Start with a scan. Test an integration when you need to.

In my runs, one vendor was rejected nineteen times out of twenty and chosen once. One signup endpoint returned 200 once and 403 four times within an hour.

I record each run separately. The published audits name the models, run counts and scaffold. They also describe what a different scaffold would have changed.

Four to six runs per cell can reveal a failure every run hits. That sample cannot reliably rank vendors with close results.

Read the four published build studies

For discovery, I give an agent a developer's problem and deadline in an empty directory without prior context about your product. I record names, mention order and wording. Discovery runs create no accounts or credentials and submit no forms. I can run them on any domain before speaking with its owner.

Build runs require your agreement and someone watching the agent. I run them within a paid audit.

Discovery measures whether an agent names you. Build runs test whether it can use your product. I have observed vendors failing either test while passing the other.

The matcher uses a published list of names. I quote ambiguous matches, such as brands that are ordinary English words, for human review. Those matches are excluded from the count.

I check production access, package and tool registries, search results and agent choices.

Edge access. I scanned all 181 domains in the corpus between 24 August 2026 and 26 September 2026. 17 refused requests at their edge from my scanner's address. Of those, 13 returned a browser challenge rather than a rate limit.

On 20 August 2026, I retried from an ordinary home connection. Three hosts answered normally; one challenged that address too. Access varied with the request's origin.

One documentation host challenged all thirteen requests, leaving three checks unmeasured while its repository was unchanged. When I retried that host alone an hour later, it answered and all three checks returned.

Registries and search. I check package types under names an agent would guess and listings in tool registries. I also read the two lines about pricing that search results show before an agent opens your page.

Agent choices. I read the reasons agents give for choosing another provider.

Scan variation. I measured 0.59 percent of verdicts changing between two clean scans with nothing changed.

Keep repository checks in CI. Use monitoring to follow production access and external services over time.