500 SaaS Sites Audited: AI Citation Visibility in 2026

Citation visibility — getting cited inside ChatGPT, Perplexity, Claude, and Gemini answers — is the new SaaS organic acquisition channel. Traditional SERP traffic is flat to declining for most B2B SaaS verticals; AI-answer traffic is the line that moves. So we ran the audit nobody had published yet.

We pulled 500 SaaS landing pages — a stratified sample of B2B, prosumer, and developer tooling vendors at $1M to $500M ARR — and tested every page against ChatGPT, Perplexity, Claude (Anthropic Search and Citations), and Gemini (AI Overviews and standalone) over a 30-day window in March-April 2026. We logged every citation, scored every page on a structural rubric, and ran the correlations.

The result is the cleanest signal we have published in the GEO space to date. The top quartile of pages gets cited 8.4× more often than the bottom half, and the lift maps to a short list of repeatable structural choices: comparison sections, llms.txt, schema, answer-format headings. The full playbook is below — including the 8-point rubric we now apply inside our agentic SEO services.

Two things changed in the SaaS acquisition stack between 2024 and 2026. First, traditional Google SERP traffic for B2B SaaS keywords flattened or declined for most verticals as AI Overviews and ChatGPT zero-click answers absorbed the easy intent. Second, the number of users who start product research inside ChatGPT or Perplexity rather than Google crossed an inflection — for developer tooling specifically, the share is now well past 40%.

The combined effect is that the SaaS pages getting cited inside AI answers are getting the qualified-intent traffic the SERP used to deliver. The pages that are not getting cited are not getting that traffic — and the gap is measurable in trial signups within a single quarter. Citation visibility has become a leading indicator for pipeline.

That made the audit worth running. We wanted to know two things: how big is the spread between the best- and worst-cited SaaS pages, and what structural choices predict the spread. The answer to the first question is roughly an order of magnitude (8.4×). The answer to the second is a list of eight repeatable factors, of which six are page-level work that any team can ship inside a single sprint.

The sample frame: 500 SaaS landing pages drawn from a stratified list of B2B (60%), prosumer (15%), and developer tooling (25%) vendors with reported ARR between $1M and $500M. The cap on ARR kept the audit focused on companies where citation visibility is an actionable lever — at the very top of the market, brand and sales-led motion dominate the acquisition stack and structural page choices matter less.

For each site, we tested the primary product landing page (the page the homepage CTA points at) plus the dedicated pricing page. We submitted a fixed list of 12 query variations per vendor — three branded, three category, three comparison, three jobs-to-be-done — across ChatGPT (GPT-5.5 with web search), Perplexity (Sonar Large), Claude (Anthropic Search and Citations on Sonnet 4.5), and Gemini (3 Pro with AI Overviews). The window was 30 days, with three sampling passes to control for routing variance.

We logged every citation as a binary at the URL level (page X cited or not cited in answer Y) and aggregated to a per-page citation rate. We then scored each page on the 8-point rubric (described in section 06) and ran point-biserial correlations between each rubric item and the citation rate. The lift figures in section 04 are the per-factor correlation expressed as a percentage uplift in citation rate, after controlling for domain authority, page recency, and content depth.

The audit explicitly excluded brand-search behavior — we did not credit a vendor for being cited on a query that included its own brand name, since that signal collapses to whether the engine indexed the homepage at all. The interesting variance is in the unbranded queries, where the engines have to choose a citation and structural signals decide what gets picked. About 75% of the citation events in the dataset are from unbranded queries.

The headline finding is the spread. Bucketing the 500 sites by citation rate quartile, the top quartile averages 31 citations per month across the four engines; the bottom quartile averages 3.7. The middle two quartiles cluster between 8 and 14. The distribution is power-law shaped, not normal — a small share of sites accumulate most of the citation volume, and the long tail of under-cited pages is much longer than the over-cited head.

The cohort labels are not arbitrary. Cluster analysis on the rubric scores produces three naturally separating groups across the 500 sites — what we are calling AI-native, retrofit, and legacy. AI-native sites (~12%) were built or rebuilt with AI citation in mind, typically post-mid-2024; their pages score 7 or 8 of 8 on the rubric and they sit almost entirely in the top quartile of citations. Retrofit sites (~38%) added GEO investment after mid-2024 and score 4-7 of 8; they sit across Q2 and Q3. Legacy sites (~50%) have not invested in GEO at all and score 0-3 of 8; they sit overwhelmingly in Q1 with a long tail into Q2.

The mobility implication is the part that matters. Retrofit sites that added the four highest-leverage rubric items (llms.txt, comparison page, schema, answer-format H2s) moved an average of 1.6 quartiles in 90 days — Q1 sites mostly into Q2 or Q3. The top quartile is reachable on a roadmap, not just by AI-native rebuilds.

The eight factors below are the rubric items we scored every page against, ranked by their per-factor lift on citation rate after controlling for domain authority, recency, and content depth. Each lift is the percentage uplift in citation rate associated with the factor being present versus absent — derived from the 500-site point-biserial correlation, then expressed as a per-factor uplift for readability.

A few observations about the ranking. The top three factors (comparison, llms.txt, answer-format H2s) compound multiplicatively — pages that have all three score in the top decile across all four engines, not just one or two. The bottom two factors (use-cases, plain-language hero) are weaker individually but they map to a different mechanism: they make the page parse cleanly into LLM context, which lifts citation rate even on pages that already score well on the structural items.

The factor that is conspicuously not on the list is page recency. We expected fresh content to outperform — it does not, materially. Median cited page age in the dataset is 14 months. Engines treat SaaS landing pages as evergreen reference material rather than news, and the citation behavior reflects that.

The four engines do not behave the same way. ChatGPT cites SaaS pages most aggressively (average 6.1 SaaS citations per AI Overview-equivalent answer); Perplexity is at 4.8; Claude at 3.6; Gemini at 2.9. More importantly, the engines weight the rubric differently — what wins citation in one engine does not always win in another. The dominant pattern shifts from comparison-bias in ChatGPT to depth-bias in Perplexity to methodology-bias in Claude to schema-bias in Gemini.

The pragmatic implication is that an audit needs to weight the engines by traffic mix. A SaaS vendor where most product research happens in ChatGPT should over-invest in comparison pages and structured comparison blocks. A vendor selling into developer audiences (where Perplexity and Claude over-index) should over-invest in depth — long-form pillar content, methodology pages, named frameworks. A vendor optimizing for AI Overviews traffic from Google should over-invest in schema and markdown docs.

The 8-point rubric is the cross-engine baseline. The engine-specific tilt is the second-pass optimization once the baseline is in place.

The rubric below is the same one we ran across the 500 sites. Every item is a binary — present or absent — and every item maps to a measured citation lift in the audit. Out of 500 sites, the median score was 3 of 8. Anything above 5 puts a site into the top half of the dataset; anything at 7 or 8 puts it into the top quartile. We use this rubric inside our own engagements as the opening diagnostic.