We audited every page AI assistants cite about our own industry. Most of their numbers have no source.

When a buyer asks an AI assistant which AI-visibility tool to use, the assistant does not know the answer. It retrieves a handful of pages and synthesizes them. On our week-1 panel scan (2026-08-06, 130 vendor-selection answers across five engines), the retrieved layer was remarkably narrow: about nineteen third-party comparison articles supplied most of the citations, with the single most-cited article appearing in 12 of 130 answers. Whoever writes those pages is, functionally, writing the AI's answer. So we audited them.

What did we audit, and how?

We took the ten most-cited pages about our own category and counted three things on each, with uniform definitions: quantitative claims (body sentences asserting something about the world with a number), how many of those carried any source (a link or named attribution near the claim), and how many traced to a primary source (the original study, dataset, or announcement rather than another blog). Counts are approximate within two, applied identically to every page, our own included. All pages were read on 2026-08-07; the audit file is part of our research record.

What did we find?

Panel-wide: roughly 464 quantitative claims, about 25% sourced at all, about 14% primary-sourced. And the distribution matters more than the average: 52 of the 65 primary-sourced claims sit on just two pages, the vendor research studies from Profound and Semrush, which are first-party data and cite themselves legitimately. Exclude those two, and the third-party comparison layer that feeds AI answers sources 15% of its numbers (62 of 412) and traces 3% (13 of 412) to a primary source.

The pages that write the AI's answers publish statistics at scale. Sixty-two of their 412 numbers carry a source. Thirteen trace to the original.

What practices did the audit surface?

We describe practices here without attaching negative findings to named companies; the audit file preserves specifics. Observed across the panel: pages whose published and modified timestamps sit 12 to 15 milliseconds apart, the signature of fully programmatic publishing; a page targeting a cost query whose body contains almost no prices; a page whose modified date advanced the same day we audited it, nine days after publication, with no visible content change; invisible zero-width characters embedded inside headings, an artifact consistent with machine-generated text pasted from another tool; a market statistic cited to a link that opens a chatbot conversation rather than a study; and an affiliate link styled as a citation.

Two practices deserve credit by name, because they are what good looks like. Semrush's AI Overviews study is the genre's standard: a stated methodology (10M+ keywords, disclosed data windows), a named author with contributors, and a refresh-in-place canonical URL that accrues citations at one address instead of splitting them across yearly copies. And Profound's citation-sources research is built on a genuinely large first-party dataset, 11.8 billion citations by its own description; its weakness is the inverse of everyone else's, strong data with almost no machine-readable hygiene (no schema, no dates, no author markup). Question-form titles and honest self-placement, where we saw them, also measurably coexist with citation growth.

Does source quality even matter to the engines?

The uncomfortable research answer: less than you would hope, today. Wan, Wallace and Klein found that when language models choose between conflicting web sources, they respond strongly to topical relevance and largely ignore the credibility signals humans value, like scientific references and neutral tone. Venkit and colleagues measured answer engines retrieving a mean of 4.31 sources but citing only 3.0. Algaba and colleagues found LLM-generated references skew heavily toward already-highly-cited work. And the original GEO paper measured content-side optimizations lifting answer visibility by up to 40% on its benchmark. Sourcing rigor is not currently the engines' selection criterion; relevance-shaped, quotable content is. That is exactly why the citation supply looks the way it does.

We publish this knowing both readings. The cynical reading says unsourced listicles win, so why source anything. Our reading is that the citation supply is a market with a quality vacuum at the exact moment usage is exploding, and Cloudflare's 2025 measurements say the machines are already a measurable share of the web's readership, with AI bots at 4.2% of HTML requests and Anthropic's crawl-to-refer ratio reaching as high as 500,000 to 1. When engines sharpen their source discrimination, and the research direction points that way, the pages that survive re-evaluation will be the ones whose numbers hold. Ours are built for that reader.

The self-audit

The same rubric, pointed at us. Our previous research post carries 41 ledgered claims: every instrument figure traces to a dated scan, every external statistic was verified against its primary source before publication, and the two claims our verifier could not confirm at a primary source were held out of the article until they were. This post's claims ledger is built the same way. We also state what we are: a measurement vendor writing about its own category, ranked at 0.0% in the very answers this supply chain produces, which is the fact that makes the audit worth publishing and the reason to check our work.

What happens next

This page's own citation uptake is now a tracked experiment: our weekly panel measures which sources AI engines cite on these exact questions, and this article either enters that supply or it does not. Either result gets published on the scoreboard. That is the difference between writing about the citation supply and being measurable inside it.

Sources

1. Aggarwal et al., "GEO: Generative Engine Optimization", KDD '24
2. Wan, Wallace & Klein, "What Evidence Do Language Models Find Convincing?", ACL 2024
3. Venkit et al., "Search Engines in the AI Era", FAccT 2025
4. Algaba et al., "Large Language Models Reflect Human Citation Patterns with a Heightened Citation Bias", NAACL 2025 Findings
5. Cloudflare Radar 2025 Year in Review
6. Semrush AI Overviews Study (10M+ keywords, 2025)

Every external figure above was verified against the primary source before publication. Our own figures come from the instrument described on the methodology page; the claims ledger for this post is part of our research record.