🌟 In today’s Issue
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.
Three Questions... Three letters, three different jobs, and three questions you can ask tonight to find out whether the machines recommending restaurants in your suburb have ever heard of you.
Strategic Marketing:
The Invisible 83%... Uberall's four-market international benchmark — US, UK, France and Germany — found 83% of restaurant locations never appear in an AI recommendation. In the same data, 86% keep a Google presence [1]. A profile is not a recommendation…
The Second Opinion... A complete-market census landed three points away using a different method: 85.6% of venues are never recommended by any system [2]. Two studies. Opposite hemispheres. Same answer…
The Real Risk... The machines are not inventing things about your venue. Fabrication ran at 0.08% of mentions. But permanently closed venues were recommended 93 times [2]. They are faithfully repeating something that was true eighteen months ago..
Practical AI Implementation:
The Three-Question Audit... Pick the assistant you already use. Ask three questions. Get an honest answer about where you stand…
The Answer Block Writer... Short, straight answers machines quote word for word…
The Off-Site Evidence Map... The pages deciding your name, none of which you own…
Actionable Growth Tactic:
The Day-0 Baseline... Activation, Collaboration and Transformation, pointed at the number nobody in your kitchen has ever said out loud…
The Savvy Operator Mindset:
From Listing Keeper to Evidence Builder... The visible thing is never the thing. The thing with memory is…
The Ghost on the Footpath
It is 6:40 on a Thursday. Four people stand on the footpath two streets from your door, deciding where to eat. Work is done and they are hungry. Nothing is booked.
One of them lifts a phone and does not open Maps. Not Instagram either. She opens ChatGPT and types the thing people type now. Where should we eat around here, something good, nothing fancy.
Three names come back, with a reason under each one. They read it, nod, and walk the other way.
You are 200 metres away. Your kitchen beats two of the three, and your pasta was made that morning. None of it came up, because none of it was ever in the conversation.
Here is the part that stings. Nobody outbid you. A machine simply had more evidence about them than about you. It never tasted your food — it read what the internet had already written down. About you, the internet was quiet.
Sydney's average restaurant spend is $32.28 a visit [3]. Four people is about $129. Twice a week is $13,400 a year, walking past a room with tables free.
So you go home that night and ask ChatGPT about your own suburb. And there you are, second on the list. Your chest lets go.
That relief is the trap.
Ask again on Tuesday and you may be gone. Researchers put the same restaurant question to the same engine twice. It agreed with itself 46.7% of the time [4]. You were not recommended. You were rolled for.
Nobody runs a business on a coin toss they cannot see.
I Ran This On Myself. I Scored Zero
Before I ask you to do anything, here is my own number.
On 26 May 2026 I ran our own audit on Strategic AI Marketing. Four assistants. Three questions a real buyer would ask. Twelve chances for our own name to come up.
It came up in none of them. 0 out of 12.
I published it that day, on a page anyone can open, and I have published every reading since. Nine June: 0 out of 12. Seven July: 0 out of 12. Two August: 0 out of 12.
Sixty-eight days. Four readings. Still zero.
Then, on 30 July, we found out why. Our own web host was blocking three of the four assistants at the network level. Not the content. Not the writing. The front door.
Nine of our twelve zeros were our host. Three were us.
We did not edit the old scores. We added a column recording that on those dates, three of the engines could not reach the site at all. The number stands. You add context to a row, you never change what was measured.
I am telling you this for one reason. Is my website even readable by these things? is a question almost nobody thinks to ask, and I did not ask it either. I sell this work for a living, and I still spent 68 days staring at a zero and blaming the wrong layer.
Check the door before you redecorate the room.
The version I run on myself uses four assistants and gives me a score out of twelve. You do not need four. You need one, and three questions. Pick whichever assistant you already have open and start there. Everything that matters about the method is the same.
Strategic Marketing: Three Letters, Three Different Jobs
Most operators think this is one problem with one fix. Sort the Google profile. Tick the boxes. Done.
