Here is a question we get in some form from almost every new customer: “Our documentation is good. Why does the assistant get this wrong?”
The documentation usually is good. The answer is wrong anyway, and the reason is almost never the sentence the assistant quoted. It is what was missing around it.
The assistant never sees your page
When a visitor asks a question, the assistant does not open your site and read. Long before that, every page was cut into passages of a few hundred words, and at question time a handful of those passages are picked because they look relevant. Those passages are all the AI gets.
A passage is not a page. On the page, the sentence “Set the limit to 50” sat under a heading that said which limit, inside a section that said which product, on a page whose URL said which version, next to a sidebar date that said when it was true. Cut the sentence out and every one of those facts is gone, unless something carried them along.
That is the whole problem. Your content is full of signals that give each sentence its meaning: the page title, the breadcrumb, the heading path, the version, the publication date, the author, the one-line description. Whether the AI answers correctly depends less on how well the sentence is written than on how many of those signals are still attached when the sentence arrives.
What it looks like when they are lost
Four cases from the past year, each one a good sentence that lost its context on the way.
The forecast reported as the result. A customer publishes years of reports and commentary. Asked what a key figure was for a given year, the assistant quoted a forecast published early that year as the final number. The sentence was accurate. The page said it was a forecast, in the title and the date. The passage did not. The fix was to carry the page’s date with the passage, tell the model today’s date, and use a model strong enough to act on both. Nothing about the writing had to change.
The contact who left years ago. Same customer. Asked who to contact about a topic, the assistant named someone in a senior role they no longer held, from a press release that was years old. Again the page carried the date. The passage arrived without it, and the model did not know what day it was, so it had no reason to doubt a years-old sentence.
Twenty versions of the same section. A software company publishes its documentation for every release it supports. Each version has the same section with the same heading and nearly the same words. Asked a question, the assistant picked the passage from whichever version happened to score highest, usually not the one the visitor was on. Once each passage carried its version and the assistant kept only one version per section, the share of questions that found the right passage went from three in four to nineteen in twenty. The pages were not touched.
The banner quoted as the answer. A documentation site showed a release announcement at the top of every page. On short pages that banner was most of the text, so it was cut into passages like everything else and competed with the real content for every question. The signal that reached the AI was a real signal. It was the wrong one.
In three of the four, the fix was on the assistant’s side: keep the date, keep the version, keep only one copy. In the fourth it was on the page’s side: keep the banner out of the article. None of them were fixed by rewriting a sentence.
Two halves of one job
This is why “AI-ready content” is a misleading phrase. It puts the whole job on the writer. The job has two halves, and the writer owns one.
The page’s half is to carry the signals at all. A date on anything that goes stale. A version on anything that varies by version. Headings that say what the section is about, so “Step 1” becomes “Step 1: Create the storage bucket”. A one-line description. Navigation and banners kept outside the article, so that what gets cut into passages is the article. This is not a style guide. It is making sure the facts that give a sentence its meaning exist somewhere a machine can find them.
The assistant’s half is to keep them. Every passage should travel with its page title, its place in the page, its version, its date, and whatever else the page offered. An assistant that drops those is the reason so much advice tells you to repeat the product name in every paragraph and never write “see above”. That advice is compensation. It is asking the writer to do by hand what the software should have done.
Most of the advice you will read about AI-ready documentation is about the first half, because that is the half the reader can act on. Most of the failures we see come from the second.
Our half
In ChatbotIQ, every passage carries its page’s title, its one-line description, its breadcrumb, the full path of headings above it, its version where the site publishes several, and the date the page was published or last changed, read from the page itself when the site does not declare one. The model sees those with every passage it is given, and it is told today’s date, so “is this person still in that role?” has something to compare against. Where documentation exists in many versions, the assistant keeps one per section. Where a site puts a banner inside every article, we strip it before the page is cut, once we have seen it.
This is why the three fixes above took no edits on the customer’s pages. The signals were already there. We had to stop losing them.
And the other half, if you want it
The page’s half still has to be done by someone who knows the content, and most teams do not know which of their signals are reaching the AI and which are not. There is no error when a date is missing. The answer just comes back confident and wrong.
On annual Standard and Business plans, the Bot Success Program is where we do that part with you. At onboarding we read your site the way the assistant does, show you which pages arrive with their meaning intact and which do not, and tell you what to change, which is usually short. Business customers get the same review every quarter, driven by the questions the assistant could not answer.
Good documentation is necessary. It was never sufficient. The sentence has to arrive with its meaning, and that is a job for both of us.