How to read the evidence in AI search visibility: sources we trust, sources we don't

The AI visibility category has a sourcing problem, and a public tier system is the way through it. Vendor blogs cite each other in circular patterns, and statistics with no disclosed methodology get repeated until they look authoritative. Our three-tier system separates evidence that survives audit from evidence that does not. Tier A is citable without qualification, Tier B is citable with a caveat, and Tier C is provisional. The sections that follow set out the tiers, name the sources used most, and give the reader a four-step check.

Sourcing transparency is the category's missing layer

The category repeats numbers that do not hold up. A claim originates in a single vendor case study and travels into an agency deck. It gets quoted in a conference talk, then arrives in a brand-team briefing with the original sample size stripped away. By the time a brand manager reads it, the figure looks like consensus.

Two structural habits drive the problem. Vendor blogs cite other vendor blogs, which manufactures the appearance of corroboration without adding evidence. Real research sits next to marketing collateral in the same search results, and the two are formatted to look alike. A brand team making a budget decision deserves a way to tell them apart.

The fix is not a longer reading list. It is a published standard for what counts as evidence, applied openly, including to this library's own provisional sources. Sourcing transparency is the layer the category is missing, and publishing it is the move that earns external citation.

Three tiers cover the practical space

The system has three tiers, defined by what a source discloses and how it was produced. Tier A is citable without qualification. It covers peer-reviewed academic work, enterprise-scale analytics providers with public methodology, major consultancies, large first-party studies with disclosed methodology, and government or regulatory sources. The Princeton GEO paper is a Tier A example, peer-reviewed and method-disclosed Aggarwal et al. · Princeton University / Georgia Tech / Allen Institute for AI / IIT Delhi · 2024A.

Tier B is citable with attribution and a one-line caveat. It covers credible trade press, case studies from named companies with disclosed methodology, and single-vendor studies with disclosed methodology. The control-matched schema study is a Tier B example, a single vendor's analysis with a clearly stated difference-in-differences design Ahrefs · 2026B.

Tier C is provisional. It covers vendor blogs, individual LinkedIn or Substack posts, and case studies with a single data point. A Tier C source is used only when a finding is genuinely novel and no Tier A or B source exists. It is labelled as provisional in the body. Finer-grained distinctions than these three do not survive contact with the work.

Tier A is the list of sources whose methodology survives audit

Tier A is not a wish list. It is the set of sources whose methodology can be checked. The most-used Tier A sources in this library fall into four groups.

The first group is enterprise-scale telemetry with public methodology. Pew Research Center publishes its panel and opt-in tracking method Pew Research Center · 2025A. Adobe Digital Insights aggregates Analytics data across trillions of visits with disclosed method Adobe Digital Insights · 2025A. Cloudflare Radar reports network-scale crawler telemetry Cloudflare · 2025A.

The second group is major consultancies with stated samples. Bain surveyed more than 1,000 consumers alongside proprietary analysis Bain & Company · 2025A. McKinsey published its forecast method for the AI search revenue projection McKinsey & Company · 2025A. The third group is peer-reviewed academic work: the GEO-16 empirical analysis arXiv · 2025A, the citation-absorption framework Zhang et al. · arXiv (cs.IR) · 2026A, and the OpenAI usage study Chatterji et al. · NBER / OpenAI / Harvard University · 2025A. The fourth group is primary-source documentation and government work: the Stanford AI Index Stanford Institute for Human-Centered AI (HAI) · 2026A, the Tow Center accuracy study Tow Center for Digital Journalism, Columbia · 2025A, and first-party crawler docs OpenAI · 2025A.

The statistics we no longer cite, and why

Publishing retractions is the move nobody else in the category makes. Several figures that circulate widely are no longer used here, each for a specific, checkable reason.

The first is the "23× AI conversion rate" claim. It generalises from a single SaaS case study with one data point, and the headline multiple shrinks dramatically once the underlying rates are disclosed. The replacement evidence is a study reporting AI traffic converting at roughly three times other channels, with its method stated Microsoft Clarity · 2026B. The second is the family of "FAQ schema lifts citations 2.5 to 3.2 times" claims, sourced to vendor blogs. The control-matched study found schema produced a +2.4% lift on Google AI Mode, statistically indistinguishable from zero Ahrefs · 2026B.

The third is the "44% citation lift from FAQ schema" figure, a single vendor study with unclear methodology. It is contradicted by the same control-matched evidence and by a 300,000-domain analysis that found no relationship between the relevant file and citation rate SE Ranking · 2025B. The fourth is a representative vendor-reported case of a 19.72% rise in AI Overview visibility after entity linking van Berkel · Schema App · 2025C. The figure is a single-data-point case study; treat it as provisional. The pattern across all four is the same: a clean primary source replaces a recycled vendor figure.

You can check a statistic yourself in four steps

A reader does not need this library to evaluate a claim. The check has four steps, and any one of them can fail a statistic.

The first step is to find the primary source. If a vendor blog cites another vendor blog, and that one cites a third, stop. A claim with no primary source is not a claim, it is a rumour with a number attached. The second step is to check the sample size and method. A study of one customer does not generalise to a category, and a percentage with no denominator hides its own weakness.

The third step is to check the date. AI search behaviour moves fast, so a 2023 figure may already be obsolete, and single measurements are unstable in the first place Mustahsan · arXiv · 2025A. The fourth step is to check whether the claim survives recasting. A "23×" multiple that collapses to "1.66% versus 0.15%" once both rates are shown was overstated by its framing. A reliable reading reports the figure with its method and variance Lior et al. · arXiv (cs.CL) / EMNLP 2025 · 2025A.

What we still don't know

Naming the gaps openly is the same discipline as naming the tiers. Several important questions have evidence too thin for a confident answer, and pretending otherwise would fail this article's own test.

The first gap is real-world prompt data. What users actually type into AI assistants is proprietary to the AI companies, so the category works from public proxies rather than direct observation. The second is the long-term effect of model updates on visibility trajectories, where no longitudinal Tier A study yet exists. The third is how agentic browsers reshape attribution and measurement, documented as products but not yet as behaviour OpenAI · 2025A. The behaviour of the second major agentic browser is equally undocumented at Tier A Perplexity AI · 2025A.

The fourth gap is how AI engines weight first-party against third-party content in equilibrium, which current evidence describes only in snapshots. The fifth is how regulation will affect retrieval in practice. The EU Code of Practice covers copyright and transparency European Commission / EU AI Office · 2025A. The US Copyright Office has acknowledged that fair use in AI training is fact-specific U.S. Copyright Office · 2025A. The operational effects on retrieval are still emerging.

The commitment behind every source list

Every article in this library applies the same rule. Tier A sources are used where they exist, Tier B with attribution and a caveat, and Tier C only as provisional and labelled as such. The source list at the bottom of each article shows the tier of every entry. The reader can audit the evidence without leaving the page.

The standard applies to this library's own weak spots. Where a finding rests on a single vendor case study, the article says so in the same sentence as the claim. Where a number is provisional, the word "provisional" sits next to it, not in a footnote. The crawler data this library leans on most is named with its source, so a reader can check the method directly Cloudflare · 2025A.

Credibility is the category's scarcest resource. A public tier system, an open retraction list, and a named gap list are how this library spends it carefully. If a figure here turns out to be wrong, the correction is published. A research surface that hides its errors is just another vendor blog.