AI Visibility

We Stress-Tested Our Own AI Visibility Audit. Here's Everything We Broke and Fixed.

By Adev Aarons · August 1, 2026

Most free “AI visibility checkers” are a robots.txt lookup wearing a trench coat. Before we pushed ours harder, we tried to break it — the same way a skeptic would. This is the full log: what we threw at it, where it was wrong, and what changed.

Test 1: Why does reddit.com score 9/100?

The obvious “gotcha”: run one of the biggest sites on the internet through the audit and screenshot the F grade. People told us “Reddit allows crawlers.” So we went deeper and probed Reddit with four different identities. Live results:

As a normal browser. robots.txt loads fine. It opens with “Reddit believes in an open internet, but not the misuse of public content” — and then blocks every crawler on the planet in two lines: User-agent: * / Disallow: /.

As GPTBot (ChatGPT’s crawler). HTTP 403. Refused at the network level, before robots.txt even matters.

As a fake Googlebot (right user-agent, non-Google IP). A block page that literally says “whoa there, pardner — your request has been blocked due to a network policy.” Reddit verifies Google’s crawler by IP address. Imposters get walled.

So who gets in? Google’s real crawler — because Google pays Reddit a reported $60M per year for access. That’s why Reddit dominates Google results and AI Overviews. OpenAI also pays, but receives Reddit data through Reddit’s Data API, not crawling. Which resolves the paradox: ChatGPT knows everything about Reddit while Reddit 403s ChatGPT’s own crawler.

Verdict: the 9/100 is correct for what it measures — open-web retrievability, what an AI engine can fetch from a site without a licensing contract. Reddit can charge for access because it’s Reddit. A local business with those same signals is simply invisible. The audit now explains this distinction inside the report whenever it detects a deliberate blocker.

Test 2: A founder said his score was wrong. He was right.

A founder ran his site through the audit, got a low score, and pushed back — his site was genuinely optimized. We investigated instead of defending.

His hosting provider’s bot protection was serving our auditor an empty challenge page, so the audit was scoring a void: no title, no schema, no content. Meanwhile the real site — served to browsers — had all of it.

We rebuilt the fetcher: browser-grade requests, and when a firewall still challenges the direct fetch, an automatic residential-grade proxy fallback reads the real page. His score nearly doubled on re-run.

Then the fixed audit surfaced something he had no idea about: his host silently kills ChatGPT’s crawler at the connection level while his robots.txt says “welcome.” Robots.txt shows intent; only a live probe shows enforcement. His site is invisible to ChatGPT in real time, and nothing in a standard SEO stack would ever tell him.

Two rules came out of this and are now permanent: the audit never treats an unreadable page as proof of missing optimization, and every audit includes live fetches AS GPTBot and AS PerplexityBot — from datacenter IPs, the same network conditions real AI fetchers face.

Test 3: Calibration against reality

A score is only as good as its correlation with the thing it claims to predict. So we tested the prediction directly.

We asked Perplexity: “Who are the best real estate agents in Naples, Florida?” It named names and cited sources. We took one agent Perplexity actually cited and one comparable Naples agent it ignored, and ran both through the audit blind.

The cited agent outscored the ignored one — and the gap came from the exact pillars the research says drive AI citations: entity schema (17/20 vs 4/20) and content structure (14/20 vs 9/20). The audit’s built-in ask-the-AI check called both correctly: true positive on the cited agent, true negative on the ignored one.

The stress test also caught two bugs in our own tool — a token limit that truncated reports on snippet-heavy sites, and a test query that accidentally included the agent’s own name (making citation self-fulfilling). Both fixed the same day. That’s the point of stress testing your own product before your prospects do.

What the audit checks now

Crawler access with live bot impersonation probes. WAF and bot-protection detection with proxy-verified fallback. Entity and schema depth. Content citability using the Princeton GEO study factors (statistics density, quotable passages, source citations, chunk structure). JavaScript-rendering invisibility (AI crawlers never execute JS). Content freshness. Off-site corroboration via real search. Reddit and YouTube community presence. Category-query rankings and Google AI Overview citations. Wikipedia and Wikidata entity presence. And a live ask-the-AI test: two buying-intent questions put to Perplexity, with the report showing who made the answer and who didn’t.

Run it on your own site: specularisinc.com/free-audit. Two minutes, report in your inbox. If you think your score is wrong, tell us — the last person who did made the product better.