Last year we had a documentation set we were genuinely proud of. Thorough, well-written, nicely laid out.
Then I ran it through the models. Every answer that mentioned us cited third-party pages. Not a single one of our own docs made it into an answer.
The reason turned out to be embarrassingly simple: those pages were structured as essays. To pull one clean answer out of them, the model had to assemble it from three sections. Given a choice, models prefer sources where the answer is already sitting there in one piece.
Here's what I changed, and how to retrofit it in an afternoon.
Retrieval-augmented answering works roughly like this:
Two consequences follow directly:
Chunks must be self-contained. If your conclusion is in the intro and the supporting facts are 800 words later, the middle chunk contains an argument with no conclusion. The model extracts it and doesn't know what to do with it.
Key sentences must look like answers. Generation prefers sentences that can be dropped into a response verbatim. "Let's now explore the various aspects of this topic" contributes nothing.
This was my biggest mistake. We had a single "product overview" page covering what it is, pricing, competitor differences, and architecture. Four questions on one page means every chunk is half an answer.
Split it: /what-is-x, /x-pricing, /x-vs-y, /x-architecture — each title is the question. A month later, the pricing page started showing up in citations.
Every page should state the conclusion in its opening sentence, written to be quotable.
Before:
Before we discuss pricing, it's worth understanding the market context. In recent years, the SaaS landscape has become increasingly competitive...
After:
Bidvix's free tier covers 20 AI visibility checks per day across 5 models. Paid plans start at $29/month and add historical trends and citation clustering.
The answer exists by word 40. Everything after is context — useful for the model, but don't expect extraction from it.
Models extract facts from structured content far more reliably. Same information, different container:
| Format | Extraction difficulty |
|---|---|
| Three plans described in flowing paragraphs | High |
| Three-row comparison table | Low |
| FAQ, one Q per block | Lowest |
FAQs deserve special mention. The question is the user's prompt and the answer is the sentence the model wants to quote. That's not coincidence — the FAQ structure is isomorphic to the structure of a generated answer.
"We", "it", "this product" break at chunk boundaries. A chunk that reads "it supports 5 models" tells the model nothing about who "it" is.
Use the product name in every key block, and on first mention add a short descriptor: "Bidvix (an AI visibility monitoring tool)". This is entity consistency applied inside a page — same principle as making sure your one-liner matches everywhere else on the web. Don't rewrite the site. Fix the five pages that matter:
Verify by re-running the same prompts two to four weeks later and checking what share of cited URLs are pages you control. Look at the trend, not a single run.
"Extractable" is not "written for robots". These rules make pages better for humans too — direct answers, fewer pleasantries, cleaner tables. That's just good technical writing. Where extraction preferences and human reading preferences conflict, it's rarer than people assume, and you don't need separate content for the two audiences.
I built bidvix to handle the measurement half — mention rate and cited URLs across five models, so I can see whether the retrofit actually moved anything. The rewriting above needs no tooling at all.
Also in this series: The 20-minute AI visibility audit · The six kinds of pages LLMs actually cite · Your robots.txt might be hiding you