Short answer: Feed a chat widget the content that answers customer questions — services, pricing, process, policies, coverage area, FAQs and booking rules. Leave out marketing filler, blog archives written for search engines, legal boilerplate and anything outdated. More content is not better content; a small, accurate, specific knowledge base beats a large vague one every time.
Why volume is the wrong instinct
It is tempting to index everything, on the theory that more information means more coverage. Retrieval does not work that way.
Every answer is built from a handful of passages selected as most relevant to the question. If your index is full of near-duplicate marketing pages that all say "we deliver excellence with a personal touch," those pages compete for retrieval slots with the one page that actually states your call-out fee. Padding does not add coverage. It adds noise, and noise pushes the good passage out of the window.
The mental model to hold: you are not filling a library, you are curating a reference shelf.
A second consequence explains almost every disappointing widget. Answers are assembled from your material, so the widget can only be as good as what you have written. It will not reason its way to your call-out fee or invent your cancellation policy — the point of grounding, and also its cost. If the answer lives only in your head, the widget will decline the question. That is correct behaviour, and it is a to-do list. See retrieval-augmented generation for the mechanics.
The content that earns its place
Services, described concretely. Not "comprehensive solutions" — what you do, what it involves, what it excludes, how long it takes.
Pricing, in whatever form you can commit to. A real number is best. A range is fine. "Typical projects run between X and Y depending on Z" is far more useful than "contact us for a quote," because the widget can help the visitor decide. Price is the most-asked question in almost every business, and a widget that cannot address it will disappoint most visitors.
Process. What happens after someone gets in touch. First visit, timeline, what to prepare, who turns up. This is where anxiety lives, and it is usually documented nowhere.
Coverage and availability. Areas served, hours, lead times, weekends and emergencies.
Policies. Cancellations, deposits, guarantees, returns, rescheduling. Boring, high-stakes, and what people are most reluctant to phone about.
Your real FAQ — not the invented one. Whatever your team actually gets asked. If nobody has written it down, an hour with the person who answers the phone beats a month of blogging.
Qualifying information. What you do not do, who you are not right for. A widget that can say "we do not handle commercial properties" saves time and improves lead quality.
What a good source page looks like
Two versions of the same information, to make the difference concrete.
The version that produces bad answers:
Our Approach
At our practice we believe every client deserves a bespoke experience. Our highly trained team takes the time to understand your unique needs and delivers outstanding results. Pricing is competitive and tailored to each case. Get in touch to find out more.
Ask that page what an initial consultation costs and there is no correct answer to give, because the page does not contain one. The widget will decline, and it will be right to.
The version that produces good answers:
Initial consultation
Sixty minutes, £65. Includes assessment, a written treatment plan, and your first session of hands-on treatment.
Home visits
Available within five miles of the clinic, £20 surcharge. Not available on Saturdays.
Cancellations
Free up to twenty-four hours before. After that we charge half the appointment fee.
The second version is shorter, plainer, and does three jobs the first cannot. It answers the question. It carries the qualifier that would otherwise cause an argument later. And it is structured, so retrieval finds the right chunk instead of dragging in the whole page.
None of that is a chatbot technique. It is just good web copy, which is the quiet truth of this exercise: the content that makes a widget useful was making your website useful all along.
What to leave out
About-us prose on your journey and values. Nobody asks a chat widget about your journey.
SEO blog archives. Long, keyword-padded posts dilute your index and rarely contain a fact a customer needs. If a post does hold real specifics, extract those into a proper page.
Full legal documents. Terms and privacy policies are written to be defensible, not readable, and passages retrieved from them produce answers that are technically correct and useless. Summarise the customer-relevant parts and link to the full document.
Anything superseded. Old pricing, discontinued services, expired campaigns. Delete rather than archive — an unlinked page is still crawlable.
Testimonials as a bulk dump. A few are fine as social proof. Hundreds crowd the index.
Duplicated boilerplate. The same three paragraphs across twelve location pages give you twelve near-identical chunks competing for every question.
Where knowledge files fit
Some of your best content is not on your website and should not be. Internal FAQs, service specifications, policy documents, price sheets you send on request. Clerkzo takes ten file types at up to ten megabytes each, so this material goes straight into the knowledge base without being published.
Two rules make uploads work. Prefer clean text — a text-based document beats a scanned PDF, because a scan is an image and the text may not be extractable. And prefer one topic per file, because chunking cuts a mega-document in places that split the meaning.
