AI is changing hospitality growth in two separate ways, and most restaurant groups are resourcing only one of them. The first is AI running inside the business: forecasting demand, building rotas, tracking inventory, drafting replies to guest messages. The second is AI acting as the discovery layer, deciding whether a diner who has never heard of your group ever gets your name put in front of them. Nearly every AI vendor pitch to hospitality groups right now is about the first. The second is the one with no clear owner on most org charts, and it is the one compounding every month it goes unmanaged.
TL;DR
- AI inside the business (forecasting, rotas, inventory, guest comms) and AI as the discovery layer (what ChatGPT, Google AI Overviews, Perplexity and Gemini tell a diner about you) are different problems with different owners and different metrics. Most groups only staff the first.
- 22% of US consumers have already used an AI tool to help choose a restaurant, per DoorDash's March 2026 survey, and AI use for local business recommendations generally jumped from 6% to 45% in a single year, per BrightLocal.
- Multi-unit groups are structurally more exposed than single venues: more entities to keep consistent, brand-versus-venue confusion, new openings starting from zero, and venues cannibalizing each other's AI answers.
- A worked, assumption-labeled example for a group of 7 venues puts the plausible annual revenue at stake in the low tens of thousands of pounds, not a headline ROI figure.
- This sits alongside your existing PR or brand agency, not instead of it, and needs a named internal owner from day one.
What's the actual difference between AI inside the business and AI as the discovery layer?
AI inside the business takes decisions you already make (forecasting a Friday night, scheduling front of house, reordering stock, drafting a guest reply) and does them faster or more consistently. It touches operations that already have a customer. AI as the discovery layer sits earlier: it decides whether a guest who has never booked with you shows up at all, because it answers "best Sunday roast near Borough Market" or "where should we take the team for a group booking in Manchester" before your website or booking system gets a look.
| AI inside the business | AI as the discovery layer | |
|---|---|---|
| What it affects | Forecasting, rotas, inventory, guest messaging, existing-guest experience | Whether a new guest ever hears your name, across ChatGPT, Google AI Overviews, Perplexity, Gemini |
| Who typically owns it today | Operations director, general managers, a dedicated ops/tech lead | Nobody specific in most groups; falls between marketing, PR and IT |
| How you measure it | Labor cost %, waste %, stockouts, response time, covers forecast accuracy | Named-in-answer rate per venue per engine, share of citations against named competitors, wrong-branch recommendations |
| Who sells into it | POS, EPOS, workforce and inventory software vendors, well established | AI visibility trackers and GEO services, a newer and less understood category |
| Failure mode if unmanaged | Higher costs, wasted stock, scheduling chaos | Guests you never see, because the group was never in the answer |
The first row has been sold into hospitality for years; every group runs some version of it. The second row barely existed as a category eighteen months ago, and the buying decision doesn't map onto any existing budget line, which is a large part of why it goes unowned.
What actually changed in guest discovery?
Three things changed, all inside the past year, all with dates attached.
First, the behavior arrived at scale. DoorDash's 2026 Restaurant Industry Trends Report, a survey of 3,001 US consumers and 509 operators run by Dynata in March 2026, found 22% of consumers have already used an AI tool such as ChatGPT or Google Gemini to help choose a restaurant. BrightLocal's Local Consumer Review Survey, published March 2026 from a survey of 1,002 US adults, found AI use for local business recommendations jumped from 6% in 2025 to 45% in 2026, with ChatGPT used by 31% and Google's AI Mode by 23%. AI is now the third most used local discovery channel, ahead of Yelp and TripAdvisor. Both are US figures, so they describe your market directly.
Second, we now know where these engines pull the answer from, and it isn't evenly spread. Yext's research on food service citations, drawn from 6.8 million citations across ChatGPT, Perplexity and Gemini and published in October 2025, states that food service "presents a unique challenge, with a significant dependency on both listings (41.63%) and reviews (13.28%)", the highest user-generated-content reliance of any industry Yext measured. Yext publishes no first-party website figure for food service, so we won't invent one. The operational read holds anyway: over half of what an engine cites about your venues sits on platforms you don't own.
