What we found when we tested 50 agent-readiness signals against three LLMs

Two of the three major LLMs disagree about what makes a brand worth listing

Imagine you run a B2B SaaS company. You read the latest agent-readiness guidance and decide to invest. You add a clean robots.txt. You publish a sitemap. You expose OAuth discovery metadata at .well-known/openid-configuration. You configure your server to respond appropriately to Accept: text/markdown requests.

You ship. Then you check whether AI assistants list your company in response to "top brands in [your category]" queries. According to our 2026-04 study, here's the pattern in the data:

• Claude is more likely to list you if your site has these features
• GPT is more likely to list you if it doesn't

Both correlations are statistically significant after multiple-testing correction. We're not measuring noise. The two LLMs select brands by structurally different criteria — at least on these four specific checks.

This is the most counterintuitive finding of our study, and we want to lead with it because the practical implication matters: a single agent-readiness score that optimizes for visibility across all major LLMs is not achievable — at least not with the signals we measured.

The rest of this article unpacks what we found, what we didn't find, and what it means for anyone trying to think about agent-readiness in 2026.

What we measured

We took 908 brands across 50 B2B SaaS categories — every brand that at least one of three LLMs (Claude, GPT, Gemini) mentioned when asked for the top 10–20 brands in a category. Then we ran three independent agent-readiness scanners against each brand's website:

• Respectarium's Agent-Adoption Check (the closed-source tool we built and operate)
• Cloudflare's public isitagentready.com API
• Fern's open-source afdocs CLI

That gave us 66 per-check measurements per brand — 25 from Respectarium, 19 from Cloudflare, 22 from Fern. Plus each scanner's aggregate readiness score and one derived “is the site blocking our scanner?” flag — bringing the total to 72 predictors going into the analysis.

Before testing, we excluded 22 of those 72. Twenty had less than 5% real-world adoption (no variance to detect anything statistical — we cover those separately below in the ecosystem-nascency section). Two more were dropped by additional pre-registered data-quality filters. That left 50 predictors evaluated against five LLM-visibility outcomes.

Combined with the LLM rank data, we ran 11 statistical analyses on the merged dataset to test which signals predict whether an LLM lists a brand and where it ranks.

Important caveat baked into how we collected the data: every brand in our dataset was already mentioned by at least one LLM. Brands that no LLM mentioned never entered the study. So our findings are about relative ranking among already-mentioned brands, not about whether agent-readiness gets your brand mentioned in the first place. We can't answer the second question without a different sampling design.

We pre-registered the rules before looking at results

Before any analysis began, we wrote down the criteria a signal must meet to be classified as "predictive enough to score." The rules were committed in writing on 2026-04-24 — two days before we ran the analysis. They are applied mechanically by our analysis script.

For a signal to graduate to PROMOTE_SCORED status, all four of these must be true:

1. The signal correlates with at least one outcome with FDR-adjusted p < 0.05 (a standard multiple-testing correction)
2. The effect remains statistically significant after we control for category (which is why something looks predictive across the dataset can disappear when you compare apples to apples within a single industry segment)
3. At least 2 of 3 LLMs reward the signal in the same direction
4. The signal isn't redundant with another, stronger signal that already qualifies

Why pre-register? Because without pre-registration, statistical analysis becomes a fishing expedition where you can find "significant" results just by trying enough combinations. With pre-registered thresholds applied mechanically, the result you get is the result you get — no shopping for a story.

The result: 2 of 50 tested signals graduate.

What graduated, and what it means

Two predictors passed all four pre-registered criteria:

These two signals capture different layers of agent-readiness: protocol-level clarity (robots.txt, AI-bot rules) and content-presentation flexibility (markdown alternatives to HTML). Both modestly predict where a brand lands in LLM rankings — once that brand has been included in the LLM's pool of candidates.

The effects are real but small

Here is every signal that survives the strictest test we ran (Welch's t-test with FDR correction), expressed in the original outcome scale rather than abstract correlation coefficients:

These are real differences. A 10-point improvement on a 0-100 quality score is observable — readers can feel the difference between a brand that scores 65 vs 75. A 3-position rank improvement is noticeable when the listing is 1–20 deep.

But these are also clearly not transformative. None of these effects reach the conventional "large effect" threshold (Cohen's d > 0.8). The honest framing is "agent-readiness is a small but real contributor to LLM visibility" — not "implement these checks and watch your brand suddenly rank higher."

The narrative space between those two claims is where careful interpretation lives. Industry voices that promise dramatic LLM visibility wins from agent-readiness adoption are getting ahead of the evidence.

Why Claude and GPT disagree (some hypotheses)

The four signals where Claude and GPT diverge — sitemap-exists, oauth-discovery, robots-txt-exists, markdown-negotiation — are all checks from the Respectarium scanner. Claude associates them positively with listing a brand. GPT associates them negatively. Three plausible mechanisms:

• Hypothesis 1 — Training-data recency: GPT's training corpus includes a lot of older web content. Many household-name B2B brands appeared in that older content extensively, regardless of their current agent-readiness posture. GPT may rely more on training-time brand recognition; the absence of agent-readiness signals doesn't penalize a brand it already "knows." Claude, with newer training data, may weight current-state crawler signals more directly.
• Hypothesis 2 — Crawler accessibility: GPT's crawler may be blocked by sites with stricter robots.txt configurations — including AI-bot-specific blocking rules. The "negative" direction of the GPT correlation might reflect "sites without robots.txt are implicitly permissive to all crawlers, including GPT's." Sites that bother having a robots.txt file are more likely to also have rules that block GPT specifically.
• Hypothesis 3 — Selection by listing process: When asked for "top brands in category X," each LLM applies its own implicit selection function — combining training-data recall, retrieval, and ranking. Different LLMs converge on different brand subsets. The agent-readiness signals correlate with category structure differently for different LLMs because each LLM's category-scope is structurally different.

