How to brief an AI-search agency for a restaurant group

If guests keep finding the wrong branch, put that problem at the top of your agency brief. Give the agency your venue list, correct menu and booking links, and the questions guests ask. Agree which pages it will check, what it may change and how your team will verify the result.

The brief should help your team buy useful work without surrendering ownership of your website, listings or guest data. It should also prevent a familiar problem: the agency reports more mentions while the host team still receives enquiries for the wrong branch.

What decision should the brief make possible?

Start with the commercial problem. You may need guests to discover a new venue, distinguish two similarly named sites, find private-dining information or reach the correct booking page. Each problem needs different evidence and changes.

Write a short opening statement using this shape:

We want [guest type] looking for [real occasion] in [place] to find [venue or venue set], receive accurate information and reach [menu, enquiry or booking destination]. The current problem is [observed gap]. We need the agency to establish a baseline, recommend specific changes and verify the work within [agreed window].

Keep it concrete. “More visibility” does not identify a guest, a place or a useful next step. “Guests asking for a weekday lunch near the station keep finding our other branch” gives the agency something to investigate.

Which venue information should you provide?

Create one row per venue. Use information approved by the people running the site, rather than asking the agency to infer facts from scattered pages.

Google's LocalBusiness structured-data guidance covers location-specific information including address, hours, menus and URLs. It is a useful checklist for the agency's technical inspection, not a substitute for a complete visible venue page.

If several sites use the same restaurant name, add the distinction guests use: neighbourhood, street, station or concept. If guests confuse your branches, our branch-discovery guide explains how clearer venue pages help them choose the right site.

How do you choose the questions to test?

Use questions a guest could ask without knowing your name. Collect them from reservations staff, private-dining enquiries, calls and the questions people ask before arriving. Remove personal details before sharing examples.

A useful starter set might include these patterns, adapted only to facts the venue can support:

• Where can four people eat before a show near a named theatre?
• Which restaurant near a particular station serves lunch on a weekday?
• Where can a larger party book a private room in this neighbourhood?
• Which venue has a published vegetarian menu and a suitable reservation route?
• Where can a guest find clear step-free access information before booking?

These are patterns, not claims that every restaurant offers those services. Do not add wheelchair access, dietary suitability or a private room because the query seems attractive. Ask the venue team to verify the fact first.

Agree the city, language and AI service for each question. Record branded questions separately. Asking whether an assistant knows the group's name tests a different job from discovering a restaurant for dinner.

Google says that AI Overviews and AI Mode may use different models and techniques and show different responses and links. The brief should name the product being measured rather than combining every service into one unexplained percentage.

What should the agency deliver?

Ask for outputs your team can inspect and use. Specify whether the engagement covers research, implementation or both.

1. A baseline: exact questions, market, language, AI service, dates and captured answers.
2. An accuracy review: correct branch, opening status, menu, factual claims and destination.
3. A source and page review: which pages and listings support or contradict those facts.
4. A prioritised plan: affected URL, proposed change, intended guest benefit, owner and dependency.
5. Implemented changes, where commissioned: a list of what changed, with approval and readback.
6. A retest: comparable checks with differences and remaining limitations explained.
7. A handover: retained results, account access and instructions another team can use.

“Twenty recommendations” is not a useful deliverable if none identifies a page or owner. Equally, a new venue page is not verified simply because the agency has created a file. Check that it is live, readable on a phone, accurate and reachable from the relevant part of your site.

OpenAI's search guidance explains how search systems read public pages and use multiple relevance and reliability factors. A supplier can check access and improve available information; it should not present crawler permission as a guarantee that ChatGPT will recommend a venue.

Where do approval and access belong?

Keep ownership with the group and grant the permissions required for the work. Google's Business Profile owner and manager guidance allows individual accounts to manage profiles without sharing passwords. Its business-group guidance describes group-level access separately. Agree whether the agency needs a single profile or all the agreed venues, and remove access when the engagement ends.

Use a simple responsibility table:

This is where a practical brief earns its keep. It prevents a supplier from announcing an offer the venue cannot honour, answering an unverified complaint or changing a listing beyond the agreed work.

Keep the group’s important accounts protected with passkeys or strong passwords and two-step verification, and remove agency access when it is no longer needed. The NCSC’s account-security guide explains these safeguards. Agree an access review at handover so a completed project does not leave forgotten accounts behind.

What data should stay inside the business?

Visibility research can often start with public pages, public answers and an approved venue list. Do not send a guest database merely because the agency asks for “context”. Provide aggregate enquiry themes or anonymised examples where those are enough.

The ICO's data-minimisation principle requires personal data to be limited to what is necessary. If the supplier processes personal data on the group's behalf, consider the controller and processor contract requirements. Define the purpose, access, retention and return or deletion of that information.

Include review work explicitly. Google's review policies prohibit fake engagement and specified forms of review manipulation. A brief should never ask an agency to buy reviews, hide genuine criticism or post as guests.

Which measurements belong in the brief?

Separate discovery, arrival and confirmed business outcomes. A correct recommendation is useful; it is not a booking.

The GA4 traffic-acquisition report measures sessions using traffic-source dimensions. Google's explanation of traffic-source dimensions distinguishes acquisition and attribution information. Ask for the exact metric and scope, rather than accepting a chart labelled “AI customers”.

Search Console provides Google search and indexing information. It is useful alongside website analytics; it is not a complete record of guests' reservations. Use the existing restaurant bookings-measurement guide for the fuller chain.

What should you ask before accepting the proposal?

Ask the supplier to show a representative anonymised baseline, a repair plan and a completed retest. Check whether the same question set and venues appear throughout. Ask how you receive exports and what happens to access when the contract ends.

Then resolve five buying questions:

• Which sites and languages are included, and which are sampled or excluded?
• Who performs the corrections, and does the fee include that work?
• Which public actions require the group's approval?
• What evidence will establish completion of each deliverable?
• Which outcomes are measured, and which are ambitions rather than guarantees?

The ASA's substantiation guidance explains the need for evidence behind objective advertising claims. Ask the same focused question of a supplier claiming a large uplift: what was measured, for whom, over what period, and with which comparison? A result from one account does not determine your group's outcome.

A short brief you can use

Our group operates [venues and concepts] in [cities]. We need guests looking for [verified occasions] to find the right site and reach [destinations]. Please establish a documented baseline using an agreed non-branded question set, inspect the listed venue pages and profiles, identify factual and journey errors, and propose or implement the agreed repairs. State engines, markets, languages, dates, exclusions, access needs, approval boundaries, deliverables and retest method. We retain ownership of accounts and results. Count recommendations, visits and confirmed bookings separately.

Bring that brief to a free AI-search audit or a supplier conversation. It will make the discussion more useful before anyone starts writing pages, buying software or promising a percentage.

Sources

1. LocalBusiness structured-data guidance.
2. AI Overviews and AI Mode.
3. OpenAI's search guidance.
4. Business Profile owner and manager guidance.
5. business-group guidance.
6. NCSC’s account-security guide.
7. ICO's data-minimisation principle.
8. controller and processor contract requirements.
9. review policies.
10. GA4 traffic-acquisition report.
11. Google's explanation of traffic-source dimensions.
12. Search Console.
13. ASA's substantiation guidance.