Third, we have a working model of how the engines decide what to surface, and it rewards depth most owners don't have. Google's own documentation on AI features confirms that AI Overviews and AI Mode use "query fan-out": a single question gets split into several related sub-questions, run in parallel, before the answer is assembled. A December 2025 study by Surfer SEO, analyzing 10,000 keywords and 173,902 URLs, found pages that rank for several of those fanned-out sub-queries are 161% more likely to be cited than pages ranking only for the original question. The same study found about 68% of cited pages didn't rank in Google's top 10 for the main query or for any fan-out query, which is the part worth sitting with: ranking is not the gate here. For a restaurant group, holding position for "best restaurant in Leeds" says little about whether you're cited for the sub-questions an assistant actually runs, like "which Leeds restaurants take large group bookings on a Friday" or "vegetarian-friendly fine dining near Leeds station."
The scale of the gap is worth stating plainly. Uberall's Fast Food, Faster Discovery: 2026 GEO Playbook, reported in May 2026, found 83% of restaurant locations entirely invisible in AI-generated recommendations, though 86% hold some presence on Google. Being findable on Google and being recommended by an assistant are now separate achievements.
Ratings are part of the picture, but the numbers get misquoted constantly, so here is what they actually say. SOCi's 2026 Local Visibility Index, covering nearly 350,000 locations across 2,751 multi-location brands and published on January 28, 2026, reports that "locations recommended by ChatGPT averaged 4.3 stars, compared with 3.9 stars on Gemini and 4.1 stars on Perplexity". Those are averages of the venues each engine chose, not thresholds you must clear. Read them as the competitive band your venues are judged inside: a venue below it is not disqualified, and one above it is not admitted. Anyone selling you a star target off these figures has misread them.
Two other signals matter more than the average. A 2026 MyPlace study, reported by Metricus, puts a critical visibility threshold at around 2,000 reviews, with venues below it rarely appearing in AI suggestions regardless of food quality. MyPlace's methodology is not disclosed, so treat that as indicative. And pricing is the weakest thing engines understand about a restaurant: Courtyard AI's State of AI Visibility 2026: Restaurants, covering 120 restaurants across 50 US and Canadian markets against ChatGPT, Gemini, Perplexity and Google AI Mode in the week of July 10, 2026, scored assistants' grasp of "what it costs" at 0.06 out of 1, the lowest attribute it measured. If your menu prices live inside a PDF, that is a separate leak from the entity and listings problem, and a cheap one to fix. We've covered the invisibility numbers and the star-rating trap in our sourced statistics reference and the star rating trap piece.
Why are multi-unit groups more exposed than single sites?
A single independent restaurant has one name, one address, one menu and one set of reviews. A group multiplies every one of those signals across however many venues it runs, and multiplication is not the same as strength: more surface area for something to go wrong, and more places for authority to get split rather than concentrated. We've written the full mechanics in why your restaurant group's venues cannibalize each other on ChatGPT. The point worth adding for a growth conversation is why this structure sits at the center of a group's growth ceiling, not at the edge of it.
More entities to keep consistent. Seven venues means seven names, addresses and phone numbers that must match exactly across your own site, every booking platform and every listing. One mismatch reads as risk to an engine deciding whether to cite you, and a group has seven times the chances to get one wrong.
Brand-versus-venue confusion. An engine can know your group's name in the abstract while having no confident link to a specific, bookable address. That's fine for awareness and useless for the diner trying to book tonight, which is the moment that converts.
New openings starting from zero. Every new site launches with no reviews, no local citations and no resolved entity of its own, while carrying a group name diners already half recognize. That mismatch, brand equity attached to an unverified local entity, is exactly the pattern an engine hesitates to trust, so a new opening fights its hardest growth battle with the least AI footprint of anywhere in the group.
Cannibalization between venues. The reviews, mentions and citations that would push one venue over an engine's confidence threshold get scattered across several, so no single venue accumulates enough corroboration to be named. Worse, an engine can name the wrong branch entirely, which reads to a diner as the group not having its act together, when the real problem sits upstream in how the entities were structured.
None of this earns sympathy from an engine. It's a reason to treat AI discovery as a structural, group-level problem, not a page-by-page task handed to whoever updates the website.
