Issue #56: The Review Gap... The restaurants AI names most, and the ones it names once, are a tenth of a star apart...

This weekly dispatch is specifically designed for restaurant and small business owners who are trapped in the daily grind and ready to move from "struggling artist" to strategic operator.

One note first. Last week I promised you The Golden Hour. It ships next Tuesday. This one could not wait, and you will want the seven days.

You are at 4.6.

It took three years. It took answering the bad ones at eleven at night. It took the week you changed the coffee because two people in a row said the same thing.

The place four doors down is at 4.5.

You won. On the one number you have been watching, you won.

Now ask an assistant where to eat around here on Saturday. They get named. You do not.

For scale: the average restaurant brand is recommended by ChatGPT 5.3% of the time, by Gemini 7.1% and by Perplexity 7.6% [11]. Nobody is being named often. Somebody is being named more than you.

That is not bad luck and it is not a conspiracy. It is that you have been playing the scoreboard humans use, and the machine is reading a different one.

Both scoreboards are real. Both matter. They are not the same scoreboard.

Last week I told you this issue was coming, and I described it wrong.

I wrote that restaurant brands turn up in AI recommendations far more often when their review base looks a particular way, and that "the gap between the top and the bottom is not small." I had a figure in mind: 1% at one end, 53.5% at the other.

That figure is real. It is from Seer Interactive, published 14 May 2026, built on 804,491 AI responses across four assistants [2].

Three things are wrong with the way I put it.

One. 53.5% is not the top. It is near the bottom. It is the second-lowest of four tiers. The top tier is about 75%. The jump from nothing to 53.5% happens when a brand goes from no review profile at all to as few as one to thirteen reviews.

That is a better finding than the one I described. The money is at the very bottom of the ladder, not the top.

Two. It is not about restaurants. Eight verticals were studied. Money, shopping, travel, home services, health, education, electronics, business services. The word restaurant does not appear in it.

Three. It does not measure being recommended. It measures how often a brand turned up alongside a citation of one particular review platform. Close to what you want to know. Not the same thing.

So I am not building this issue on it. I am telling you it exists, what it actually says, and why it is not yours.

This took me one afternoon to check. It is the difference between a newsletter and a newsletter you can act on.

Here is the study that is actually about us.

Pluspoint published it on 24 August 2026 [1]. They asked ChatGPT, Gemini and Perplexity eight standard restaurant questions, twice each, across twelve cities. 551 answers. 1,917 different restaurants named.

Then they took the 36 restaurants named most often, and the 34 named exactly once, and looked at what was different.

Read that top row again.

A tenth of a star. That is the whole rating difference between the restaurants the machines cannot stop naming and the ones they mention once and forget.

Now read the bottom row. Among the most-named, one in nine had under a thousand reviews. Among the named-once, nearly half did.

The rating is flat. The volume is not.

It is American and it is small. Twelve US cities, 551 answers. Australia was not in it. Nobody has published an Australian version, and I looked.

It is a correlation and the authors say so themselves. Their words: the study "cannot say whether the photos caused it" or whether both follow from being popular. I would rather hand you an honest correlation than a confident invention.

And it is vendor research. Pluspoint sells review and listing management to restaurants. The study's conclusion is, conveniently, the product.

I am using it because it is the only dataset that is actually about restaurants. Not because the people who published it have no stake in the answer.

Now the other one, and this is where the rating earns its keep.

Michael Luca at Harvard Business School matched Yelp ratings against actual state revenue records for 3,582 restaurants. A one-star increase raised revenue between 5 and 9% [5].

And the detail everyone leaves out. That effect held for independent restaurants only. For chains it was statistically indistinguishable from zero.

The work is old — Seattle, 2003 to 2009 — and it is Yelp, not Google. It is a working paper, not a journal publication, and I am not going to dress it up as one.

It is also causal by design, and about exactly the venue you run.

So: the star rating is worth real money with humans and almost nothing with machines. The review count is close to the reverse.

You do not get to pick one. You have been running hard at one of them and calling it the whole job.

The evidence here is unsettled and I am not going to pretend otherwise.

Uberall, published 10 September 2026, found the same thing Pluspoint did — review volume predicts AI mentions better than rating does [6]. That is the newest large dataset available.

Yext, from December 2025, measured something adjacent and found reviews falling as a share of the sources AI cites [6].

Those two are not a straight contradiction. One is about who gets named. The other is about what gets quoted. A thing can matter more and be cited less at the same time.

But all of it is vendor research. Pluspoint, Uberall and Yext all sell something that gets easier to sell depending on the answer.

What I can tell you is this. The one study that is actually about restaurants and the largest multi-location dataset point the same way. The largest dataset of all points the same way again, and is not about restaurants. One vendor points elsewhere, at a different question.

