How Answers Work
How the agent finds a passage, the sources it shows, and what it does when it has no answer.
Your agent is instructed to answer only from the content you connected — your crawled pages, your uploaded files, and the notes you typed. This page explains what happens behind the reply: how NovixAgent AI finds the right passage for a question, what it shows the visitor as proof, and what it does when your content genuinely does not cover what was asked.
If you are still filling the knowledge base, start with Adding Content. This page is about what happens to that content once a visitor asks something.
Finding the right passage
For a visitor question, NovixAgent AI runs two searches at once over your knowledge base:
- Meaning search looks for passages that mean the same thing as the question, even when they share no words with it. A visitor asking "can I send it back if it doesn't fit" reaches a page headed "Returns and exchanges".
- Keyword search looks for the words themselves. This is what keeps product names, model numbers, SKUs, and policy titles findable — the things a meaning search can blur together.
The two result lists are then merged by how highly each search ranked a passage, not by the raw scores the two produce. That detail matters in practice: the two kinds of search score on scales that are not comparable, so merging by position is what stops one of them from drowning out the other. A passage both searches liked rises to the top; a passage only one of them found can still get through. If the meaning search is briefly unavailable, the keyword search runs on its own.
The practical consequence for you is that you do not have to guess visitor phrasing: visitors can word a question their own way. What you do need is for the fact itself to be in your crawled pages, uploaded files, or typed notes — search can match a differently worded question, but it can only find what your content says.
Crawled pages, typed notes, and uploaded files are searched together, and you do not have to decide in advance which source should win.
Putting the best evidence first
Search returns a pool of candidate passages. A second stage then reorders that pool so the strongest evidence is what the answer gets written from.
The first ranking pass is tuned to your business type — what counts as a strong match for a restaurant menu is not what counts for a legal services page. On top of that, higher plans add a premium reranking stage that reads the question and each candidate passage together and re-sorts them:
| Starter | Pro | Business | |
|---|---|---|---|
| Premium reranking | — | When the leading candidates score closely | On every question |
| Reranked searches per month | — | 4,000 | 20,000 |
Where the table says when the leading candidates score closely, the premium stage is used selectively — only when the order is genuinely in doubt. When one passage is a clear winner, the extra stage would only confirm what the first pass already decided, so it is skipped.
The monthly figure is a ceiling on the premium stage, not on answers. Once a month's reranked searches are used, answers keep coming with the first-pass ranking, and the count starts again when your monthly allowance renews.
The table above describes a deployment where the premium reranking service is connected. Where it is not, the plan distinction does not apply: in place of the premium stage, a second pass asks a language model to sort the leading candidates. It is a different mechanism, not a downgrade of the paid one, and it is bounded by the same answer-time budget: if that budget runs out, or the call fails, the first-pass ranking is used as-is. Nothing about this changes what the agent reads or how confidence is judged. If you are comparing plans, compare them on the same deployment.
What upgrading does not change is how much your agent reads. Every plan retrieves from the same candidate pool and hands the answer generator the same evidence budget. The paid stage changes the order of that evidence, not its quantity, and it is deliberately barred from changing whether an answer counts as well-supported — so the confidence you see in analytics means the same thing whichever plan you are on.
If the premium stage is briefly unavailable, the answer is still produced from the first-pass ranking. NovixAgent AI treats such an answer as less certain rather than presenting it as fully confirmed.
The Sources control
When an answer is supported by retrieved evidence, the widget shows a small Sources (N) control under the reply, once the reply has finished appearing.
Here is what a visitor actually experiences:
- The control is collapsed by default and shows only the count, so a reply stays short. The answer itself is shown in full — the sources are an extra, not a truncation.
- Opening it reveals up to three source chips. Three is a deliberate cap: a longer list stops being evidence a person will check.
- A chip for a crawled page is a link and opens that page in a new tab.
- A chip for an uploaded file shows the file name and is not clickable. Your internal manuals and price lists stay internal — the visitor sees which document backed the answer without being handed the document.
- A chip for the note you typed under "Anything we missed" shows your website's address and is not clickable. A site with no address on file shows a numbered label there instead.
- Hovering a chip shows a short preview of the passage that was used.
Two honest limits worth knowing before you quote this to a customer. First, the list is what the agent read to write the answer, not a per-sentence footnote — there is no claim-by-claim mapping from one line of the reply to one source. Second, a reply can come without one: the control is hidden when a turn is recorded as having no usable evidence, when the visitor asked for a person, and when the agent has no reply to give because of a technical fault. It may also be hidden when the agent itself says it is unable to answer, but that depends on the AI model in use and is not guaranteed. As the next section explains, some of the turns recorded as having no usable evidence are ones the agent did answer.
Visual result cards do not affect this either way. Retrieval runs before cards are chosen, and cards are shown underneath a normal text reply rather than instead of it, so an answer with cards carries its sources as a text-only answer does.
Source chips are not a paid add-on, and no setting hides them.
When the agent has no answer
Behind an answer from your content, NovixAgent AI records one of three internal results: the evidence clearly supported the answer, it supported it weakly, or there was no usable evidence. This judgement is made from the retrieved material — how strong the best match was, how much of it there was, and how many different pages agreed — and not from asking the model how sure it feels.
