Research · Report
State of AI Search: the engines disagree
By Logan Adams, Founder Reviewed & updated Measurement-first: figures are median share-of-model with 95% confidence intervals. How we measure.
When buyers ask AI for the best tool in a category, the answer depends on which AI they ask — and the sources behind it are mostly not the vendor's own site.
You're reading the online version. Download PDF (print) →
Abstract
Buyers increasingly ask generative AI assistants which software to use, and the answer they get depends on which assistant they ask. This report measures that disagreement directly. The same buyer-prompt set is run against ChatGPT, Perplexity, Gemini, Claude and Grok, each prompt repeated at least ten times per engine because a single AI answer is a sample of one, and every vendor named is recorded. From those answers we compute each vendor's share of model — the percentage of qualifying answers that mention it — with a 95% Wilson confidence interval over the pooled mention denominator, so two vendors whose intervals overlap are reported as indistinguishable rather than ranked. We find that no single vendor leads on every engine in the categories measured, and that the sources an engine cites behind its answer are predominantly third-party pages rather than the vendor's own site. The practical consequence is that AI visibility is engine-specific and source-driven: it cannot be inferred from one screenshot, and it cannot be worked on as if all engines were one channel. Consumer applications and APIs are measured as separate streams and never summed. The underlying leaderboards are published openly under CC BY 4.0.
The finding
We run the same buyer-prompt set against ChatGPT, Perplexity, Gemini, Claude and Grok — each prompt repeated 10+ times per engine, because a single AI answer is noise, not data. We record which vendors are named and how often, then compute each vendor's share of model: the percentage of qualifying answers that mention them, with a confidence interval.
No single vendor led on all engines. A name that dominated one engine was frequently absent from another. The "best tool" a buyer hears depends less on the product and more on which assistant they happened to ask — and on which third-party sources that assistant tends to cite.
What the leaderboards show
Measured category leaders (2026-07-01) from our AI Visibility Index — adaptive per-engine sampling across 5 engines (5-run floor; share-of-model Wilson 95% CI (pooled mention denominator), presence Wilson 95% CI). See each leaderboard for the full ranked table and per-engine spread. Every public Index category is now a live measurement — no illustrative previews.
Why it matters
If your visibility is strong on one engine and invisible on another, you lose shortlist spots you will never see in your analytics. The flip side: because the inputs are knowable — structured data, entity signals, and the specific third-party sources each engine pulls from — this is fixable. It just has to be measured per engine and worked per engine.
Method & honesty
Last reviewed: July 6, 2026. We re-check figures on a monthly cadence because AI engines change continuously.
Get the full dataset
The per-category CSV/JSON datasets (CC BY 4.0) plus the methodology notes. Tell us where to send them.
Logan Adams · Founder, Clear Cited
Writes on how AI answer engines pick what to recommend, share-of-model methodology, and reproducible AI-visibility measurement. About Clear Cited →