Keep an inventory of what you upload. Files are the easiest thing to forget and the likeliest source of a confidently outdated answer. A useful test first: could a new member of staff answer a customer question correctly using only this document? If not, write the missing context in.
Corrections are content too
The highest-value content in your knowledge base is often the answer you wrote by hand after seeing a bad one. Flag the response, type the correct answer, and it becomes an authoritative override.
This is worth planning for rather than treating as cleanup. Some facts are awkward to publish — a nuanced pricing exception, a policy that needs framing — and a correction lets the widget give the right answer without a public page about it. It is the practical core of chatbot training for retrieval systems.
Keep them narrow and under review. One fact per correction. And when the same theme comes up three times, stop correcting and write the page. Keeping chatbot answers up to date covers that rhythm.
How to find your gaps
Write down the last twenty questions your team was asked and ask each in the playground, which mirrors the live widget. Anything answered thinly is a content gap, and a gap for your human readers too.
Then keep doing it with real conversations. Every conversation is saved with a thumbs up or down rating, and the down-votes plus the questions that end in a human handoff are your content backlog, written by your customers.
Fixing those gaps improves your website, not just your chat. It forces you to write down the things you have been explaining verbally for years. If you are still weighing tools, our guide to choosing an AI receptionist covers what to look for.
The content audit checklist
Run this before launch, and again each quarter. About half an hour once your site is in decent shape.
- Can a stranger find your price? A number, a range, or the structure of how you charge. If none of those exist, nothing else here matters as much.
- Are your hours stated in one place only? Duplicated hours drift apart. Pick a canonical page.
- Is your coverage area explicit? Named towns or postcodes beat "the local area".
- Are your services described in customer words? Search your pages for the phrase a customer would type. If it is not there, neither they nor retrieval will find it.
- Is your process written down? What happens after they enquire, in order, with rough timings.
- Are your policies published? Cancellations, deposits, guarantees. Plain language, one page.
- Have you written what you do not do? The most underrated page for lead quality.
- Is every page focused on one topic, with headings? Omnibus pages retrieve badly, and a fact with no heading above it rarely survives as a chunk.
- Have you deleted superseded pages, and do your files still match your live ones? Conflicts produce inconsistent answers.
- Is anything important trapped in an image or a scan? A price list as a JPEG is invisible to a chatbot and a search engine alike.
A cadence for adding content
Content work fails when it is a project and succeeds when it is a small recurring habit.
Before launch. Fix items one to seven above. Do not aim for a complete FAQ, aim for the seven facts every customer needs.
Week one. Read every conversation and note each question answered badly. That list is your content plan, and it beats anything you would have guessed.
Monthly, fifteen minutes. Write one page from the top of that backlog. Twelve pages a year of exactly what customers ask about beats a hundred posts nobody reads.
Quarterly, thirty minutes. Run the checklist. Audit files and corrections. Delete what has expired.
Whenever the business changes. Update the page that day, then recrawl. This is the habit that keeps everything else honest.
Common mistakes
Writing an FAQ from imagination before launch. You will guess the wrong twenty questions. Let real conversations tell you.
Uploading the whole website as files. The crawler already reads your site. Duplicating it doubles the noise and creates conflicts when one copy is updated and the other is not.
Treating the blog as knowledge. Marketing posts attract, they do not answer. A post arguing that you should service your boiler annually does not say what you charge to do it.
Leaving prices out on principle. If you cannot publish a number, publish the structure. Silence is not a negotiating position, it is a bounce.
Never deleting anything. An index that only grows gets vaguer. Deletion is maintenance.
Frequently asked questions
Do I need a big FAQ page before I start?
No. Start with what exists, go live, and let real questions tell you what to write. Guessing at an FAQ before launch usually produces the wrong twenty questions.
Should I upload my whole website as a file?
No. The crawler already reads it. Upload only what is not published.
What if I genuinely cannot publish prices?
Publish the structure — what drives cost, typical ranges, what a quote includes. That is enough for the widget to be genuinely helpful without committing you to a number.
Does adding more content always improve answers?
No. Relevant content improves answers; irrelevant content competes with it during retrieval and makes answers worse.
What about content in PDFs my customers already download?
Upload them, but check they are real text rather than scans, and that they do not contradict a live page. A brochure with last year's prices in it is a common source of a confidently wrong answer.