A chatbot on a business website can only answer what the website already says. It has no access to your Slack history, the policy locked in a manager's head, or the page you updated last March and forgot to republish. Before you install anything, the work that matters most is the content audit that tells you what your site does not cover.
Why the chatbot is only as good as your content
When someone asks your chatbot a question, it searches across whatever documents it has been given. If the answer lives in a well-structured page, the chatbot finds it and returns something accurate. If the answer is buried three menus deep on a page titled something unrelated, it may find it and it may not. If the answer exists only in a support ticket, or in the memory of a colleague who left last year, the chatbot cannot produce it because it does not exist in any document it can read.
This means the quality of your chatbot is a content problem before it is a technology problem. Buying the most capable AI on the market and connecting it to a site with gaps, contradictions and outdated pages does not fix those problems. It makes them faster.
The audit described on this page does not require any software. It requires access to your email inbox, one hundred tickets, and roughly two hours. It produces a list of every question your website cannot answer today.
Think of it this way: the chatbot is a reader with perfect recall and no context outside what you give it. It does not know your business. It cannot infer what you meant to write. It can only work with the words that are actually there. If those words are missing, incomplete, or wrong, the chatbot's answer will be missing, incomplete, or wrong. The audit is the process of finding out which questions have no good answer in the materials you have handed the chatbot so far.
The one hundred emails method
Go to your support inbox and open the last one hundred email conversations that came in through your website contact form, live chat, or a linked help desk. Do not include internal forwarded messages or repeat contacts about the same issue. One hundred distinct conversations.
Read each one and mark it in one of two categories:
- The answer to this question exists somewhere on your website, in a form a customer could find.
- The answer to this question does not exist on your website, or exists but is not findable by a customer.
Work through all one hundred before you stop. The pattern that emerges in the second category is the map of your content gaps.
What you are building here is a gap inventory: a list of questions that sent a customer to email or chat because the website let them down. The chatbot cannot answer these questions unless the underlying content first exists in a retrievable form.
A useful refinement as you work through the emails: note not just whether the answer exists, but how long it took you to find it. If you had to search for more than a minute, or had to ask a colleague, that question still belongs in the second category. The test is not whether the answer exists somewhere, but whether a customer arriving on the site with no internal knowledge could locate it in a reasonable time.
The three shapes of a content gap
When you read the one hundred unanswered questions, they will fall into three types. Each requires different treatment, and knowing which is which changes what you do next.
Shape one: the page exists but buries the answer
The most common gap is not a missing page. It is a page that contains the answer but does not surface it in a way a customer can use. The return policy exists. It is linked from the footer in eight-point type. The answer to "how long does delivery take" is in the third paragraph of the shipping page, below a block of promotional text and above a table of express options that are not available in your region.
A customer arriving at the site with this question cannot find the answer quickly, so they email you. The chatbot trained on this page will find the answer because it searches the whole document, not just what the customer can see. But only if the page exists in the first place.
Fix the page before you automate it. Move the answer higher. Use the customer's language in the heading. Strip out the filler that sits between the question and the answer. Then connect the chatbot to the revised page.
Common variations of this shape include the FAQ buried at the bottom of a product page when customers ask the same question in every email, the pricing page that states the base fee but omits the setup charge that every customer later asks about, and the terms page that answers the cancellation question in legal language that no customer actually reads. Each of these has an answer on the site. None of them surfaces the answer where the customer looks for it.
Shape two: the answer lives in one person's head
The second gap is the answer that exists nowhere in writing. A colleague knows it. A manager makes the decision on the spot. The engineer who built the product can answer the technical question, but no document can be found.
This gap is harder to fix because it requires first extracting the knowledge and then writing it down. It also requires ongoing maintenance, because when that person leaves or changes roles, the document is the only place the answer survives.
The one hundred emails will show you which topics fall into this category: look for questions that a specific person tends to handle, or questions where the customer's phrasing suggests they are describing a situation the site does not address at all.
Document the answer. Turn it into a page or a section within an existing page. Then the chatbot has something to read.
This shape is particularly common in businesses that have grown quickly. The founders handled everything personally, built up expertise through thousands of individual conversations, and never sat down to write the answers down because they always just answered them directly. The gap inventory will surface these topics as questions with no clear source page and no obvious place to send the customer.
Shape three: the answer changed and was never republished
The third gap is the page that once told the truth but no longer does. The price list was updated. The product was discontinued. The team moved to a new phone number. The page still exists, still answers the question, but the answer is wrong.
This is the most damaging gap because it produces confident wrong answers. The chatbot will read the outdated page and return a response that sounds authoritative and is completely incorrect. A customer told their refund takes three days when it now takes ten will be unhappy before they even speak to you.
Your audit will flag this gap if a question appears repeatedly and your team finds themselves saying "the website says X but it is actually Y." Every one of those responses is an outdated page that needs correcting before it goes near a chatbot.
