Before AI Understands Your SiteIt Must Translate It FirstWE SPEAK AI™

How'd we do it?
Well, We Asked.
...100,000 times.

• We asked all the major AIs how we could maximize the AI probability of our experimental site, top10lists.us, about 20,000 times (each) over a period of 5 months.
• The result? They all consider it the gold-standard exemplar of AI usability — 4M+ crawls a month, with a user-initiated ratio twice the industry standard. See the receipts.

Along the way, we found 12 signals that affect AI citability that we can measure. Once measured, we found that we could tune the levers to maximize AI ingestion, reasoning and trust.

AnswerShare reroutes bot traffic at the edge

[image: Diagram showing AnswerShare™ detecting AI/bot traffic at the edge and rerouting it to optimized JSON-LD content (useless characters stripped, dataset-optimized, recordings transcribed, Query Fan-Out survives, ultra-fast delivery, source-grounded, auto-updated by SOT) while human traffic passes through to the production site unchanged. Production site remains source of truth — no CMS change, no workflow change, no security impact, ~30-minute implementation.]

AI Opinion -- three NPS-family metrics.

What AI systems actually think of the brand, compared to what the human community says. Direct measurement on both sides.

ΔNPS™ (delta-NPS) is the comparison: positive means AI rates the brand better than the human community does; negative means worse.

Measures the human community's opinion of the brand online -- the corpus baseline AI systems are trained against. Direct measurement.

Measures the AI's opinion of the brand in live retrieval. Direct measurement against the live frontier models.

ΔNPS = λNPS − μNPS. Positive = AI rates the brand better than the human community does. Negative = AI rates the brand worse than the human community does.

Infrastructure Readiness -- four metrics, as measured through the cache.

What AI bots actually receive when they hit the property -- composite infrastructure health, edge-cache-served latency on both ends of the response, and survivability across query fan-out.

Composite rollup of 8 underlying binary AI-bot-accessibility signals. Pass = 8/8 passing. Partial = 2-7/8 passing. Fail = 0-1/8 passing. The eight underlying signals are itemized below.

Edge-cache-served time to first byte from a clean IP. Pass if <= 200ms. AI crawlers operate on finite session budgets -- slow first bytes shrink how many pages get reached per visit.

Edge-cache-served time to last byte from a clean IP. Pass if <= 500ms. TTLB matters more than TTFB for AI crawlers because they consume the fully-rendered payload, not just the initial response.

When an AI arrives at your site, it doesn't just use one prompt — it also fires off 5–10 sub-queries on the same topic. For example: you ask, "Recommend an SEO agency in my town." The AI lands on your site, but it also runs prompts like "Recommend an SEO agency in [state]," "in adjoining cities," and "best SEO agency for [profession]" — and so on. Your QFS score is the percentage of those sub-queries that your site answers.

The 8 binary signals inside Infra Score

Each is a Pass/Fail individually but rolls up into the Infra Score above. AI crawlers check for these signals on first contact -- every one is independently verifiable from public endpoints.

robots.txt explicitly allows the AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, etc.). Default-blocked sites are invisible to AI by construction.

/llms.txt exists and lists the canonical pages an AI should ingest. The AI equivalent of a sitemap-for-attention.

/llms-full.txt exists with full-text dumps of those canonical pages. Lets AI ingest authoritative content in one hop instead of crawling the whole site.

sitemap.xml resolves with <lastmod> timestamps inside the last 60 days. Stale sitemaps tell AI the site isn't maintained -- it deprioritizes citation.

Schema.org JSON-LD is present and valid on every page. AI uses structured data to resolve entities (Person, Organization, Place) without natural-language inference.

The page renders meaningful content (>= 1500 chars) in the initial HTML, not via client-side JavaScript. Most AI crawlers don't execute JS; SPA shells without prerender are invisible.

/.well-known/mcp.json resolves and the declared MCP endpoints respond. The Model Context Protocol surface lets AI clients query data live during inference.

/ai-content-index.json exists and enumerates the machine-fluent payloads (artifact protocol). Gives AI a one-request manifest of everything you want it to ingest.

Reasoning Keys -- five continuous metrics.

How cleanly the content parses to AI once Infrastructure Readiness is in place. Each metric has a defined threshold -- below threshold means AI is actively penalizing the site in retrieval. All five are direct measurement.

How closely the bot-served HTML matches the human-served HTML. When the two diverge, AI is reading less than half the content. Low RR is the silent killer of citation eligibility.

Live retrieval and RAG require that assertions be grounded to a reputable 3rd party. Without these, the AI considers it “unverified” and deprecates it as “Marketing Fluff.”

Ratio of page-chrome/overhead tokens to useful-content tokens. Higher RTC means AI burns its token budget on navigation, scripts, and ads instead of reasoning over your content -- making citation less likely.

RPC™ (Retrieval Pages Crawled) measures how much of a site an AI system can realistically traverse during a single answer-generation cycle. Modern AI systems commonly fan out into multiple retrieval probes while composing an answer. Each probe operates under bounded crawl time, limited concurrency, and probabilistic retrieval budgets. RPC estimates how many pages the system can successfully fetch, parse, and evaluate before generation completes, given the site's response speed, error rate, crawlability, and retrieval structure. Higher RPC means more of the site's content reaches the AI citation and reasoning pool. Lower RPC means the AI forms opinions from a shallow subset of the available corpus. RPC replaces traditional requests-per-second metrics, which measured server throughput rather than practical AI retrieval reach.

Per-visit crawl capacity for AI bots -- a concrete page count, comparable across sites of any size.

• • T = bot session-burst duration (35s, calibrated) -- duration of one bot session-burst, not time per page
• • C = per-host concurrent fetches (3, calibrated)
• • P_success = probe success rate from 8-request bounded crawl simulation from a clean IP
• • TTLB_p75 = 75th-percentile time-to-last-byte (seconds) -- TTLB is used instead of TTFB because AI crawlers consume fully rendered payload time, not just initial response latency.
• • RPC is the absolute number of pages a bot can read per visit -- there is no fan-out multiplier.
• • N_eligible (no longer in the headline formula) is retained only for the secondary Coverage diagnostic: Coverage = RPC / N_eligible x 100.

T = 35s and C = 3 derived from 23,905 AI-bot session bursts (PerplexityBot, ChatGPT-User, OAI-SearchBot, GPTBot, ClaudeBot) observed across 64 days on Top10Lists.us, full per-request instrumentation, 2026-03-23 -> 2026-05-26.

Median age of the Last-Modified header on canonical URLs. Live-retrieval AI deprioritizes content older than 30 days for time-sensitive queries. Stale pages get demoted regardless of quality.

Methodology documentation by signal.

Every methodology page includes frozen receipts -- the exact commands, responses, and timestamps used to produce the published benchmark. Measurement is reproducible by design.

The translation layer between your website and AI systems.

Contact Us

Quick Links

Measurement is methodology-driven and reproducible. Frozen receipts published per metric.

© 2026 AnswerShare. All rights reserved.