How to Measure Your Restaurant's Visibility in AI Answer Engines

What does restaurant visibility in AI answer engines actually mean?

It means how often, how prominently, how accurately, and how favourably an engine such as ChatGPT, Gemini, and Google AI Mode includes your restaurant when someone asks a question like "where should I take my parents for their anniversary" or "what's open near me right now."

There is no shared index and no public rank position to read, so visibility is not observed, it is sampled. You choose the questions, run them repeatedly, record what comes back, and score it against a written rubric. The rubric and the prompt set are the instrument, which is why keeping them stable matters more than which tool you use to run them.

Our own measurement work is described in the ReleVantage methodology, and the same scoring is what an AI search audit produces as a first baseline.

How do I build a prompt set I can trust?

Build it from language real diners use, then freeze the core set. If prompts change every month, nothing is comparable and every movement is an artefact of the instrument. New prompts go into a separate tracked group with their own start date.

Size the set so it is representative of how your diners actually ask, and so no single answer can swing a percentage on its own. Keep it small enough that every repeat run can still be reviewed honestly rather than skimmed.

• •Category prompts: what someone asks before they know any restaurant name
• •Comparison prompts: your restaurant against named alternatives
• •Factual prompts: hours, menu, price range, private dining capacity, dietary accommodations, credentials
• •Constraint prompts: neighborhood, budget, party size, urgency, specific dietary requirement
• •Branded prompts: what the engine says when asked about you directly

Which metrics should I use, and how is each one calculated?

Score these separately, per engine, per period. Rolling them into one visibility number hides the detail that tells you what to fix. In each formula below, a prompt counts once per period after its repeat runs have been consolidated.

Why do two runs of the same prompt give different answers?

Generated answers are probabilistic, retrieval changes between runs, and personalisation, location, and account context all shift results. A single response is an anecdote.

Run every prompt multiple times per engine each period, enough repetitions to see the spread rather than a single draw, record each response in full, and report the consolidated result alongside that spread. If a prompt names you in most runs one week and hardly any the next, the spread is the finding worth investigating.

• •Keep wording, schedule, and location context identical between periods
• •Record the full answer text, not a summary, so scoring can be re-checked
• •Note engine, model version where visible, date, and any location used
• •Treat a change smaller than your observed run-to-run spread as noise

Should I report one number or one per engine?

One per engine. The engines retrieve differently, cite differently, and refresh on their own schedules, so a blended average can hide a collapse on one platform behind a gain on another.

Keep a per-engine table for every metric, and use a blended figure only as a headline that links through to the detail.

How do I compare a baseline against the current period?

Record a baseline before any work starts: the full prompt set, every answer, and every metric, per engine. That run is the reference point for everything reported later.

Each period after that, report baseline, previous period, current period, and the change, alongside the specific work shipped in between. Where a movement cannot be explained by work you did, say so. Engines update independently of anything you publish, and correlation is not proof.

What are the limits of attribution here?

Assistant referral data is incomplete by design. Some engines pass an identifiable referrer, some pass none, and many answers create demand that arrives later as a branded search, a direct visit, or an offline enquiry.

Capture landing page, referrer, and campaign parameters at session level with first-party tracking, pass that context into enquiry forms, and add a self-reported source field. Treat the result as directional evidence. Any report claiming complete revenue attribution to AI answers is overstating what the data supports.

How often should this be measured?

Choose a cadence your team can repeat consistently. One practical option is weekly repeat runs on a priority subset and a monthly run of the full set; faster-moving categories may justify a shorter interval on a smaller group.

Consistency beats frequency. A disciplined monthly run with the same prompts and rubric is worth more than daily runs that keep changing shape. Ongoing AI search monitoring exists to keep that cadence steady once the baseline is set.

Frequently asked questions