How Many Chatbot Conversations Will Your Website Get? A Sizing Formula

Every AI chatbot plan is sold in a unit: replies, message credits, conversations or resolutions. You know your website traffic, but not how many visitors will open the chat, type, or keep talking, so the plan size starts as a guess.

The guide is for owners, founders and support leads choosing a plan before launch, or checking whether a free tier will last the month. It gives the sizing formula, the published benchmark for each variable with its source, year and sample, worked examples from 2,000 to 200,000 visits a month, what pushes the numbers up or down, and how to measure your own figures in the first 30 days. Every source was opened on 19 October 2026. Several are old: the best public studies of how many visitors start a chat date from 2014 and 2015, and the tables say so.

PepoChat, which we build, counts AI replies, so the plan examples use its prices. PepoChat publishes no usage statistics and none appear here; every rate below comes from a third party. For adoption and resolution figures beyond volume, the AI customer service statistics roundup lists the sourced numbers.

How do you estimate chatbot conversations from website traffic?

Multiply your monthly visits by the share of visits that become a conversation, then by the AI replies each conversation uses. Written out with every step, the formula has four inputs:

Monthly AI replies = monthly visits × chat-open rate × message rate × AI replies per conversation

Monthly visits are the sessions on pages that load the chat widget, taken from your web analytics. Use visits rather than unique visitors, because the published benchmarks divide chats by visits and a returning visitor can chat twice. If the widget runs only on some pages, count only the sessions that include those pages.

The chat-open rate is the share of visits in which someone opens the widget. The message rate is the share of opened widgets in which the visitor sends at least one message. The two multiplied together give the conversation rate, the share of visits that turn into a conversation. Public studies only measure that combined rate, so in practice the formula has three inputs:

Monthly AI replies = monthly visits × conversation rate × AI replies per conversation

AI replies per conversation is the number of answers the AI writes in an average conversation. A visitor who asks one question and leaves uses one reply; a visitor who asks, follows up and says thanks can use three. Plugging in typical values: 10,000 visits × 1.8% = 180 conversations, and 180 × 2 replies = 360 AI replies a month.

What are realistic values for each variable?

Start with about 2 conversations per 100 visits and 2 AI replies per conversation, then replace both with your own numbers after a month. The table gives the published low, typical and high values and where each comes from.

Sources opened 19 October 2026. "Typical" is a published median or our division of published totals; the low and high columns are published extremes, not hard limits.

Read the table with three cautions. The denominators differ: BoldChat and LiveChat divide chats by visits, while Tidio divides by widget impressions. Most of the conversation-rate data comes from live chat answered by people, where the widget showed "offline" part of the day. And every figure is a vendor reporting on its own customers.

What percentage of website visitors start a chat?

About one or two in every hundred visits, in the two public datasets that divide chats by visits, falling below one in a hundred on high-traffic sites. The spread between individual sites is wide, so treat any single number as a starting point rather than a forecast.

The most detailed study is BoldChat's Live Chat Performance Benchmarks 2015, the third edition of a report first run in 2009. It drew on thousands of customers, tens of millions of chat records and hundreds of millions of website visit records. The median site chatted with 1.6% of its visitors (1.7% on desktop, 1.4% on mobile), essentially unchanged since 2009. BoldChat gives a 95% interval of under 1% to 10%, and its "elite engagement" group chatted with 7% to 38% of visitors. BoldChat counted chats an agent actually answered, so attempts made while nobody was online are not in the 1.6%; an AI agent that answers at any hour catches those as well.

BoldChat also split the median by traffic, which is the table to use for sizing:

Source: BoldChat Live Chat Performance Benchmarks 2015, figure 1, read on 19 October 2026.

LiveChat's Customer Service Report, with data last updated in 2024, says it examined more than 87 billion website visits and 2 billion chats. Dividing one by the other gives about 2.3 chats per 100 visits; that ratio is our arithmetic, not a figure LiveChat publishes, and using the 1.68 billion chats in the report's own table gives 1.9.

Tidio's live chat statistics page, updated on 25 February 2026, says about 15% of visitors join live chat conversations, calculated by comparing widget impressions with chats started on almost 300,000 websites. Tidio does not give the period, or say what counts as a chat, so treat 15% as the top of the plausible range rather than a typical value.

How many messages does a chatbot conversation have?

AI conversations are short. Crisp's 2026 data puts the median at 2.2 to 4.2 messages per AI conversation, depending on the industry. Human live chats ran to about 10 messages in Zendesk's 2015 benchmark, counting both the agent and the visitor.

Crisp's State of AI Customer Support 2026 analysed 484.8 million conversations from 27,318 workspaces on its Plus and Essentials plans between May 2025 and August 2026, reporting workspace-level medians and no figure from fewer than 30 teams. Telecommunications had the longest AI conversations, at 4.15 median messages, then education at 3.05; entertainment was among the shortest at 2.18. Crisp does not say whether the customer's own messages are in the count, and the data spans all of Crisp's channels, not only website chat. If both sides are counted, a typical AI conversation holds one or two AI replies; if only the AI's messages are, two to four.

Zendesk's benchmark report on live chat, from the first quarter of 2015, used 2,261 Zendesk customers on Zopim Live Chat with at least 25 chats in the quarter. Its benchmark was 10 messages exchanged per chat between agent and visitor, with software companies the longest at 12.0 and healthcare the shortest at 8.3. Those counts include the agent's side; if both sides wrote about equally, about five messages came from the visitor, which is why the table's high value is 5.

