Two weeks ago I published a checklist for making sites readable to AI crawlers. A few people asked the fair question: does any of it actually change what AI assistants say? So here are two before/after cases from my own projects, with live files you can inspect.
Case 1: comiclens.cloud — 40/100 → 100/100
A comic-reading app. Before: no llms.txt, no structured data, robots.txt silent on AI bots. When you asked assistants about it, they either knew nothing or guessed wrong.
What shipped (about an hour, mostly copywriting):
- robots.txt: explicit Allow for GPTBot, ClaudeBot, PerplexityBot
- llms.txt: one markdown page of hard facts — what the app does, platforms, pricing model, feature list (live file)
- JSON-LD: SoftwareApplication schema with real attributes
After re-crawl: engines that previously answered "I don't have information about this" started describing the product accurately and quoting the fact sheet.
Case 2: filmaro.app — 85/100 → 100/100
Already decent (good semantic HTML got it to 85). The gap was verifiability: no single canonical fact source, and the FAQ wasn't marked up.
- Added llms.txt (live) and FAQPage JSON-LD
- Rewrote vague marketing lines into checkable statements (numbers, platforms, dates)
The FAQ markup mattered most — assistants answer in Q&A format, and FAQPage schema hands them ready-made pairs.
What I'd generalize from this
- The floor is lower than you think. Most sites fail at step zero: the crawler is blocked or the site renders everything client-side with no fallback text.
- Facts beat prose. Every line an assistant actually quoted was a concrete fact: a year, a price, a platform name. Zero slogans survived.
- Re-check after model updates. Answers drifted between engine versions during the two weeks — what Perplexity says is not what Gemini says.
Verify it yourself
Both scores above come from the scoring tool I built, AnswerRank — it asks ChatGPT and Perplexity your buyers' questions and shows whether your site is in the answers (free, no signup). But you don't need any tool to sanity-check: open ChatGPT with browsing, ask "What do you know about {your product} at {your URL}?", ship the checklist, ask again in a week. The delta speaks for itself.
Top comments (2)
im curious if you actually saw a traffic bump from AI sources after adding the json-ld or if it was just cleaner logs
Honest answer: no traffic bump I can attribute to it. Both sites are small and AI referrals stayed at noise level — on my own site the analytics show 1 human session over the last 7 days, so there is nothing statistically real to report on that axis. I'd rather say that than dress it up.
What did change measurably was the answer, not the traffic. Before: asking ChatGPT/Perplexity with browsing "what do you know about {site}" produced either nothing or invented details. After: both quoted actual facts from the page — price, platform, year. That's what the 40 → 100 tracks: retrievability and quotability, not sessions.
Your suspicion is right in one specific place, though. I can't prove an LLM parsed the JSON-LD itself. What I can observe is that the facts the assistants quoted existed in both the schema and the server-rendered visible text, so my working assumption is that plain rendered facts did most of the work and JSON-LD mainly kept them unambiguous. llms.txt is weaker still: no engine has publicly committed to consuming it — cheap to ship, no evidence of a direct effect yet.
On the traffic side specifically, the only clean signal I know is grepping logs for chatgpt.com and perplexity.ai referrers. For me that's been ~0 so far. I'll post real numbers when there are any rather than implying there already are.