We can't disambiguate these mechanisms with cross-sectional data alone. What we can confidently say: the disagreement is real, statistically significant, and has practical implications for anyone designing a "universal agent-readiness score." Such a score is structurally constrained by these reversal effects — any single number will trade Claude visibility for GPT visibility on these specific signals. Universal optimization is unreachable; weighted compromises remain possible.

Specs exist; practice has barely arrived

Of the 66 per-check signals we measured (the bulk of our 72 predictors), 20 had less than 5% real-world adoption in our 908-brand sample — they were among the 22 we excluded before testing. We couldn't even include them in the correlation analysis — there's no statistical signal to detect when 95% of brands sit in the same status bucket.

The list of un-adopted checks reads like a tour through the bleeding-edge agent-protocol future:

• MCP server cards (Model Context Protocol — for agentic interactions)
• A2A agent cards (Agent-to-Agent communication declarations)
• OAuth protected-resource metadata (formal OAuth 2.0 endpoint discovery)
• AGENTS.md files (proposed convention for agent guidelines per repo)
• web-bot-auth (cryptographic proof-of-bot identity)
• content-signals (declared content-usage permissions for AI training)
• The Cloudflare commerce-protocol stack: x402, mpp, ucp, acp, ap2 (agent-to-payment protocols)

Adoption rates for these range from 0.0% to 3.6%. The specifications are public, documented, and championed by major industry players. The practice has not yet arrived.

This is itself a publishable finding. The agent-readiness standards landscape in 2026 is comparable to where the AMP standard was in 2016 or where structured data was around 2012-2013: the technology exists, the documentation is available, the major players are advocating for it, and almost nobody has actually implemented it yet. Adoption follows specifications by years, not months — and we now have a baseline measurement to track that adoption against in future quarterly studies.

The unflattering finding we published anyway

Respectarium operates one of the three scanners we evaluated. We told you that upfront in §1 of the preprint. Here's the consequence: our scanner produces a 0–100 aggregate score for each brand it scans. We tested whether that score predicts LLM visibility outcomes.

It doesn't.

Specifically: the Respectarium scanner's score aggregate has a mean correlation of 0.016 across all five outcome variables. The FDR-adjusted p-value is 0.69 — far above any reasonable significance threshold. Effectively no predictive power.

For comparison, Cloudflare's level aggregate has a mean correlation of about 0.11 — almost an order of magnitude stronger. Some of the Respectarium scanner's individual checks (like markdown-negotiation) graduate to PROMOTE_SCORED status. But the way the v1 weighting scheme combines those checks into a single score dilutes the signal of the few actually-predictive ones.

We're publishing this finding because methodological transparency is the credentialing mechanism of this whole research program. If we quietly buried the result that our own scanner's aggregate score doesn't predict outcomes, we'd be exactly the kind of vendor-research operation we want to distinguish ourselves from.

Practically, this finding is the empirical foundation for a v2 specification. The next version of the Agent-Adoption Specification will rebuild the scoring formula around the signals that survived our pre-registered tests, rather than the more comprehensive but signal-diluted v1 weighting.

What this means for you

What's next

We will repeat this study quarterly. The next release (study-2026-07) will:

• Re-measure the same 908-brand universe (or expanded to ~1500 if our sample expansion succeeds)
• Use the same 11-script analytical pipeline applied unchanged
• Compare directly to Q1: which signals strengthened, which weakened, which newly-adopted checks now have measurable variance
• Publish to the same GitHub repository as a new immutable tag

Past tags remain accessible forever. The comparative time series across quarterly tags is itself a research output — we expect to be able to make stronger longitudinal claims by Q3 (~6 months out) and meaningful causal-direction claims by Q4 (~9 months).

We're also working on a v2 of the Agent-Adoption Specification that incorporates these findings into the scoring formula. Specifically: rebuild the score around the predictively-validated signals, treat bot-protection as a covariate rather than a quality penalty, and demote the bleeding-edge agent-protocol checks to "informational" until adoption produces variance to test against.

Where to read more

• Formal preprint: respectarium.com/research/correlation-2026-04 — full methodology, complete results tables, references, conflict-of-interest disclosure
• Source code, data, results: github.com/respectarium/agent-adoption-research — MIT-licensed code, CC-BY 4.0-licensed data, full reproducibility instructions
• Agent-Adoption Specification V1: respectarium.com/spec/agent-adoption/v1 — the open specification the Respectarium scanner implements
• Run the Agent-Adoption Check on your own site: respectarium.com/agent-adoption-check

For updates on quarterly re-runs and v2 spec development, watch the GitHub repository.