Anyone selling you certainty on this has not read all four.

Everything above is somebody measuring from the outside.

Here is Google describing its own plumbing.

In the documentation for grounding Gemini in Google Maps, current as of 23 September 2026, Google states the service queries Maps for "places, reviews, photos, addresses, opening hours" across more than 250 million places [3].

Reviews. Photos. In a list of what the machine goes and fetches. Written by Google, not inferred by an agency.

And it is live here. Ask Maps — conversational search inside Google Maps, running on Gemini — rolled out in Australia on 7 August 2026 [4]. Google's own example in the Australian announcement is looking for a restaurant where you can hear each other talk.

That is nearly two months ago. Most operators I speak to have not registered it.

Put the two together. A diner in your suburb can now ask Maps a question in plain English, and the thing answering reaches for reviews and photos to build the answer.

Your review base stopped being a reputation asset. It became an input.

I am not going to hand you a target of 2,559 reviews. That number came from big-name restaurants in American cities and it would be a lie to put it in front of a forty-seat venue in Sydney.

Here is the only review number you control. Not the total. The rate.

Say you do 300 covers a week. That is 15,600 a year.

At one review per 200 covers, you earn 78 a year. At one per 50, you earn 312 a year.

Same room. Same food. Same rating. The only thing that changed is whether anybody asked.

Over three years that is 234 reviews against 936. One of those venues has a thin record. The other has a thick one. Neither of them got there by being better.

Label on that, stated plainly: the covers and the ratios are an illustration, not a measurement. Run it on your own numbers. The first prompt below does it properly.

And the reason it matters, in the one place it is measured: among the restaurants AI named most often, 11% had under a thousand reviews. Among the ones named once, 47% did [1].

Thin records lose. Not to worse food. To less evidence.

The photo gap in that study is the widest of the raw-count gaps. 3,979 against 1,044. Nearly four to one, and a wider separation than the review count.

Before you go and upload two hundred plating shots, two things.

One. A lot of those photos are not the venue's. Customers upload to a busy listing constantly. You cannot upload your way to 3,979 on your own — you can only be the kind of place people photograph.

Two. There is better evidence than a correlation, and it says something more interesting.

Zhang and Luo published in Management Science in 2023, working from 755,758 customer photos and 1.1 million reviews across 17,719 restaurants [8]. They found customer-posted photos predict whether a restaurant survives — beyond what the reviews alone explain.

Food photos carried the strongest signal. Then exteriors. Then interiors.

Read that as what it is. Photos are not a marketing lever you pull. They are a reading you take. A room people photograph is a room people are having something in.

You have probably seen this: businesses with photos get 42% more direction requests and 35% more website clicks. It is on hundreds of agency pages, usually attributed to "a study by Google and Ipsos."

I went looking for that study. It is not on Google's own help page, it is not in the Ipsos local search work it is credited to, and no version of it names a sample, a date or a market.

It appears to have escaped from a sales deck around 2016 and been copied ever since [10].

Delete it. The Management Science paper does the same job and it is real.

Three prompts. All three are new ground — nothing in the fifty-one from last week's Special Edition covers rate, photos or recency.

If you want the ones that already exist, run them alongside: The Review Gold Miner and The Human Reply Desk from #39, and The Sentiment Scanner from #49. Those handle mining, replying and root cause. These three handle arithmetic.

Set up the project. Call it The Review Desk. Upload: your last 90 days of covers by week from your POS; your Google review export; a screenshot or count of the photos on your listing; your brand notes; your house rules for replies.

Stop counting reviews. Start counting the rate.

This one turns your covers and your reviews into a single number — how many reviews you earn per hundred people through the door. Then it tells you where in your service the ask is going missing.

You need: ninety days of covers from your POS, and your Google review export.

You get back: a week-by-week table, and what your current rate adds up to over a year. Then three moments in your service where an ask could sit, ranked by how little they would cost your team.

Run this one first. It is the one that tells you whether you have a problem at all.

Your listing has photos on it. This works out what they add up to.

Not whether they are nice photos. Whether a machine reading your listing would find anything missing. And whether a diner who has never walked in could tell what walking in feels like.

You need: a count of the photos on your listing and who posted them, plus the totals for the two nearest comparable venues. Ten minutes of counting.

You get back: a breakdown across food, room, street and the practical shots people check before booking. Five shots to take this week. And one change to the room that would get guests photographing it themselves.

The cheapest of the three. Most of what it finds you can fix on a Tuesday afternoon with your phone.

Your all-time rating is an average of everything you have ever been.

This one shows you what you are now. It ages your review base and reads it the way a stranger reads it. A stranger is asking whether the place is still good, not whether it was ever good.

You need: your Google review export with dates.