It is three problems that need three different kinds of work. Most restaurants do the first, skip the second, and have never heard of the third.
Picture a busy street on a Friday night.
SEO is your street sign. It gets you on the map and in the list, so someone already looking can find you and click. Address, hours, categories, menu, a site that loads before they give up. That was Issue #32. It still matters, and it is no longer the finish line.
AEO is the concierge's answer. Answer Engine Optimisation is a fancy name for a plain thing: getting your name to come up when someone asks their phone. They ask about gluten-free, about walk-ins at eight, about parking, and the machine answers. Nobody visits your site to find out. No short, clean answer anywhere, and the machine guesses or moves on.
GEO is the mate at the pub. Generative Engine Optimisation means being named inside a recommendation someone else's machine writes about you. Not a link. A sentence. Try the wood-fired place on the corner, people rave about the lamb. That sentence gets built from evidence scattered across the internet, and almost none of it lives on your website.
And underneath all three sits the thing I missed. Can the machines read your site at all? Some hosting setups block AI crawlers by default and nobody sends you a letter. If you are locked out, none of the three letters matter. It takes 20 minutes to check. It is either yes or no.
Now The Brutal Math.
Uberall benchmarked ChatGPT, Gemini, Perplexity, Copilot and Google AI Overviews across four markets — the US, UK, France and Germany. It found 83% of restaurant locations never appear in an AI recommendation. In the same data, 86% of those locations keep a Google presence [1].
Read those two numbers side by side. Nearly everyone has done the SEO, and nearly nobody is getting the GEO.
A separate team ran a complete-market census in a different hemisphere and published on 7 August. They found 85.6% of venues are never recommended by any system [2]. Three points apart, two methods, same verdict.
The work most operators finished in 2023 quietly stopped being enough, and nothing sent them a letter about it.
It gets tighter. Those engines name three to five businesses per query [1]. Not ten blue links. Not a page to scroll. Three to five. Across restaurant categories, the top three names take 53.4% of the share of voice [1]. That is not a long tail. It is a very short one, and it is closing.
Here is the finding that should change your Monday. Researchers analysed 551 AI answers and tracked which sources the engines leaned on. Reddit 58%. TripAdvisor 33%. The Infatuation 27%. Timeout 24%. YouTube 22%. OpenTable 20% [4].
Your website is not on that list. Neither is mine.
Your visibility is being decided on pages you do not own, cannot edit, and have probably never read.
That is uncomfortable, and it is also the most useful line in this issue. Once you know where the evidence actually gets gathered, you stop polishing the wrong surface.
One more, and it kills the excuse that this is about quality. In that study the restaurants named most and the restaurants named once both averaged 4.5 to 4.6 stars [4]. Identical. What separated them was the paper trail. The most-named carried a median of 3,979 photos against 1,044 [4].
The machines are not picking the better restaurant. They are picking the better-documented one.
That is not a food problem. That is an evidence problem.
Mentioned is not recommended.
Getting mentioned is now the easy part. Roughly four in five brands get cited at least once, and about one in seven get the primary recommendation — the one name in the answer [5]. Those are two different results, and they get reported as one all the time.
Cited puts you in the undifferentiated middle. Recommended puts you in the room. Count them separately.
The risk is not the one you expect.
Everyone worries the machines will make things up, and they mostly do not. That census found outright fabrication at 0.08% of mentions. Then it found systems recommending permanently closed venues 93 times [2]. Uberall found 68% of local businesses showing up incorrectly on missing, inconsistent or out-of-date information [1].
The machines are not inventing your venue. They are faithfully repeating something that was true eighteen months ago.
Old hours. A dish you dropped two menus ago. A phone number from the last fit-out. That is not a fear story, it is a maintenance story, and you already service a fridge.
