dromad

Research

Dromad treats claims about AI search as hypotheses. When we think something matters, we try to measure it; when the evidence is weak, we say so; when a study goes against what we expected, we publish that too. Every Dromad audit finding says how strong its evidence is — deterministic, supported, suggestive or experimental — and these studies are why.

Each study froze its design in version control before collecting any data, and every number in it is recomputed from the raw data by the study's analysis script.

Do Dromad's checks predict what AI engines cite?

August 2026 · 50 questions × 3 engines × 3 repetitions · 1,834 pages audited

The question. Among pages that are all plausible search results for the same developer question, do the pages AI answer engines cite pass Dromad's checks more often than the pages they pass over?

Method. 50 developer questions (5 intents × 10 ecosystems) ran three times each through Claude, GPT and Gemini, with each engine's own web search forced on — left to choose, models answer from memory and cite nothing. The comparison set was the live Google top 10 for each question. Cited pages and retrieved-but-uncited pages were kept separate, and every page was audited with the checks unchanged.

What we found. At the page level, most of Dromad's checks did not separate cited from uncited pages, and not for one reason:

  1. Most checks barely fail on pages that compete for citations. A check that fires on 1% of real search results cannot tell them apart.
  2. The strongest differences were access and renderability — bot walls and JavaScript shells — which sit upstream of every content check.
  3. One structural check, heading_level_skip, showed a consistent association with citation. Two others flipped direction across groups and look like confounds.
  4. Domain authority ran backwards from the SEO intuition: at the same Google rank, engines cited pages on weaker domains more often.

What changed because of it. Audit facts come first and findings are derived from them; access and rendering are reported as facts (likely_js_dependent, robots.txt, the redirect chain); and findings carry an evidence level instead of implying every check matters equally.

llms.txt: presence, quality, discoverability

August 2026 · 104 developer-tool documentation sites

The question. Three questions usually get collapsed into one: is there a file at /llms.txt (presence), is it any good (quality), and does anything tell an agent it exists (discoverability)?

What we expected, and what we found. We expected that almost nobody makes llms.txt discoverable. The data did not support that: among sites that publish one, 66% point at it explicitly. The more useful finding was about Dromad itself. Its presence check counted any 200 response as a file, and sites that answer unknown paths with an HTML page were passing.

What changed because of it. Dromad reports llms.txt as found only when the response is a 200 with a non-HTML body, and the check that would have been cheapest to build — "is it linked from the homepage" — was not built, because the data says it would be wrong more often than right.

Dogfooding the toolkit: Exa on "AI search API" questions

September 2026 · 30 generated questions, 8 run on 3 engines, 5 page audits

We ran the investigation from the agent guide end to end through the public CLI, the way a coding agent would: a generated question set (the command was then called fanout), 24 prompt runs, one citation analysis and a handful of audits, for $1.42 in provider cost.

Counts from that run: exa.ai was cited in 9 of 24 runs (38%) and on 5 of 8 questions; Exa was named in 17 of 24 answers, as often as Tavily; the most-cited pages were third-party comparison posts rather than any vendor's own pages.

It also found two defects in Dromad's audits, both fixed before release: a "no h1" finding on pages whose h1 sat outside the main content region, and a homepage whose "main content" was one testimonial card out of nine. A finding that claims to be deterministic and is wrong is the worst kind, which is why the audit reports its facts next to every finding.