You get back: your reviews sorted into four age bands, with your last-ninety-days rating sitting next to your all-time rating. Plus the one thing a guest reading only your recent reviews would wrongly believe about you today.

The uncomfortable one. Run it when you have twenty minutes and nobody needs you.

Pro Tip. Take the output and argue with it. If it names an ask moment your floor staff will never do at eight on a Friday, say so and get the next one. The model does the arithmetic. You know the room.

The ACT Model. Four weeks, one number.

The struggling artist polishes the score. They check it every morning. They get to 4.5, defend 4.5, and panic at 4.3. They answer the bad ones at eleven at night and they take it personally. Ask them how many reviews came in last month and they will not know, because nobody ever told them that was a number.

That training was correct, and it is still correct for the human reading your page and deciding where to take his in-laws. It is close to useless for the thing now standing between him and you.

The Savvy Operator builds the record. Not by buying anything. By asking, on purpose, at the moment the guest is already happy, week after week, until there is enough there to be worth reading.

Here is the whole thing in a line.

A rating is an opinion about you. A record is evidence that you exist.

Two weeks ago it was a machine cannot book a table that exists only in your head. This is the same sentence from the other side. A machine cannot recommend a venue it can find nothing to say about.

You are not being marked down. You are being skipped for lack of material.

One more thing, and then I will let you get back to it.

There is no Australian benchmark anywhere in this issue, and that is not an oversight.

I went looking for the average review count and the average rating for an Australian restaurant. Every figure I found was either unsourced, or attributed to the company publishing it with no sample and no method.

So there is nothing honest to hand you. The only review benchmark that exists for your venue is your own last ninety days. That is why every prompt above compares you against yourself and against nothing else.

Somebody should build the Australian one. It might end up being me.

Open your POS. Open your listing. Write down two numbers: covers last week, and reviews last week.

That is the whole first step and it takes four minutes.

Here is the honest problem with everything above. It is another job.

You already run a kitchen, a roster, a supplier list and a family. Nobody needs a fourteenth tab open.

So do not run it yourself. Set your AI up as a second brain for the business and let it hold the detail.

Not a toy. Not a chatbot you poke when you remember it exists.

A working folder that knows your menu, your prices, your hours, your reviews and your suburb. And the twenty questions your team answers on the phone every week.

Feed it those things once and it stops handing you ideas. It starts handing you a scoreboard.

That is why I built Strategic AI Marketing. We set your AI up properly, on your real numbers, so it thinks with your business instead of guessing at it.

You do the cooking. It does the remembering.

One, and it is free. Get your AI Visibility Snapshot. Four assistants, twelve questions about your venue, written down and dated. You will know exactly where you stand by the end of the week. It is at strategicai.marketing/free-ai-marketing-snapshot.

Two, if you want to know where your gap actually is. Reply to this email with one word: GAP, and tell me your venue name and suburb.

I will send back three numbers: your review count and photo count, and the same two for the two nearest comparable venues. Plus which of the three prompts above to run first, and why.

No pitch deck. No discovery call unless you ask for one. One reply, one useful answer.

Next Week on The Savvy Operator: Somebody Else Is Uploading Photos To Your Listing. In September, scammers were caught putting AI-generated images carrying fraudulent phone numbers onto business profiles through Google's own contributor programme. Google also started emailing owners when it detects a spike in spam reviews [12]. Two changes, one month, and what to check on your own listing tonight.

Until then, remember. The rating is what people think of you. The record is whether the machine can find anything to say. Serve your community. Be the operator.

Talk soon,

Rowan Shead

The Editor

The Savvy Operator

Owner of Strategic Ai Marketing

[1] Pluspoint. The 2026 Restaurant AI Visibility Study: 551 Answers From ChatGPT, Gemini and Perplexity, Dmytro Semonov, 24/08/2026. 551 AI answers (ChatGPT 189, Perplexity 192, Gemini 170) to eight standardised questions asked twice per engine per city, across twelve United States cities; 1,917 distinct restaurants named; data collected 30/07–03/08/2026. Comparison is the 36 most-named restaurants against the 34 named exactly once. United States data. Australia was not sampled and no Australian equivalent has been published. The authors state the study "cannot say whether the photos caused it" — it is correlation, and is presented here as correlation.

[2] Seer Interactive. Study of 800K AI Responses: How Review Profiles Shape Brand Presence in AI Search, Nick Haigler, 14/05/2026, in partnership with Trustpilot. 804,491 AI responses, 1,926 brands, 15,783 prompts, four platforms (ChatGPT, Google AI Mode, Gemini, Perplexity), data window March 2026. Eight verticals — none of them hospitality; the word "restaurant" does not appear in the study. Tier 0 (no verified profile) median citation rate 1%; Tier 1 (1–13 reviews) 53.5%; top tier approximately 75%. The metric is co-occurrence with a Trustpilot citation, not being recommended. Star rating, review recency and response rate were not tested as variables. Geography is not stated in the published write-up. Cited here to correct the description given in Issue #55, not as evidence about restaurants.