What is a shift in discovery share plausibly worth to a group of 7 venues?
This needs a worked example with the assumptions on the table, because the honest answer is "it depends," and a made-up single number would be worse than no number at all. The revenue benchmarks below are in pounds, because that is where the per-site figures are actually published. Read the pounds as a sized illustration rather than a conversion, and the method as the part that transfers.
| Step | Assumption | Basis |
|---|---|---|
| 1. Venues | 7 | Given |
| 2. Average annual revenue per venue | £1.2m | Rounded down from the £1.276m midpoint of a UK casual dining range of £912k to £1.64m, annualized from daily revenue benchmarks published by Primewise, a marketing agency, so a secondary source rather than a primary study |
| 3. Total group revenue | £8.4m | Step 1 × Step 2 |
| 4. New-customer share of revenue | 20% to 30% | Widely cited hospitality rule of thumb; no single primary study confirms an exact figure, so treated as a range, not a fact |
| 5. New-customer revenue | £1.68m to £2.52m | Step 3 × Step 4 |
| 6. Share of new-customer decisions already touched by an AI tool | 22% | DoorDash 2026 Restaurant Industry Trends Report, US data |
| 7. AI-influenced new-customer revenue pool | £369,600 to £554,400 | Step 5 × Step 6 |
| 8. Illustrative capture-rate improvement from fixing entity and citation gaps | 10% to 33% | Stated planning assumption, not a sourced figure, chosen as a deliberately modest range |
| 9. Plausible annual revenue at stake | roughly £37,000 to £183,000 | Step 7 × Step 8 |
Read the last line as a range to plan against, not a promise: roughly £5,000 to £26,000 per venue per year, smaller than most pitch decks in this category would like, which is deliberate. Step 8 isn't there to flatter the outcome. It shows that even a cautious fix to how consistently seven venues show up in AI answers is worth doing on its own terms, without an invented headline figure. If your cover value or new-customer mix differs, rebuild the table with your numbers.
Who should own this inside a group?
Someone specific and internal, sitting close to but not inside your PR or brand agency relationship. The reason is scope, not competence. An agency retainer is built around earned coverage, launches and brand voice, work that runs in weeks and months. AI discovery needs someone checking on a standing cadence what ChatGPT, Google AI Overviews, Perplexity and Gemini say about each of your seven venues this week, then feeding fixes back into entity data, listings and content. Most retainers were never scoped for that rhythm, so asking the agency to absorb it without changing the brief just means it doesn't happen.
In practice this lands with a marketing director or head of growth, because they already sit between the brand story the agency tells and the reality of seven separate bookable venues. Their job isn't to replace the agency's earned-media work. It's to own the entity layer, so every venue is a clean, distinct thing an engine can resolve with confidence, track the citation pattern across engines, and hand the agency a sharper brief for the "best of" and editorial coverage that still drives most citations. It's a new line of ownership alongside the existing agency, not a second agency stacked on the first.
What should a multi-unit group actually do in the next 90 days?
Days 1 to 30: find out where you stand. Run the real questions diners ask, not just your brand name, through ChatGPT, Perplexity, Google AI Overviews and Gemini for every venue individually, and record which get named correctly, which get confused with a sister site, and which don't appear. Cross-check every venue's name, address and phone number across your own site, Google Business Profile and every booking and delivery platform. This pass usually shows whether the core problem is entity inconsistency, thin citations, or both.
Days 31 to 60: fix the entity layer. Give every venue its own page and Restaurant schema with an unambiguous address and area, resolve every name-address-phone mismatch the audit found, and align every listing (Google Business Profile, TheFork, OpenTable and equivalents) so they say the same thing everywhere. Push reviews toward the platforms engines actually pull from, given listings and reviews carry 41.63% and 13.28% of food service citations. Publish crawlable menu text with prices, since pricing is the weakest signal most restaurant sites give an engine.
Days 61 to 90: earn the citations that compound. Brief your agency with the specific gaps found, pitching venue-specific "best of" and local editorial coverage rather than group-level brand stories, since citations attach to resolvable venues, not an abstract brand. Seed honest presence in the local forums engines already lean on. Then agree who owns the standing check of what the engines say, and put it on a calendar with the same seriousness as a P&L review: this is where the exercise either becomes a habit or quietly stops.
