We kept a schema Google had stopped rewarding, and it paid somewhere else

In 2023 Google stopped showing FAQ rich results for almost every site. The structured data that fed them kept working, kept validating, and stopped producing anything visible. Most sites removed it, which was the rational response to the stated change.

On one property we kept shipping it, for a reason recorded in source before any outcome was known: not for Google rich results, but as a machine-readable statement of what each service answers, aimed at systems that read pages rather than render them.

The property

A lead generation site in a specialist professional field. We built the website; a third party runs the service and the client relationships, and these figures are published with their permission. Nothing here describes their business beyond the traffic and the contact actions on the site itself.

FAQPage markup shipped on every service page from launch. The instrumentation, which recorded where each contact action originated, shipped at the same time.

What happened

The highest converting source on the site was an AI assistant, at roughly twenty six percent of its sessions producing a contact action, against five to six percent for direct and organic traffic. On a site with a hundred and fifty three sessions in total.

What this is not

This is not proof that FAQ markup caused the assistant traffic. The schema shipped first and the traffic followed, and that ordering is recorded in version control, but nothing here isolates the schema from everything else that shipped alongside it. It is a sequence, not an experiment.

The numbers on the channel itself are small enough to name exactly: twenty seven sessions and seven contact actions. A conversion rate computed off seven events is not a regime change. We have a recorded case on another property where a rate that looked like a step change on a similar sample reverted completely in the following window, and the honest position is that this one may do the same.

Contact actions are messaging app opens and form submissions. They are not confirmed clients. One of the thirteen was confirmed as a client, and the other twelve are not claimed as anything more than what they are.

The part that holds regardless

The measurement existed. A site of this size, in this field, would normally have no idea which channel produced its enquiries, because the instrumentation to answer that is usually added after someone starts asking, by which point the events are gone.

There is a companion finding on this site measuring the assistant channel on a utility property, where it came to about one percent of traffic and was worth nothing. Same channel, two properties, opposite conclusions. The difference is not the channel. It is that one property sells a considered professional service where a single enquiry is worth a great deal, and the other serves single-answer queries monetised by advertising. A channel is worth what a session on it is worth, and that is a property of the business, not of the channel.

How to check your own

Instrument the contact action, not the page view, and record the surface it came from. Then let it run before drawing on it. If the channel you care about has produced fewer than about thirty events, you do not yet have a rate, you have an anecdote with a percentage sign on it. Publish it anyway, with the count next to it, and let the reader judge.