Practical AI Implementation: Building Your Visibility Desk With Claude
You will not check this by hand every month. You will do it once, feel briefly sick, and never open it again — which is what happens to every audit that lives in someone's head.
So build it as a working folder instead. A Claude Project that knows your restaurant, remembers last month's score, and hands you this month's. It is the quiet operator in the back room again. Last week that operator looked after your guests; this week it watches the machines deciding whether new guests ever hear your name.
How to build it.
Create the Project... Open Claude. Call it The Visibility Desk or [Your Restaurant Name] Search Brain.
Feed the Brain... Upload your menu with prices, your full Google profile including every category and attribute, your last 50 reviews, your hours including public holidays, your function and dietary details, and a plain list of the questions your team answers on the phone. Built the GBP Brain from Issue #32? Add this to it.
Lock your three questions... This is the part that matters most and takes the longest to accept. Pick three questions a stranger would ask. One for your highest-value booking. One for the urgent, tonight, nothing-booked moment. One comparison question — who is the best… — phrased how a person speaks, not how a keyword sounds. Never use your restaurant's name in one. Is Nonna's any good measures nothing, because you just told it the answer.
Then never change them again. Not when the number is embarrassing, and not when a better question is obviously available.
Because if you can change the question, you can make the score say anything. And then it is worth nothing to you.
🤖 AI PROMPT #1: The Three-Question Audit
Three questions. One assistant, or four if you have the time. One honest read.
[TASK TITLE/GOAL] Build my Day-0 visibility baseline.
1. Role & Expertise (Function): You are an AI search visibility analyst with deep knowledge of answer engines, local discovery, and how restaurant recommendations get assembled. Your job is to tell me plainly whether the machines know my restaurant exists, and who is beating me.
2. Context & Background (Pre-loaded / Specific to Task): My Business: [Insert restaurant name, cuisine, suburb, what makes the room different]… My Ideal Customer Persona: [Insert who you want walking in and the occasion]… Specific Problem: I asked my three locked questions and pasted every answer below. Tell me where I stand… The Assistants I Used: [Name them — ChatGPT, Claude, Gemini, Perplexity, or whichever you tried]… The Answers: [Paste them all]…
3. Task Description & Output Requirements (Function & Modifiers): Build a Day-0 baseline using a Cited, Partial, Absent framework. Mark an answer cited only when my restaurant is named. Mark it partial when the answer describes a place like mine without naming me. Mark everything else absent. Report partial separately and never add it to the score. Give me one markdown table with a row per answer, my score as cited-out-of-total, the restaurants named more often than me, and the reason each was given. Under 500 words. End by naming the single biggest gap between me and the most-named restaurant.
4. Work Order: First, mark every answer cited, partial or absent… Then, total only the cited ones… Next, count every competitor named and rank them… Finally, group the reasons the assistants gave into themes…
5. Examples: Answer: "For a relaxed dinner in Newtown, locals rate Casa Verde for the wood-fired lamb and the courtyard." Finding: Casa Verde cited. Reason theme: signature dish plus outdoor space…
6. Warnings/What to Avoid (Modifiers): Do not guess at answers I did not paste. Do not count a partial as a citation. Do not soften the result to make me feel better. Do not recommend a fix here — this report only measures.
🤖 AI PROMPT #2: The Answer Block Writer
Short, straight answers a machine can lift, instead of paragraphs it skips
[TASK TITLE/GOAL] Write 20 answer blocks for my website.
1. Role & Expertise (Function): You are a plain-English writer with deep knowledge of how AI systems pull short factual passages out of a page. Your job is to turn what my team already knows into text a machine can quote without changing the meaning.
