Every answer links back to the page that actually said it.
A support bot that will not show its work is asking for blind trust, and a visitor has no reason to give it any. Helpforge's agent does the opposite: whenever an answer draws on your website or the files you gave us, it links the exact page underneath, so a customer, or you, can check it in one click instead of taking the bot's word for it.
Sources are resolved from what the model actually retrieved, never typed in after the fact, so a link can only ever point at a page the answer really drew on.
What's your return window on sale items?
Sale items can be returned within 14 days of delivery for store credit, rather than the 30-day window on full-price orders.
An answer with no source is a promise you cannot check.
"Trust me" is not good enough when the thing talking is a language model, in your name, to your customers.
A generic chatbot states a policy and moves on. If it is right, you got lucky. If it is wrong, neither you nor the visitor has any way to notice before it becomes a complaint or a refund you did not plan for.
A confident sentence and a correct one read identically from the outside.Without a link back to the page, checking a transcript means hunting through your site by hand to guess where an answer might have come from, assuming it came from anywhere at all.
If you cannot audit it, you are trusting it, not verifying it.An answer with nothing under it feels like it came from the model's general knowledge, even when it happens to be correct. A citation is what tells a skeptical visitor "this came from the business's own page," not from a guess.
Being right and looking trustworthy are not the same thing without proof.The citation is the retrieval, not a guess bolted on afterward.
The mechanism, not a promise: this is what the code actually does between a question and the "Sources:" row under the reply.
Every page and file we index gets a number when it's handed to the model as context, and the model cites by that number. A citation resolves back to one of the specific passages the agent was actually given for that question, so it cannot point at a page the answer never touched.
The link is a record of what was read, not a label added after the fact.A long page is often split into several passages, and an answer can draw on more than one of them. Citations are resolved back to the page they came from and deduplicated by that page's address, so a visitor sees one clean link, not the same page listed three times.
One page, one link, how many times the model actually read it."I'm not able to offer a discount" is not a claim about your business, so it gets no source, even if the model tried to attach one. A separate, deterministic check reads every sentence, tells the assistant talking about itself apart from the business stating a policy, and strips a citation that has nothing to do with what it is glued to before the reply ever ships. Read the full detail on our hallucinations page.
A link that supports nothing is worse than no link at all.The "Sources:" row shows at most three links, in the order the answer actually used them. A wall of ten citations under a two-sentence answer reads as noise, not proof, so the row stays to the handful that matter.
Enough to check the claim, not so many nobody clicks any of them.Two content sources. One of them makes a link, the other doesn't, and that's on purpose.
Your crawled pages and any files you hand us both feed the same agent, but only one of them has a public address for a citation to point to.
Pages and uploaded files
A crawled page and a document you upload (a PDF policy, a CSV price list) are ranked the same way, on how well they answer the question, and either can end up in the "Sources:" row with a real link a visitor can open.
A hand-written answer
When you or we teach the bot a specific answer directly, it grounds the reply exactly as well as a page does, and it never gets marked as an unanswered gap, but it has no public URL for a citation to link. It shows up in your own report as an answer that grounded the reply, just not as a clickable source to the visitor, because there is no page to send them to.
A source link is untrusted content until it's proven safe to render.
A page's own title and address reach the widget as data, not markup, the same rule this product applies to a visitor's own words.
Only a link starting with http or https is ever rendered. A javascript: or data: address, however it got into a source's url field, is simply dropped rather than turned into a clickable link.
A source's title comes from your own crawled page, so it is treated as untrusted the same way visitor input is. It is written into the link as plain text, never parsed as HTML, so a page titled with something that looks like a tag cannot inject anything into your widget.
Your own website's content gets the same scrutiny as a stranger's.Every source link opens in a new tab with noopener noreferrer. The page it links to can't reach back into the tab it opened from, standard hygiene for a link generated from content we do not control.
One flat build fee, one flat monthly.
Citations are part of the agent we build for you, not a higher tier. Every plan gets the same "Sources:" row, on every grounded answer, from day one.
Free demo on your real site first. Ask it something your site actually answers, and follow the link yourself.
Source citations, answered honestly.
Does the AI chatbot show where its answers come from?
Yes. Whenever a reply draws on your website or a file you uploaded, a "Sources:" row appears underneath it linking the exact page or document used, so a visitor, or you, can check the answer against the original rather than taking the bot's word for it.
Can the bot cite a page it didn't actually use?
No. Citations resolve back to the numbered passages the model was actually given as context for that question, so a link can only point at something the answer really drew on. A separate check also strips a citation off any sentence that is the assistant talking about itself, a decline or an offer of help, rather than a claim about your business, so a "sorry, I can't do that" never wears a source it doesn't earn.
What if the same page gets cited more than once?
It shows up once. A long page is often split into several passages, and citations are deduplicated by the page's own address before they reach the visitor, so a reply that drew on three passages of one page still shows one link, not three copies of it.
Do hand-written answers get a citation too?
They ground the reply exactly as well as a page does, and never get counted as an unanswered gap, but they have no public web address to link to, so no clickable source appears for that part of the answer. It still shows up correctly in your own report as a grounded reply.
How many sources does one answer show?
Up to three, in the order the reply actually used them. It is a deliberate cap: enough to let someone verify the claim without turning the row into a wall of links nobody reads.
Is a source link safe to click?
The link itself only ever uses http or https, since any other scheme is dropped before rendering, and the page title is written in as plain text rather than parsed as markup, so a crawled page cannot use its own title to inject anything into the widget. It also opens in a new tab without handing that tab a reference back to yours.
See the citations on your real site.
Give us your website URL and we build a working demo of your agent, usually within 2 business days. You get a private link to test it, ask a real question, and follow the source link yourself before any commitment. No call, no credit card.