Setup
Indexing your whole site is the most common way to make an assistant worse.
The instinct on setup is to point the crawler at the domain and let it read everything. It is the single decision that most reliably degrades answer quality, and it is almost always made in the first ten minutes.
- Private launchOnboarding selected teams now
- Built by CreoglyphSeven client engagements behind the product
Why more sources make answers worse
Retrieval ranks passages by similarity to the question. Every additional source adds candidates. If those candidates are outdated blog posts, superseded policies or marketing pages that describe an aspirational version of the product, they compete with the correct passage and sometimes win.
When a stale passage wins, the model does exactly what it should: it reads the supplied text and answers from it, confidently, because from its position that text was the relevant context. The failure happened before the model was involved, and no amount of model quality fixes it.
This is why teams who upgrade their model to fix accuracy usually see no improvement. The problem was never phrasing. It was that retrieval returned the wrong passage from a source that should not have been indexed.
What changes
Today
- Point the crawler at the domain, index everything
- Assistant confidently quotes a 2023 blog post as current policy
- Team assumes the model is weak and upgrades it
- Nothing improves, because the wrong passage still wins
- Trust in the assistant drops and it gets removed
With an assistant
- Index the pages you would send a customer to, and nothing else
- Every source has an owner and a reason
- Retrieval competes between good candidates only
- Refusals rise, which is honest, and answer quality rises with them
- Gaps are logged so the source list grows deliberately
A source list you can defend
Start with the sitemap
It is a list you already maintain of pages you consider canonical. It is the best starting point and it is usually already too broad.
Cut to pages you would link a customer to
If you would not paste the URL into a reply, it should not be a source. This one rule removes most of the problem.
Remove anything dated
Blog archives, changelogs older than the current version, superseded policies. If it was true and no longer is, it is worse than absent.
Add the files that are not on the site
Handbooks, specification sheets, price lists. Often the highest value sources and usually not crawlable.
Give each source an owner
Not a tooling feature, a working practice. A source with nobody responsible for it is a stale passage waiting to happen.
Reindex on publish, not on a schedule
Scheduled reindexing is simple and always slightly wrong. Event triggered is correct and requires your publishing system to emit an event.
A pruning pass you can run in an afternoon
Export the list of URLs you are indexing. For each one, answer three questions in writing.
Would I send a customer this link today
If no, remove it. This alone typically removes a third of an over broad index, and it removes the third most likely to produce a confidently wrong answer.
Is anything here contradicted by another page
If yes, fix the contradiction before indexing either. An assistant with two conflicting sources will pick one and will not tell you it had a choice.
Who notices if this goes out of date
If the answer is nobody, either assign someone or remove it. Ownership is the only durable defence against staleness.
Run the same pass quarterly. An index is not a setup step, it is a maintained artefact, and the failure mode is always the same: it was right on the day it was built.
Where the data goes
The assistant answers only from sources you nominate. Content behind a login, inside an image, or rendered only by client side script is not reachable by the crawler and will silently be absent rather than fail loudly. Security and data boundaries.
Questions
Signal
Is this your situation?
Access is by waitlist, demo request or public-content pilot. Describe the question your visitors keep asking and we will tell you honestly whether it fits.