When we set out to make our own site — and the small-business sites we build — show up in ChatGPT, Perplexity, and Google's AI Overviews, we did something slightly obsessive: we ran our plan past three independent reviews and only kept the tactics all of them agreed actually move the needle. A surprising amount of the popular "optimize for AI search" advice didn't survive.

Here's the short version, because the short version is the whole point.

The stuff we deliberately skipped

These get written about constantly. We don't spend time on them:

  • FAQ and HowTo schema as a ranking trick. Google deprecated these rich results for commercial sites back in 2023 — FAQ rich results are health/government-only now, and HowTo is gone entirely. Marking up a page with FAQPage won't win you a SERP feature. (Schema still matters — just for a different reason; more below.)
  • llms.txt as a citation channel. No major answer engine reads it for live answers today. It's cheap to publish and spec-shaped, so we have one — but treating it as a growth lever is wishful thinking.
  • Fiddling with the robots allowlist. Allow: / already lets every crawler in. And Google-Extended is a training opt-out token, not a control over whether you appear in AI Overviews — those are built from the regular Google index.
  • Content treadmills and bought links. Publishing volume for its own sake, "schema everywhere" on pages with nothing citable, purchased backlinks, doorway city-by-category pages. All footprint, no signal.

What actually moves AI citations

Ranked by leverage, this is what survived all three reviews:

  1. Be in the index the engine reads. AI Overviews and Gemini ground on Google's core index; ChatGPT and Perplexity lean on their own crawl plus Bing. Not indexed means not citable. Boring, classic indexation — crawlable HTML, a clean sitemap, search-console verification — is the prerequisite, not a separate "AI" track.
  2. Original, specific, verifiable facts. Generative answers quote numbers, definitions, and comparisons — not slogans. A page that states a concrete fact an engine can repeat gets cited; a page of adjectives doesn't.
  3. Entity clarity off-site. Engines trust an entity they see described consistently in more than one place: the same name/address everywhere, links to real profiles, links from real domains.
  4. Structure built for extraction. Answer-first paragraphs, question-shaped headings, comparison tables (engines lift these straight into answer boxes), and clean server-rendered HTML — not content hidden behind JavaScript or iframes, which crawlers and LLM scrapers see as blank.

And yes — schema. Just not as a rich-result hack. We use Organization, LocalBusiness, Service, and Article markup so a machine can understand what a page is, which makes a fact easier to extract with confidence. That's the job it's actually good for.

The honest caveat

AI search compounds over months, not days, and attribution stays murky — engines won't tell you why they cited someone else. So we treat this as hygiene: get the foundations right once, keep the facts current and specific, and put the marginal hour into doing real work worth citing rather than chasing the next tactic.

Which is the quiet punchline. The single best thing we can publish is a true, specific thing nobody else has said — like this post. Being worth quoting is the optimization.