Citation audits now find far more of a location's listings

We measured our citation audits against an independent scan of the same locations, found we were recovering a small fraction of what it knew about, and spent the day closing the gap. Audits also come back sooner and read pages the cheap way first.

A day spent entirely under the surface. We put a number on how good our citation audits actually are, did not like it, and rebuilt the parts the number pointed at.

For our partners

Nothing moved on the surface today. The one thing you might notice is a citation audit that comes back sooner and with more listings on it.

Behind the scenes

1. We measured our citation audits against an independent scan of the same locations and found we were recovering a small fraction of the listings it knew about. Most of the day went into closing that gap.
2. An audit that sat for ten to twelve minutes now spends a fraction of that waiting, because searches and page reads happen side by side and long lookups are started and collected later instead of held open.
3. Listing pages are read the cheap way first and only escalate when the page will not give up a phone and a street.
4. When an agent asks a question, the options are now the edits it would make rather than the facts it is missing.
5. The keyword, content, crawl and citation services are now built, checked and updated the same way as the rest of the platform.

1. What the measurement changed

The rules that decide whether a page belongs to a business are strict on purpose, and they were throwing away pages a person would recognise at once: an old trading name with the right phone, a suite number the record does not carry. Those now get a second look, where an agent reads the page beside the record and answers whether it is the same business, and only when the page states a name and either a phone or a street. We also record the Google Business Profile as a listing, search every directory we hold a working search for instead of a shortlist, retry a town-scoped search without the town, and look up Apple Maps places directly, since nothing indexes them and nothing readable renders on the page.

Recall against the outside scan went from a handful of listings to just under half, measured the same way each time. Not finished, but we can now tell which way the number moves.

3. A ladder for reading pages

Every listing page used to go through a paid extraction step. On a sample of real listing pages, most already stated a phone and a street in the page itself, so the read now starts with a plain request, then a rendered page, then a light fetch, and only pays for extraction when the page comes up short. Pages behind bot walls get the heavier route automatically instead of being given up on. Each read records which rung answered, so we can see where the cost actually sits.

4. Asks that offer edits

An ask once offered two phone numbers and the option “both are correct, keep both”. Both were correct, so picking one resolved nothing. Every option is now something the agent would do with the answer, with the one it recommends first and a few words on why. The agent that plans page work went through repeated dry runs on live campaigns in the same spirit: it leaves intents and pages already at the top alone, and any decision resting on a campaign fact names where that fact is written.