What Actually Moves the Needle for LLM-Era SEO, Tested

'LLM SEO' has become a category full of confident advice with very little testing behind it. We ran a dozen commonly recommended tactics across a set of client content pages over four months, tracking both AI Overview citation presence and referral traffic from AI chat tools where it's measurable. Results were a lot more mixed than the advice articles suggested.

What didn't move anything we could measure

Adding explicit 'AI-friendly' formatting markers, FAQ schema stuffed onto pages that didn't naturally have FAQ content, and keyword-stuffing variations of a query phrase all showed no measurable change in citation rate. These are tactics that sound plausible because they mimic what a crawler-era SEO tactic looked like, applied to a new target, without evidence they actually affect how an LLM selects source content.

Most LLM SEO advice right now is crawler-era SEO tactics relabeled for a new audience, with no evidence the mechanism actually transfers.

What actually moved citation rates

• Clear, extractable direct answers near the top of the page — a concise, self-contained answer to the exact question the page targets, before any narrative buildup — correlated strongly with citation pickup.
• Original data or a specific tested number (our own survey result, our own benchmark) got cited noticeably more often than restated general knowledge, likely because it's the one thing an LLM can't synthesize from elsewhere.
• Clean, unambiguous page structure with a single clear topic per page outperformed long pages trying to cover several related questions at once.
• Fast, reliably crawlable pages (no JS-rendering dependency for the core content) still mattered — a surprising number of AI crawlers behave more like simple fetchers than modern Googlebot.

The pattern across what worked is less about gaming a new system and more about making content genuinely easy to extract a correct, specific answer from. That's not really a new SEO discipline — it's good writing with a verifiable fact in it, which happens to be exactly what both a human skimmer and an LLM summarizer reward.