Schema that moves the needle (and the kind that doesn't).
FAQ, HowTo, LocalBusiness, Service. We tested combinations across 24 sites. Here's the short list that consistently worked.
Structured data has been around for over a decade and most agencies still treat it like a checklist: ship every type schema.org has, declare victory. In the AI era, that's worse than wrong; it actively dilutes the few signals that matter.
Over the last six months we've tested combinations of JSON-LD schema across 24 client sites, measuring AI citation rate before and after each variant. Here's what consistently moved the needle, what was neutral, and what occasionally hurt.
What worked.
LocalBusiness + Service. The pairing of LocalBusiness (with full address, phone, opening hours, geo coords, area served) and one Service entry per primary offering was the single most consistent lift. AI models use this combination to confidently say "X serves Y in the Z region", which is exactly the sentence you want them generating about you.
FAQPage on long-form service pages. Critical caveat: the questions and answers in the schema need to actually appear visibly on the page (not hidden in CSS, not in tabs that only open on click). Done right, FAQPage gave us the cleanest extraction pattern of any schema type. Done wrong (questions hidden until interaction), it was a wash.
Article schema on real publications. If you publish actual writing (case studies, research, technical guides), Article schema with author, dateModified, datePublished and publisher is meaningful. Don't apply it to thin product or service pages.
What was neutral.
Breadcrumb schema. It's good for Google SERP appearance and was once a small SEO win. AI models don't seem to use it. Zero measured lift in our tests.
Organization schema. Useful for the brand-name knowledge panel in Google, but redundant if you have a thorough LocalBusiness entry. We include it but don't credit it with any of the lift we measure.
SiteNavigationElement. Largely ignored by AI crawlers. Don't bother.
What occasionally hurt.
HowTo schema on the wrong content. HowTo is meant for actual step-by-step instructional content. We saw a handful of clients applying it to sales pages with a vague three-step "how we work" graphic. AI models that tried to extract "steps" from those pages produced answers that misrepresented the service, occasionally to the point of being wrong about what the business did. We strip HowTo from non-instructional pages by default now.
Over-stuffed Service schema. Putting twelve overlapping Service entries on a single page, each with broad descriptions, made the AI's job harder, not easier. AI models couldn't distinguish the offerings and started either omitting the business entirely or picking one offering and ignoring the others.
The short list.
For most Australian SMBs, the schema setup that consistently moves AI citation rate is:
- One LocalBusiness entry on the homepage with complete trading details.
- One Service entry per page that maps cleanly to one offering. No more, no less.
- FAQPage on the two or three pages where the questions are also fully visible on the page itself.
- Article schema, properly attributed, on the writing/insights/blog pages.
That's it. Skip the rest. Less is more here. The AI is trying to extract a clean, confident answer from your site, and every extra schema type is one more thing that could either confuse it or contradict another part of your content.