[3] Google. Grounding with Google Maps, Google AI for Developers documentation, current as at 23/09/2026. States the service queries Google Maps for relevant information including places, reviews, photos, addresses and opening hours, across more than 250 million places. Google's own technical documentation.

[4] Google Australia. Ask Maps, blog.google, 07/08/2026. Gemini-powered conversational search inside Google Maps, live in Australia in English. Note: the Australian announcement does not itself name reviews as an input — the reviews-and-photos list comes from the developer documentation at [3].

[5] Luca, Michael. Reviews, Reputation, and Revenue: The Case of Yelp.com, Harvard Business School Working Paper 12-016, 2011, revised 2016. Regression discontinuity on Yelp's half-star rounding, 3,582 restaurants, revenue from Washington State Department of Revenue records. A one-star increase raised revenue 5–9%. The effect was statistically indistinguishable from zero for chain restaurants (coefficient 0.005, standard error 0.025). United States (Seattle), 2003–2009, and Yelp rather than Google. An unpublished working paper — first issued 2011, revised March 2016, never published in a peer-reviewed journal. Causal in design (regression discontinuity) and specifically about independents, which is why it is here — but it is a working paper and is labelled as one.

[6] Uberall, reported by Search Engine Journal, 10/09/2026: 120,000+ AI mentions across 3,793 locations and five models, United States; review volume predicted AI mentions better than star rating. Vendor research, labelled sponsored by the publisher, and drawn from multi-location brands rather than independents. Yext, reported by iPullRank, 16/12/2025: 6.9 million citations across 1.6 million questions; review and social citations declining as a share of AI sources. Also vendor research, with no geography or period disclosed. The two disagree and both are named here for that reason.

[7] The Insiders. People trust People: How Online Reviews shape Consumer Choices in Australia, May 2023. 300+ Australian consumers; 75% consider reviews from the past three months crucial. Australian, but a small vendor whitepaper, three years old and retail-skewed rather than hospitality. Independently echoed in a different market by BrightLocal's Local Consumer Review Survey 2026 (11/02/2026, 1,002 United States adults), where 74% wanted reviews from the last three months.

[8] Zhang, Mengxia and Luo, Lan. Can Consumer-Posted Photos Serve as a Leading Indicator of Restaurant Survival? Evidence from Yelp, Management Science 69(1), 2023. 755,758 consumer photos, 1,121,069 reviews, 17,719 United States restaurants. Consumer-posted photos predict restaurant survival beyond what reviews explain; food photos carry the strongest signal, then exteriors, then interiors. Peer-reviewed, and explicitly predictive rather than causal.

[9] Google. Maps User Generated Content Policy — prohibited and restricted content. Merchants must not offer incentives such as payment, discounts or free goods in exchange for a review, must not discourage or prohibit negative reviews, and must not selectively solicit positive reviews. Australian position: the ACCC's guidance on online reviews, last updated 15/07/2026, states it is against the law to create fake or misleading reviews, that incentives must be offered regardless of what the review says and must be disclosed, and that suppressing or editing genuine negative reviews may itself breach the law. Recent Australian enforcement: PhotobookShop paid $39,600 in penalties on 24/03/2026 over 107 occasions of undisclosed influencer reviews and the editing of a review to remove negative comments.

[10] Claim checked and rejected: "businesses with photos receive 42% more direction requests and 35% more website clicks," widely attributed to a Google and Ipsos study. No primary source found. It is absent from Google's own Business Profile photo help documentation and from the Ipsos local search research it is credited to, and no version of it states a sample, date or market. Not used in this issue.

[11] SOCi. 2026 Local Visibility Index, restaurant breakout published 17/08/2026, as cited in Issue #54. The average restaurant brand appears in Google's local 3-Pack 24.3% of the time, and is recommended by ChatGPT 5.3%, Gemini 7.1% and Perplexity 7.6%. Restaurant-specific figures; SOCi's widely quoted cross-industry numbers are different.

[12] Search Engine Roundtable, 10/09/2026 (updated 11/09): a Google Maps contributor account was found uploading 150+ AI-generated photos carrying a fraudulent phone number across unrelated business profiles; the contributor was removed about a day after disclosure. Search Engine Roundtable, 22/09/2026: Google has begun emailing owners to say it has detected a spike in spam reviews on their Business Profile and removed them. Practitioner trade reporting, dated, not a Google announcement.

Strategic AI Marketing's own citation record — 0/12 on 26/05/2026 and every reading since — is published at strategicai.marketing/we-did-this/.