Evidence over guesswork.

AgentDoor samples AI-search responses to questions a customer might ask about your services in your area.

Four separate visibility measures

Mention rate, recommendation rate, first recommendation rate, and domain citation rate each count matching valid observations divided by all valid observations. Failed, skipped, incomplete, and ungrounded checks are excluded. No valid checks means insufficient data.

A sample, at a point in time

Provider API results are samples of AI-search behaviour and may differ from consumer application responses. Answers can change across users, locations, model versions, and repeated searches. We do not claim a universal AI ranking or statistical significance from a small sample.

Conservative detection

We match business names, declared aliases, and domains. Recommendations require positive recommendation wording or an identifiable business in a recommendation list. First recommendation is counted only when a literal numbered list begins with that business. Ambiguous language may be missed. Competitors are named entities detected in the answer, not every cited website.

Strong means recommended in at least 60% of valid samples for a search; Moderate means at least 30%; Weak means some mention or citation but less than 30% recommendations; Absent means neither mention nor citation. Each search currently has one sample per configured provider, so the report shows the sample count.

Website readiness is separate

The 100-point audit has six categories: crawlability (20), entity clarity (20), service clarity (20), geographic clarity (15), structured data (15), and machine accessibility (10). Each category weights its documented checks equally and rounds to the nearest point. It measures the pages we could read; it does not guarantee AI recommendations. llms.txt is inspected but contributes no points.

Facts and findings

Extracted facts retain a source and confidence. User-supplied search geography is labeled separately from website evidence. Phone discrepancies compare a response with the website and need owner verification; other factual claims remain available for manual review. Verified issues describe observed evidence. Optimization hypotheses are possible improvements, not proven ranking factors.

Costs and limits

Calls have page, request, output, and scan budget limits. We reserve an estimated budget before each call and stop scheduling checks when that budget runs out. Pricing estimates depend on configured rates; missing prices are shown as unknown. Reservations are not an invoice guarantee, because providers control search activity and billing.

Limited crawling

We read up to 20 HTML pages, prioritize contact and service pages, respect robots.txt, and never execute a target site's scripts. Sites that require JavaScript, block crawlers, or expose incomplete structured data may produce a limited audit.