Case study · Unsolicited audit · Local restaurant · August 2026
A neighborhood Italian trattoria in the San Fernando Valley — audited at snapshot depth because the local failure mode is vivid and almost nobody has measured it. Not a client. One prompt set across five intent categories, clean logged-out sessions, every answer recorded verbatim.
Recorded answers that named the restaurant — an 87.5% appearance rate. On the surface, a business with no visibility problem at all.
Category discovery — ask "best Italian near [neighborhood]" and the engines named them every time. The hard part, already won.
Brand-direct — ask about the restaurant by name and the engines hedged, mixed it up, or missed half the time. The easy part, being lost.
The inversion is the finding. Most owners assume discovery is the hard part. Here, the engines had already handed them discovery — and were fumbling their own name.
Most visibility work is spent trying to win the top bar. This business already had it, and was losing customers on the bottom one.
The sharpest single answer in the run: asked about the restaurant with no city in the prompt, one engine confidently and fluently described a different restaurant with a nearly identical name — in a city two thousand miles away.
No hedge, no "there are several restaurants by this name." A customer who'd heard the name from a friend and asked the engine about it got the menu, the vibe, and the neighborhood of a stranger's business.
That's not an exotic failure. It's the default outcome for any business whose name isn't uniquely anchored in the sources the engines trust — and it's invisible until someone runs the prompt, because the wrong answer reads exactly as confidently as a right one.
On discovery questions — the ones the restaurant was winning — the citations went to review directories, a booking platform, Reddit threads, a mapping site, and a local blog. The restaurant's own site was cited only when the prompt contained its name. Citation share across the run: 42.9%.
So the restaurant's discovery wins were built entirely on other people's descriptions of it. The engines put it in the answer because third parties vouched — and every fact in that answer was only as current as the directories were.
They were the subject of the answer without being its source. For a local business, that's the whole game in one sentence.
Identical name, address, and neighborhood everywhere the engines look — site, directories, booking platform — so a name-only question stops resolving to a restaurant in another state.
The citations named exactly which third parties the engines trust for this category. Current menus, hours, and photos on those platforms move the answer more than anything on the restaurant's own site.
Hours, location, and the answers to the questions people actually ask, stated plainly where a machine can lift them — so the own-site citation rate has somewhere to go.
This snapshot ran eight recorded answers across two engines; the third engine wasn't run. That's below the three-engine standard, it's disclosed in the deliverable, and it's disclosed here — a finding you can't reproduce is an anecdote, so the run count travels with the numbers.
Two operational lessons came out of this run and are now standard method. Geography lives in the prompt — engines geolocate the analyst, and an audit run from the wrong city contaminates every local-intent answer. And clean sessions are non-negotiable — a signed-in account carries history and location into every result. Logged-out, memory off, fresh session per prompt.
Why anonymized: the restaurant never asked for this audit. The findings went to them; the method and the numbers are public; their name isn't attached to their gaps. Client results, when published, are named — with written permission.
One page, forty-eight hours: what the engines say when strangers ask about your category, what they say when someone asks about you by name — and which of those two is quietly broken.
I run a limited number each month, so there may be a short wait.
Takes about a minute. The more precise the last two answers, the sharper the report.
I'll only use this to run and send your snapshot. No list, no sequence you didn't ask for, and you can tell me to delete it at any time (see the privacy policy). I run a limited number each month, so there may be a short wait.