The risk here is specific to chatbots because they do not hedge. A human support agent who knows the page is wrong will naturally correct themselves or qualify the answer. A chatbot that has read only the outdated page has no way to know the information has changed, and will return the wrong answer with the same confidence it would return a correct one. This is why the content refresh must happen before the chatbot goes live, not after.
What the audit produces
After you have marked all one hundred emails, you will have a count in each category and a list of specific questions that belong in the second category. That list is your content roadmap, and it is the document that decides what the chatbot can and cannot handle on launch day.
Count how many of the one hundred fell into the second category. If fewer than twenty, your content is in reasonable shape and the gaps are mostly fixable. If fifty or more of your last one hundred contacts were questions the site could not answer, the chatbot will inherit that gap and customers will notice within the first week.
Categorise each gap by shape. If the majority are buried answers, your priority is page restructuring. If they are tribal knowledge, your priority is documentation sprints. If they are outdated pages, your priority is a content refresh before anything else.
You can run the chatbot on the content you have now, but the more gaps the audit finds, the more the chatbot will deflect to a human agent, which means your volume savings will be lower than expected and your team will still be handling the questions the site cannot answer.
The output of the audit is also a baseline for measuring progress. After you have written the missing pages and updated the outdated ones, run the same check again on the next hundred tickets. The proportion should shift. If it does not, the content you added may not be finding its way to customers, which is itself a navigation or linking problem that the chatbot cannot solve alone.
Connecting the audit findings to your chatbot volume
Once the content is in place, the next question is how many conversations your chatbot will handle and which plan that maps to.
A chatbot conversation consists of multiple messages: the opening statement, the customer's question, any clarification the chatbot asks for, the answer, and a closing exchange. A typical business-to-business chatbot session runs between four and eight messages. A consumer-facing site with shorter queries might average three to five.
Take the number of chats you currently handle per month. Multiply by the messages-per-chat that matches your use case. The result is your monthly message volume, and that is the figure that maps to an Asyntai plan.
These calculations use your own volume figures. Run them with the number of chats your team handles today, not an industry average, because your traffic pattern is the one the chatbot will serve.
If your audit found a high proportion of gaps, the chatbot will not handle every incoming conversation. Some will deflect to a human agent because the answer does not exist in the knowledge base. That reduces the effective volume the chatbot covers, but it also means those conversations were never going to be automatable until the content gap was closed.
What changes the answer
Several factors shift which plan is appropriate and how much of your volume the chatbot can absorb.
Question complexity. Simple questions with short, factual answers need fewer messages per conversation than questions requiring clarification, follow-up, or multi-step guidance. A chatbot answering "what are your opening hours" needs one exchange. A chatbot helping someone complete a multi-field form needs several.
Session length settings. Some platforms count every message in a rolling thirty-day window. Others count by conversation. A customer who contacts you three times in a week and asks one question each time will generate three separate conversations on some platforms and a single ongoing session on others. This affects how your message count relates to your actual chat volume.
Out-of-hours coverage. If your chatbot operates outside your team's working hours, it handles a larger share of your total contacts, because there is no human to pick up the phone. Sites that serve international customers across time zones tend to see higher chatbot-to-human ratios in off-peak windows.
Seasonal spikes. A business with seasonal peaks may find that chatbot volume in the quiet months fits comfortably within a lower tier, while peak months exceed it. The audit gives you the baseline; you decide whether to buy for the peak or manage overflow differently during the high season.
Industry and audience. A technical product with complex configuration questions will generate longer conversations than a retail site with product queries. A business-to-business site where customers arrive having already read the pricing page will have shorter, more specific questions than a consumer site where customers arrive with no prior research. These differences affect the messages-per-chat assumption, and getting it wrong in either direction has consequences: too high and you are paying for unused capacity, too low and you are hitting the allowance ceiling.
The mistake most people make
They install the chatbot, connect it to their existing content, and expect it to handle the full volume of contacts. When it deflects more than expected, they blame the AI.
The chatbot is doing exactly what it is supposed to do: reading the documents it has been given and answering questions that appear in those documents. If the documents do not contain the answer, it cannot produce one. This is not a model failure. It is a content gap, and the chatbot has revealed it in a way that the one hundred emails also reveal it.
The installation cost is small. The content work is where the time goes, and it is the work that determines whether the chatbot reduces your ticket volume or simply redirects it.
Do the audit. Fix the three shapes of gap. Then connect the chatbot to a knowledge base it can actually read. That is the sequence that produces a working assistant instead of an expensive deflector.
Asyntai installs via a script tag or Code Injection on Squarespace. It connects to your published website content and answers questions within a configurable scope. The free plan carries 100 messages a month, and the paid self-serve plans start at 2,500.
Sources and dates for the figures in this article:
Asyntai message allowances and prices: Asyntai pricing page, asyntai.com/pricing. Arithmetic on message counts is worked out in the article body using the figures published there. The 5 messages per chat session is a working assumption stated as such in the text, not a measured figure from a third-party source.