What We Learned Auditing 100 Sites for AI Visibility

We recently crossed 100 completed audits in GetBotRank. That's enough sites - spanning SaaS, ecommerce, local services, and content publishers - to start seeing real patterns instead of anecdotes. Some confirmed what we expected. A few surprised us.

What actually predicted citation rate

Question-shaped headings beat keyword-shaped ones. Pages with H2/H3 headings phrased as the questions users actually ask ("How much does X cost for a team of 10?") were cited noticeably more often than pages with the equivalent keyword-optimized heading ("X Pricing"). Answer engines appear to weight the semantic match between a query and a heading heavily when selecting a source to quote.

Freshness signals mattered more than we expected. Pages with a visible, accurate "last updated" date - and content that actually reflected that update - outperformed undated evergreen pages on topics where facts change (pricing, feature lists, comparisons). Several models appear to discount or avoid citing pages where they can't establish recency.

Comparison pages punched above their weight. Sites with an honest, specific "us vs. competitor" page were cited disproportionately often for exactly the multi-option queries ("what's the difference between X and Y") that make up a large share of AI search volume. Vague or marketing-only comparison pages did not get the same lift - specificity was the deciding factor, not the existence of the page.

What turned out not to matter much

Raw word count. We expected longer, more comprehensive pages to win. They didn't, beyond a fairly low threshold. A tight 400-word page that answered the question precisely often outperformed a 3,000-word page covering the same ground with more throat-clearing.

Backlink profile, as a standalone signal. Domain authority still correlates with citation rate, but weakly compared to on-page structure. Several lower-authority sites with well-structured, specific content outcited higher-authority competitors whose pages were vaguer or more marketing-heavy.

Meta descriptions. Unsurprising in hindsight - meta descriptions are written for search result snippets, not for models reading the full page - but worth saying plainly: optimizing this field had no measurable effect on AI citation behavior.

The pattern underneath the pattern

If there's one thread through all of this, it's that answer engines reward pages that behave like good direct answers to a specific question, and are largely indifferent to signals that were built to game a ranking algorithm rather than to communicate clearly. That's a genuinely encouraging finding - it means AEO/GEO isn't a new set of tricks to learn so much as an old discipline (write specifically, structure clearly, stay current) that finally has a more literal-minded reader grading it.

We'll keep sharing what we see as the audit count grows. If you want your own site audited, you know where to find us.