2. Context & Background (Pre-loaded / Specific to Task): My Business: [Insert restaurant name, cuisine, suburb, the details that matter]… My Product or Service: [Insert menu, dietary options, seating, parking, functions, hours]… Specific Problem: I have listed the 20 questions my team answers on the phone every week. My menu, hours and venue details are in this project… The Questions: [Paste your 20]…
3. Task Description & Output Requirements (Function & Modifiers): Write one block per question using a Question, Direct Answer, Supporting Detail framework. Each block opens with a complete sentence that answers the question on its own, without the heading above it. Each sits between 40 and 60 words. Each carries one checkable fact — a time, a price in Australian dollars, a seat count, a dish name. If someone screenshots just that block, it must still make sense. Give me a markdown table with a suggested page for each.
4. Work Order: First, rewrite each question the way a customer would actually say it… Then, write the answer as a standalone sentence… Next, add the detail that makes it useful… Finally, flag any block where you need a fact I have not given you…
5. Examples: Question: Do you take walk-ins on a Friday night? Block: Yes. [Restaurant] holds 6 of its 48 seats for walk-ins every Friday and Saturday from 5:30pm. The bar seats another 8 without a booking. Groups over 6 should book, as courtyard tables are held for reservations after 7pm.…
6. Warnings/What to Avoid (Modifiers): Do not write marketing copy. Do not use words like welcoming, passionate or exceptional. Do not invent a fact, a price or a policy — flag it and ask me. Do not write an answer that only makes sense under its heading
🤖 AI PROMPT #3: The Off-Site Evidence Map
Where the machines actually read about your suburb, and where you are missing.
[TASK TITLE/GOAL] Map my off-site evidence gaps.
1. Role & Expertise (Function): You are a local evidence strategist with deep knowledge of review platforms, community forums, listing sites and local media. Your job is to show me where the conversation about restaurants in my area happens, and where I am absent from it.
2. Context & Background (Pre-loaded / Specific to Task): My Business: [Insert restaurant name, cuisine, suburb]… Specific Problem: My baseline from Prompt #1 is in this project, including every source the engines cited. I need to know which off-site platforms carry weight here and where I am missing, thin or out of date… My Current Listings: [List every platform you know you are on]…
3. Task Description & Output Requirements (Function & Modifiers): Build a gap map using a Present, Partial, Absent framework. One markdown table, a row per platform, showing my status, what that platform feeds into an AI answer, and the single next action. Rank by likely impact, not alphabetically. Under 400 words. End by naming the one platform to fix this week and why.
4. Work Order: First, list every source cited in my baseline… Then, add the local platforms and community spaces that did not appear but matter here… Next, mark each Present, Partial or Absent… Finally, rank by how often that source showed up…
5. Examples: Platform: a suburb community forum thread on best dinner spots. Status: Absent. Feeds: named repeatedly in Perplexity answers as local opinion. Next action: ask three regulars who already post there whether they would mention us honestly. Never post as the venue…
6. Warnings/What to Avoid (Modifiers): Do not suggest buying reviews, posting as a customer, or creating accounts that are not real. That is fraud, the platforms catch it, and it breaches Australian Consumer Law. Do not suggest paid placement as a substitute for being present. Do not fill the table with a platform that has no real audience in my suburb
Pro Tip: Take this AI-generated content and add your personal touch. Change a word here, add a local reference there, include a quick story about a regular customer. The AI does the heavy lifting; you add the soul.
Actionable Growth Tactic: The Day-0 Baseline
You do not need a project. You need 40 minutes and a piece of paper. The ACT Model, pointed at the number nobody in your kitchen has said out loud.
Here's how it works:
Step 1: The Activation.
Action. Tonight, write down your three questions. Then pick one assistant and ask all three. Not four. One. The one you already have open, or the one you have been meaning to try.
Any of these will do:
ChatGPT — chatgpt.com…
Claude — claude.ai…
Gemini — gemini.google.com…
Perplexity — perplexity.ai…
All four are free to use and none of them needs anything installed. If you built the project above, you are already in Claude, so start there.
Three questions on one assistant is about ten minutes. Count how many times your name comes up. Screenshot every one where you appear, with the date visible. No screenshot, no citation.
