a number because everyone says it matters?

Reviews sit between two extremes. They are not irrelevant, and they are not the dominant lever either. Below is what checked evidence actually shows, where reviews sit against other signals, and what that means for where you spend the next pound.

What do AI companies say about ranking reviews?

Start with the companies building these systems. OpenAI's own help centre describes ChatGPT's web search as ranking results "using multiple factors intended to help users find relevant, reliable information," without publishing the formula, and states that placement is never guaranteed. Google's guidance for appearing in AI Overviews and AI Mode goes further: it states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary" beyond the standard requirements for appearing in Search at all. Neither page names a review-weighting figure for its AI-generated answers.

That is a meaningful gap. Google's own local-ranking guidance does say plainly that "more reviews and positive ratings can help your business's local ranking," and that prominence is partly "based on info like how many websites link to your business and how many reviews you have." That statement is about Google's conventional local pack and Maps ranking. It is not confirmed as the same mechanism a chatbot uses to decide what to say in a generated answer. Treating the two as interchangeable is exactly the overstatement this question invites, and the evidence does not support it.

What did two 2026 studies find when they actually checked?

Statements from the platforms only go so far, so two 2026 studies tried to measure the question directly rather than describe intentions.

An August 2026 academic preprint, Invisible to the Machine, audited 4,776 food and drink venues against a complete market census across two markets, then logged 2,208 recommendations from ChatGPT, Claude, Gemini and Perplexity answering 96 queries. It found review volume raised a venue's odds of being recommended at all, while star rating showed no such association for a venue's entry into the recommended set. Rating only became useful once a venue already appeared, where it helped predict which venue came first. Roughly 86% of venues studied were never recommended by any of the four systems, and nearly three-quarters of venues with 50 or more reviews stayed invisible regardless.

A separate June 2026 study by local-SEO vendor Local Falcon checked 10,000 US restaurants and found 74.9% invisible in Google's AI-generated recommendations, against 19.6% invisible on Google Maps for the same restaurants. Restaurants with more than 1,000 reviews were still excluded from AI recommendations 70.9% of the time. That study is vendor-run and explicitly correlational, not causal, but it points the same direction as the independent academic audit: a large review count does not reliably buy a restaurant a place in an AI-generated answer.

Note what neither study claims. Neither shows that adding reviews causes a restaurant to appear, and neither disproves that reviews matter at all. They show that review signal alone is a weak and inconsistent predictor, next to other signals that this question usually gets asked in isolation from.

Where do reviews fit against other signals?

The Bali preprint is the only one of the two that measured several signals side by side using the same method, so treat the comparison below as evidence from one study, not a universal ranking. It still answers the real question: reviews are not competing alone.

Listing accuracy matters in a way raw numbers miss. The same preprint found permanently closed venues recommended 93 times across the study, a staleness failure rather than a reviews failure: the venue's listing data had not caught up with reality. A group with accurate, current structured data for every location is protecting itself against a different failure mode than the one review generation addresses. Worth noting: Google's own structured-data guidance for review and rating markup documents that markup for conventional rich snippets and Knowledge Panels, and does not mention AI Overviews, AI Mode or any generative feature in connection with it, another example of a well-documented mechanism that is not confirmed to be the same one behind a generated answer.

Press coverage and third-party mentions behave similarly to reviews in one respect: they are both forms of other people talking about your restaurant, rather than your own marketing describing it. That is likely why both showed a measurable, moderate association in the study above. Neither dominated the comparison.

What should this mean for a restaurant group's priorities?

Keep review generation in your operating mix. It still shapes real diners: a Harvard Business School study of Seattle restaurants from 2003 to 2009 found a one-star increase in Yelp rating associated with a 5 to 9% increase in revenue, concentrated in independent restaurants rather than chains. That is evidence about human readers responding to Yelp, not about how an AI model weighs a recommendation, and the two should not be conflated. But it is a real reason to keep asking for genuine reviews well, even before any AI-visibility question enters the room.

What changes is the priority order. If your group is choosing where the next hour of effort goes, the evidence above points to three practical moves, in order:

• Fix a broken or missing website before chasing another hundred reviews. It measured as the strongest single signal in the one study that compared several side by side.
• Keep location data (hours, address, menu, whether the branch is still trading) accurate across every platform. The same study found permanently closed venues still being recommended, a staleness failure, not a reviews failure.
• Treat reviews as one input you maintain continuously, not the one input you optimise in a short burst. A sudden spike looks exactly like the incentivised pattern regulators check for.

For the operational side of running that review process fairly and consistently across multiple venues, our review-management guide for multi-site groups covers the practical detail, including the same evidence limits on review causation. Read it for the how-to. Settling whether the underlying bet is sound comes first, before you commit more budget to it.

There is also a compliance reason to keep review management honest rather than aggressive. Genuine reviews, properly collected, are not just good practice; UK consumer protection law requires it. The Competition and Markets Authority's guidance for businesses publishing reviews requires a trader to take reasonable and proportionate steps to prevent and remove fake reviews and concealed incentivised reviews, even when a third party manages the review queue on the business's behalf. Outsourcing the task does not remove the legal responsibility for what appears under your name.

Where does Schmitdy fit into this decision?

None of the research above tells you what is actually happening to your own locations. A general study of 10,000 restaurants, or 4,776 venues in two Bali markets, is evidence about the shape of the problem, not a diagnosis of your group's specific gaps.

Schmitdy's free AI Search audit checks the signal mix for your own venues rather than assuming the general finding applies to you: current visibility across ChatGPT, Perplexity, Claude, Gemini and Google AI Overviews, the specific prompts your category is actually asked, which domains and pages get cited in your results, and a 90 to 180 day plan naming the levers that matter for your locations specifically, whether that turns out to be your review profile, your listing accuracy, your web presence, or more often, a combination none of those general studies can tell you about in advance. The audit is free, with no obligation afterwards.

Weigh that against the realistic alternatives before picking one:

Choose the venue most likely to need attention this week, request the audit, and compare what it actually finds for your locations against the general pattern above before deciding where the next investment goes.

Sources

1. OpenAI help centre: Searching the web with ChatGPT.
2. Google Search Central: how to appear in AI features.
3. Google Business Profile help: how local ranking works.
4. Google Search Central: review snippet structured data.
5. Invisible to the Machine: Auditing AI Restaurant, Cafe, and Bar Recommendation Against a Complete Market Census, Vladimir Pitenin, August 2026 preprint.
6. Local Falcon: The Restaurant AI Visibility Index, June 2026.
7. BrightLocal: Uncovering ChatGPT Search Sources, December 2024.
8. BrightLocal: consumers trust AI as much as reviews.
9. Harvard Magazine: HBS study finds positive Yelp reviews lead to increased business, reporting Michael Luca's Harvard Business School study "Reviews, Reputation, and Revenue: The Case of Yelp.com".
10. Competition and Markets Authority: short guide for businesses publishing consumer reviews.
11. US FTC: final rule banning fake reviews and testimonials, August 2024.
12. Germany's UWG, Annex to section 3(3), Nr. 23b and 23c.

The list above covers the 12 external sources checked. Our own review-management guide for multi-site groups, linked above, is an internal cross-link, not counted in that external total.