If you'd rather see where your own venues currently stand before committing to any of this, the free AI search audit shows which of your sites the major engines already name, and which they skip.
Sources
- DoorDash, "From the Doorstep to the Dining Room: New DoorDash Survey Data Reveals the Full Picture of the Modern Restaurant Guest": 2026 Restaurant Industry Trends Report; 22% of consumers have used an AI tool to help choose a restaurant; survey of 3,001 US consumers and 509 operators by Dynata, March 2026.
- BrightLocal, "Local Consumer Review Survey 2026: half of consumers are asking AI for business recommendations": AI use for local business recommendations rose from 6% (2025) to 45% (2026); ChatGPT 31%, Google AI Mode 23%; survey of 1,002 US adults, published March 2026.
- Yext, "AI Citations, User Locations, and Query Context": food service "presents a unique challenge, with a significant dependency on both listings (41.63%) and reviews (13.28%)"; the highest user-generated-content reliance of any industry measured; analysis of 6.8 million citations across ChatGPT, Perplexity and Gemini, published October 2025. Yext's prose gives no first-party website percentage for food service, so none is quoted here.
- Google Search Central, "AI Features and Your Website": official documentation on query fan-out and how AI Overviews and AI Mode select and display sources.
- Search Engine Land, "AI Overview fan-out rankings boost citation odds by 161%: Study": coverage of a Surfer SEO study of 10,000 keywords and 173,902 URLs; pages ranking for fan-out queries are 161% more likely to be cited; about 68% of cited pages didn't rank in Google's top 10 for the main query or any fan-out query; published December 2025.
- MediaPost, "Most QSRs 'Effectively Absent' From AI-Generated Recommendations": coverage of Uberall's "Fast Food, Faster Discovery: 2026 GEO Playbook for Multi-Location QSRs"; 83% of restaurant locations entirely invisible in AI-generated recommendations while 86% hold some presence on Google; published May 2026. Used here only for those figures; the Uberall report carries no star-rating data.
- Search Engine Land, "AI local visibility is up to 30x harder than ranking in Google: Report": coverage of SOCi's 2026 Local Visibility Index, nearly 350,000 locations across 2,751 multi-location brands, published January 28, 2026; "locations recommended by ChatGPT averaged 4.3 stars, compared with 3.9 stars on Gemini and 4.1 stars on Perplexity". These are averages of recommended venues, not qualifying thresholds.
- Metricus, "Good Reviews but Empty Tables: Why Independent Restaurants Are Losing to Chains Online": secondary reporting of a 2026 MyPlace study placing a critical visibility threshold at around 2,000 reviews; MyPlace's methodology is not disclosed. Metricus sells paid visibility reports, so treat as an interested secondary source.
- Courtyard AI, "The State of AI Visibility 2026: Restaurants": assistants' understanding of "what it costs" scored 0.06 out of 1, the weakest attribute measured; 120 restaurants across 50 US and Canadian markets, tested against ChatGPT, Gemini, Perplexity and Google AI Mode, week of July 10, 2026.
- NIQ, "CGA/Reputation Study Reveals AI and Economic Pressures Driving New UK Consumer Habits in Hospitality": a quarter (26%) of consumers now use AI apps "to learn more about a venue", usage the study puts on a par with Google Maps (27%); nationally representative survey of 755 British consumers by CGA by NIQ and Reputation, August 2025.
- Primewise, "Average Restaurant Sales Per Day in the UK: Benchmarks and the 3 Numbers That Matter More": UK casual dining annual revenue range of £912k to £1.64m per site, annualized from daily revenue benchmarks. Primewise is a marketing agency, so this is a secondary source, not a primary study.
- Schmitdy, "Why Your Restaurant Group's Venues Cannibalize Each Other on ChatGPT": the entity-structure mechanics behind multi-venue disambiguation failures.
- Schmitdy, "Restaurant AI Visibility Statistics (2026): The Numbers, Sourced": the underlying invisibility, star-rating and UK/German demand figures this article builds on rather than restates.