Want the fuller picture? Ask the same three questions in a second and a third. You are welcome to do all four, and four assistants times three questions is twelve — which is exactly what I run on myself. But one is enough to start, and one done beats four you never got to.
Most operators come back with nothing. Zero is normal, and zero is the asset — it is the only way you will ever prove what changed.
Step 2: The Collaboration.
Action. Bring the number to a 15-minute huddle. Do not hide it. Then ask one question: what do people ask us on the phone that is written down nowhere? Let them talk. Your host knows the parking one. Your chef knows the gluten-free one. Your bartender knows what the Friday crowd wants. That list is your answer-block inventory and it cost you a quarter of an hour.
Step 3: The Transformation.
Action. Publish the first 10 answers this week, in plain text, on your own site, in blocks a machine can lift. Add 20 photos and ask five happy regulars for an honest review, scripting none of it. Then log what you did and the date it went live. First Saturday of next month, ask the same three and add the row.
Then keep the row. Same three questions, same assistant, first Saturday of every month. Append the new score, never edit an old one. If you miss a month, write "not run" and the reason, because a blank gap looks exactly like a hidden bad month.
Measure. Publish. Log. Measure again. Same discipline you already run on food cost, pointed at a different leak.
And I will say the part the industry will not. Nothing here works in 30 days. Days to weeks on a good run, for the engines that search the web live. Two to four months for anything consistent. Longer again for what the models learned in training. Anyone selling you 30-day AI magic is taking month one's money to spend month two inventing excuses.
If you would rather not do it by hand.
Three questions on one assistant is the free version, and it is a real measurement. Do it tonight and you will know more than most operators in your suburb.
The Snapshot is that same act, twelve times, written down and dated. We ask four assistants instead of one, including Google's AI answers, which is the one you cannot check properly by hand. And we put the date on it, so that next month means something.
That is the entire difference. It is free, and it is at strategicai.marketing/free-ai-marketing-snapshot.
Savvy Operator Mindset: From Listing Keeper to Evidence Builder.
The struggling artist keeps a listing. They filled in the fields once and believe the work is finished because the form was finished. When it goes quiet it is the economy, the weather, the new place on the corner. They maintain a record nobody reads and call it marketing.
The Savvy Operator builds evidence. They know a machine recommending a restaurant does what a nervous friend does — reaches for whatever proof is nearest. So they make sure there is a pile of it. Photos. Answers. Reviews. Real prices. Mentions in places they do not control, earned honestly, over months.
Here is the whole thing in one line, and it is not really about restaurants.
The visible thing is never the thing. The thing with memory is the thing.
Covers tonight are visible. The guest list is memory. Ads are visible. The Knowledge Moat is memory. Backlinks are visible. Training data is memory.
You cannot bribe a memory with metadata.
The Listing Keeper. | The Evidence Builder. |
|---|---|
Filled in the profile once, in 2023. | Adds evidence every week. |
Thinks the website is the destination. | Knows the website is one source among many. |
Writes paragraphs about atmosphere. | Writes short answers to real questions. |
Has never asked a machine about their own restaurant. | Knows the score, and the date it was taken. |
Blames the algorithm. | Reads what the algorithm reads. |
Hopes to be found. | Is impossible to leave out. |
Now the honest part, and it is the reason I am writing this rather than selling you something.
Most numbers above are international. Uberall's is a four-market benchmark across the US, UK, France and Germany. The 551-answer study ran across 12 US cities. The zero-click figure is US Google data [6].
The closest thing we have at home: Roy Morgan surveyed 14,646 Australians. 13.6 million of us — 58% of Australians aged 14 and over — used an AI tool in an average four weeks in the March 2026 quarter. ChatGPT alone reached 10.5 million, or 45% [7].
That is not a restaurant number. It is the size of the room the question is being asked in.
So do not take an international benchmark as your own. The only Australian number that matters for your restaurant is the one you measure yourself, and it costs you 40 minutes.
