Field research · Auckland · June–July 2026

The AI recommended it.
It closed months ago.

As part of our Auckland research we did something unglamorous: verified every single business against the live web. Every vertical contained closed venues, ghost listings, and wrong data — and the AI engines, assembling answers from stale sources, recommended some of them anyway.

SCOPE — Auckland, New Zealand · June–July 2026 snapshot. 7 verticals, 679 probed businesses: Italian, Japanese, sushi and Indian restaurants, barbers, plumbers, and board-game shops. Engines queried via their APIs — Gemini with Google Search grounding, OpenAI web_search, Claude with live search, and Apple Maps (MKLocalSearch, the layer Siri uses for local queries). APIs are clean, reproducible instruments and proxies for the consumer products, not the apps themselves. One city, one snapshot — treat this as field data, not a global law.

Everything downstream of a business list inherits that list's errors — so we validated ours the hard way: a research agent checked every venue in every run against the live web (site, socials, review platforms, news), flagging closures, duplicates, wrong addresses, and category mismatches. These are flags, not certainties — the validator can err too — and the counts mix severities: a suburb mismatch and a permanent closure each count once. The dozen confirmed closures are the hard core.

The map, it turns out, is dirty everywhere:

  • Every one of our seven runs surfaced list issues — from 14 flags among 67 plumbers to 52 among 104 barbers.
  • The most concentrated case: board-game shops, where 16 of 19 master-list entries carried some issue — closures, a shop that had become an online-only ghost, listings pointing at the wrong city.
  • We confirmed a dozen outright closures across the dataset that still existed as live-looking listings.
  • And the engines inherit it: one closed shop, The Board Gamer, was still being cited in AI answers — its old domain now silently redirecting to a big-box competitor. The recommendation looked fine. The business it recommended didn't exist.

Stale truth is the default

This is the quiet flaw in AI-mediated discovery: engines assemble answers from written sources, and written sources outlive the businesses they describe. Nobody updates the 2023 listicle when a restaurant closes. The map layer keeps the pin. The review platform keeps the page. An agent reading all of it sees a business that stopped existing months ago — and recommends it with total confidence.

For business owners the flip side is more practical: the stale data being served about you is competing with you. Old hours get quoted as current. A previous menu, a dead booking link, a moved address — whatever the written record says, that's what gets repeated to customers who never see your actual website.

What to do about it

Hunt your own ghosts. Search your business the way an agent would and check what's stale: old addresses on directories, dead booking links in listicles, an unclaimed profile with 2024 hours. Each one is being read as truth.

Keep the sources fresh, not just the site. Updating your own website doesn't update the ten other places engines learned about you from. The maintenance loop — recheck what agents say as models and sources shift — is exactly why we built monitoring into the service.

What stale story is being told about you?

We verify what every engine currently says about your business against reality — and flag what's wrong before your customers find it.

Request a shelf check →