What the evidence actually shows

Proven: agents cannot read a page that needs JavaScript

Vercel and MERJ instrumented AI crawler traffic and found that none of the major AI crawlers execute JavaScript — OpenAI's GPTBot, ChatGPT-User and OAI-SearchBot, Anthropic's ClaudeBot, Meta, ByteDance and PerplexityBot among them. Content that only exists after the client runs your bundle is content they never see.

⚠️ Two honest caveats. That study is from December 2024 and we have not found an equally rigorous replication since, so treat it as the best available rather than current. And the headline 404 rates in it (ChatGPT 34.8%, Claude 34.2%) are substantially crawlers re-requesting stale static assets — they are not evidence that a third of your pages are broken, and we will not use them that way.

Proven: markdown is dramatically cheaper to read

Cloudflare measured one of their own blog posts at 16,180 tokens as HTML and 3,150 as markdown — about an 80% reduction.

⚠️ That is Cloudflare's number, not ours. entreprenoid does not measure the size of a twin against the HTML it came from — the source is discarded once converted, so the comparison is not even recoverable afterwards. We are not going to print a percentage we did not measure.

And nobody has shown that cheaper documents get selected more often. Low token cost is a real benefit to whoever is paying for the context. It is not a ranking factor.

Proven, and causal: what a page says changes whether it is used

The one peer-reviewed, controlled result in this field. Across 10,000 queries in 25 domains, rewriting pages to add:

• quotations from relevant sources — up to +41% visibility
• statistics — +33%
• cited sources — +28%
• and keyword stuffing — −9%, making things actively worse

This is why the pages entreprenoid proposes are built out of quoted passages from your own content with sources attached, rather than generated prose. The evidence points at substance, not at format.

Not proven: that serving markdown gets you cited

Two controlled experiments have tested this, and both came back null.

• Profound — a competitor — randomised 381 pages across 6 sites for 21 days. Their conclusion: "We tested the claim that serving Markdown to AI crawlers drives more bot traffic. The data doesn't support it."
• Otterly published .md mirrors alongside their HTML for 14 days. The markdown pages received zero crawler visits and zero citations. (That tested separate .md URLs, which is a different mechanism from same-URL negotiation — but it is evidence, and we are not going to wave it away.)

A separate 44-day log study of Cloudflare's markdown feature found GPTBot, PerplexityBot and ClaudeBot sent no markdown requests at all. Today the clients asking for markdown are mostly coding agents.

Not proven: that publishing more pages helps

Across 75,000 brands, "number of pages on the site" was the weakest factor Ahrefs measured against AI visibility, at a correlation of about 0.17 — far below brand mentions elsewhere on the web. Volume is not the lever.

It is why entreprenoid caps a run at five suggestions, requires a real gap, and refuses to propose a page for a path that already answers. A tool that helped you publish at volume would be selling you into Google's scaled-content policy.

Disproven: llms.txt as a ranking tactic

Google has stated that llms.txt has no effect on Search or AI Overviews. Ahrefs looked at 137,000 domains and found 97% of llms.txt files received zero traffic in May 2026 — nothing fetched them at all.

entreprenoid still reads yours and tells you where its links point, because a file that sends agents to HTML when a twin exists is a real defect in something you already published. That is a different claim from "this file will rank you", which we do not make.

What we cannot measure at all

Whether you were cited. There is no rank tracker here and no answer scraper. Nothing in this product observes what ChatGPT, Gemini or Perplexity said about you.

What we do observe is your own server: which AI clients arrived, what they asked for, what format they got, what 404'd — and, separately, when a person arrives with a referrer belonging to an assistant. That last one is a click, not a citation. It misses every citation that produced no click, and we label it as what it is.

We learned that one the hard way on our own dashboard: a tile reading "visits this vendor sent back" showed 5 for Google, because only Gemini counts as an assistant while Google Search had sent 108. We narrowed the label rather than widening the query.

So why build it

Because two of these are true at once, and most of this category only says the first:

• A page an agent cannot read cannot be quoted. That failure is real, measurable at your own server, and fixable — whether or not the uplift is proven.
• Nobody has shown the fix produces citations. So we will tell you what changed at your server, and we will not invent a number for what changed inside a model.

And one argument that needs no study at all: a CDN that converts your HTML on the fly leaves nobody able to read, approve or withdraw what an agent was told. A twin is a document you control. That is a fact about governance, not a performance claim.