Mine was zero and I published it. Yours will probably be zero too, and that is not the bad news. Not knowing is the bad news.
Your Next Move: Three Questions
Write your three questions tonight. Open one assistant. Ask all three and count the times your name comes up. Screenshot every one.
Put the number and today's date on a piece of paper, and stick it where you will see it in a month.
That is the first honest measurement of your visibility you have ever taken.
How We Can Work Together
If you want this to work without adding another job to your already full plate, set up your AI like a second brain for the business. Not a toy. Not a chatbot you poke when you remember. A real working folder that knows your menu, your prices, your hours, your reviews, your suburb, and the 20 questions your team answers on the phone every week.
When your AI has the right inputs, it stops giving you fluffy ideas and starts giving you a scoreboard. It becomes the quiet operator in the back room, watching the machines that decide whether four people on a footpath two streets away ever hear your name. That is exactly why I created Strategic AI Marketing. It helps you set up your AI the right way, so it can think with your business, not just answer random questions.
Need help with my AI systems. Reply to this email.
Next Week on The Savvy Operator: The Golden Hour: How to Turn Your Slowest Shift into Your Most Reliable Profit Engine. As promised. We will use your quiet hours to build offers, systems and guest habits that keep cash moving when the room goes still.
Until then, remember. The machines are not tasting your food. They are reading your file. Serve your community. Be the operator.
Talk soon,
Rowan Shead
The Editor
The Savvy Operator
Owner of Strategic Ai Marketing
PS. You know another restaurant owner who's staring at empty tables right now wondering what they're doing wrong. They're not doing anything wrong. They just can't see what you just saw. Forward this newsletter to them. It takes four seconds and it might save them thousands. |
PPS. Those 37 restaurant owners I've worked with? They didn't come to me talking about "digital marketing strategy." They came to me talking about the hollow ache of a half-empty dining room on a Friday night. The 3am calculator spiral where you keep re-running the numbers hoping they'll change. The feeling of captaining a ship that takes on water faster than you can bail. |
If you're tired of fighting the digital war alone |
The Savvy Operator
References
[1] Uberall. (2026). 83% of Restaurants Are Invisible in AI Search. Published 07/05/2026 via Business Wire. GEO Studio benchmark across ChatGPT, Gemini, Perplexity, Copilot and Google AI Overviews. Four-market international benchmark — US, UK, France and Germany. Location sample size not disclosed by the publisher. Same source for the 3–5 businesses per query, the 53.4% share of voice, and the 68% incorrect-listing figure.
[2] Pitenin. (2026). Complete-market census of venue recommendation across AI systems. arXiv 2608.07069, published 07/08/2026. 85.6% of venues never recommended by any system; fabrication at 0.08% of mentions; permanently closed venues recommended 93 times.
[3] Tyro. (2026). Eat Pay Love 2026. Average spend per visit: Sydney $32.28, Melbourne $33.50. Australian data.
[4] Pluspoint. (2026). 2026 Restaurant AI Visibility Study. 551 AI answers across ChatGPT (189), Gemini (170) and Perplexity (192), 30/07/2026–03/08/2026, across 12 US cities. Source of the 46.7% self-agreement rate, the citation mix, the 4.5–4.6 star finding and the photo medians. US data.
[5] Birdeye. (2026). AI visibility research. ~80% of brands cited at least once; ~15% secure the primary recommendation position. [CONFIRM PRIMARY SOURCE BEFORE SEND]
[6] SparkToro, using Similarweb clickstream data. (2026). Zero-Click Search Study, January–April 2026. Reported by Search Engine Land 09/06/2026. 68.01% of US Google searches ended without a click, up from 60.45% in 2024. US data.
[7] Roy Morgan. (2026). Artificial Intelligence (AI) Tools Usage, March 2026. Roy Morgan Single Source, n=14,646 Australians aged 14+, January–March 2026.
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/.