This is not a badge the visitor sees. There is no confidence score, no colour, and no label in the chat. What the visitor notices is the behaviour that follows from it:
- Clear evidence. The agent answers and shows its sources, unless the visitor asked for a person.
- Weak evidence. The agent answers and shows its sources, unless the visitor asked for a person. The answer is recorded as less certain so you can find these turns later.
- No usable evidence. The sources are hidden. What happens to the answer depends on something the label does not tell you — see below.
"No usable evidence" is one label over two different replies
This is the part worth reading slowly, because the label and the behaviour do not line up the way the name suggests.
A turn is recorded as no usable evidence in two situations: when the search came back genuinely empty, and when it came back with passages whose best match scored below the bar. Those two produce different replies, because the decision to hold back is made on whether passages reached the agent — not on the recorded verdict.
- Nothing was retrieved. The agent has nothing in front of it and is instructed not to answer from general knowledge but to say, in the visitor's language, that it cannot help with this one. This is the refusal.
- Passages were retrieved but scored below the bar. The agent is handed those passages and answers from them normally. The verdict is still recorded as no usable evidence — so the turn is filed under Questions Without an Answer in your analytics — but the visitor was not refused.
In both cases the Sources control is hidden. On the second one that produces a reply you should know about: an ordinary-looking answer with no sources beneath it. The evidence it was written from exists and was used; it is simply withheld from display, because the same verdict that suppresses the list is the one the ranking was not confident in.
In both cases, where you have a way to reach a human, the reply also offers it — a Talk to a human option when an operator is online, or an Email the team option where you have a lead form. On the second case that button sits under an answer rather than under a refusal, which can read as the agent hedging. The same option also appears under a reply when the visitor has asked for a person, whatever the evidence was, and the Sources control is hidden under that reply too.
Why this matters to you. A sourceless answer is the one shape a visitor has no way to check, so it is worth knowing that it exists and is not a bug. It is also why the Questions Without an Answer list is not the same thing as a list of refusals: entries there may have been answered. Treat that list as "turns where the evidence was too weak to stand behind", and read the reply before concluding the visitor was turned away.
This is the trade the product makes deliberately: a confident wrong answer about a price, a policy, or a delivery date is far more expensive than an honest "I could not find that". A wrong answer is repeated to a colleague, screenshotted, and occasionally enforced against you. A refusal, with a way to reach your team, costs one click. The refusal path is not a limitation of your plan, and it is the reason the rest of the answers can be trusted.
Visual result cards
For supported business types, a browsing question can produce visual cards below the assistant's short reply instead of a long text list. The supported card types are:
- E-commerce — products.
- Real estate — property listings.
- Restaurants and food businesses — menu items.
- News and media — articles.
- Educational services — courses.
- Hospitality — rooms.
- Business, professional, and home services — services.
Cards are selected automatically from the business type detected during crawling. Other business types continue to receive ordinary knowledge-base answers; selecting a template or writing a prompt does not turn cards on for an unsupported type.
During the crawl, NovixAgent AI builds cards only from information it can extract from the page: structured data such as JSON-LD, Open Graph metadata where that format identifies the item, and recognizable item-page content. A card carries the link to the page it was built from and shows only the fields that page offered, such as a title, image, description, or price. It does not invent missing prices, availability, images, or attributes. Cards appear only when the visitor is browsing concrete items and at least one relevant card was found; support and policy questions still receive a normal text answer.
Seeing this in your analytics
You do not have to guess how often your agent is finding good evidence.
Dashboard → Analytics → Messages (its plans are in Which tabs are on which plan) lists Questions Without an Answer — the turns recorded as having no usable evidence. As the section above explains, that is not the same as "the agent refused": some of those turns were answered from passages that scored below the bar. Either way each entry is worth the same action — a page to write, a document to upload, or a line to type into your knowledge base. Fix it once and the next visitor who asks gets a supported, cited answer.
On a plan without that tab, it is shown locked and the per-question breakdown arrives with an upgrade. The plan decides how much you can inspect afterwards, not how answers are produced.
Conversations is a separate tab, with its plans in the same table. It shows the account owner a session turn by turn. Two things about it are worth knowing before you go looking:
- Confidence is not printed next to each reply on screen. The tab shows who spoke, when, and the text. If you want each reply's recorded status and confidence figure, use the transcript's CSV download — those are columns in the exported file.
- Turns the agent answered confidently usually have no text to show. By default your site stores message text only for turns the assistant did not answer confidently, so a well-answered turn appears as a short label ("Answered confidently") rather than a blank. That is your retention setting working, not a fault — you can change it under Chat Configuration → Transcript storage if you want full transcripts kept.
The practical consequence is a useful one: the stored text is the turns most worth reviewing. See Website Analytics for the full loop, and Adding Content for how to close a gap once you have found one.
Seeing an answer that puzzles you — a source you did not expect, or a refusal on content you know is indexed? Email support@novixagent.ai with the website and roughly when it happened, and we will trace that specific turn.