On PepoChat, an AI reply is one answer the agent generates in a visitor conversation. Your team's messages, the greeting and system messages do not count, and chats in the dashboard's Test the agent drawer are free. The agent answers each visitor message once, so a closing "thanks" uses a reply too. Once a conversation is escalated, the agent stops and the replies stop counting; the billing documentation has the exact definition.

How many AI replies does a website need? Four worked examples

At typical rates, a 2,000-visit site needs about 80 AI replies a month and a 200,000-visit site about 2,800. A site that pushes chat hard can need four to thirteen times more. The table shows the arithmetic for both cases.

Typical rates are BoldChat's 2015 medians for each traffic band. The busy case uses 6%, about the share of invited visitors who accepted a proactive invitation in BoldChat's data, as a stand-in for a site that invites every visitor or makes chat the main way to ask a question.

Plan on the typical column and check your plan against the busy one. At 200,000 visits the typical case needs 2,800 replies and the busy case 36,000, so the busy column, not the typical one, tells you whether usage-based pricing is safe. If your estimate lands near 500 conversations a month, the support stack for a bootstrapped SaaS covers the tools around the chatbot at that volume.

Notice that 200,000 visits produce only about eight times the conversations of 10,000, not twenty times. BoldChat's data shows larger sites converting a smaller share of visits into chats, so scaling a small site's rate up to a large one overstates volume.

Do ecommerce, SaaS and service businesses differ?

Less than you might expect, on the published evidence. In Crisp's 2026 data, AI conversations in retail and ecommerce (2.64 median messages), technology (2.65) and professional services (2.63) were almost the same length; real estate (2.93) and education (3.05) ran longer. Human chats showed more spread: software was the longest industry in Zendesk's 2015 benchmark at 12.0 messages, and in LiveChat's 2024 data software and services chats averaged 9 minutes 24 seconds against 8 minutes 25 seconds overall.

No public study we found splits the conversation rate by industry. Start all three business types at 2 replies per conversation, budget 3 if your chats are mostly troubleshooting or setup questions, and measure the conversation rate yourself, because that is where your site will differ most.

What makes chatbot conversation volume go up or down?

Invitations to chat, the first screen of the widget, forms in front of the first message, and the calendar move the numbers most. Here is what the data says about each, and what you control in PepoChat.

Proactive greetings and invitations

Inviting visitors to chat is the largest single lever in the published data. BoldChat's 2015 report estimated that adding proactive invitations increased a site's chat engagement by about 313% on average, about 421% on mobile and about 714% on high-traffic sites. LiveChat's Chat Greetings Efficiency Report, from April 2014, followed 810 businesses for six months: 32 million greetings produced 2.4 million chats, a 7.28% average and an 8.07% median, so about one greeting in twelve became a chat. Greetings brought smaller businesses almost half their chats, and larger ones about a quarter.

Targeting beat volume in the same report. Greetings triggered by time spent on one particular page converted 10.21% on average, against 6.89% for the most common rule, time on the whole site. The lesson for any tool: invite on the pages where people hesitate, such as pricing, shipping and checkout, rather than everywhere.

The PepoChat widget does not open by itself; the greeting appears when a visitor opens the chat. To invite visitors on a particular page, add a link or button that calls EchoWidget.show(), which opens the chat window, as the widget API documentation shows. A button is a gentler invitation than a pop-up, so no published figure applies to it directly; recheck the estimate after you add one.

Suggested questions and the first screen

PepoChat shows up to three suggested questions as tappable chips, and each one starts a conversation with its text. When a workspace has no published Help Center articles, Home is chat-first: the greeting, the chips and a message box. With articles published, Home shows a Send us a message card above a help search and your top four articles, so some visitors read an answer and never chat, which lowers both conversations and replies. No public study measures the effect of suggestion chips, so change one thing at a time and compare a week before with a week after, as the welcome message guide describes.

Forms before the first message

Asking for details before the chat starts costs about half the conversations. BoldChat found that 55% of visitors abandoned when a pre-chat form appeared, 52% on desktop and 68% on mobile, up from 39% in 2009. PepoChat chats start anonymous by default; lead capture can be set to ask for an email before the first message, and BoldChat's numbers are a fair warning of what that does to volume.

Launches, sales and seasonality

No public study gives a seasonal multiplier for chat volume. Conversations follow traffic and confusion, so a launch, a price change, a sale or an outage can push a month well above an ordinary one. Size the plan on your busiest month, not your average, as the chatbot ROI guide also recommends.

Which chatbot plan size do you need?

Convert your estimate into the vendor's unit, then pick the smallest plan that covers your busiest month. Vendors count AI replies, message credits, conversations or resolutions, and the same traffic gives very different numbers in each unit.

With replies or message credits, multiply conversations by replies per conversation. With conversations, use the conversation count. With resolutions, multiply conversations by the share the AI fully resolves. The AI chatbot cost guide works each pricing model at three volumes.

For PepoChat, the unit is the AI reply:

Prices and allowances from PepoChat's billing documentation, last verified 19 October 2026. The visits column assumes BoldChat's 2015 median rate for each traffic band and two replies per conversation.

Applied to the worked examples: 80 replies fit Free. At 360 replies, Free answers about the first 100 conversations of the month and hands the rest to your team inbox; Starter covers the whole month for $49. At 1,300 replies, Starter is $49. At 2,800 replies, Starter is $49 plus $8 of overage, $57. In the busy cases, 9,000 replies cost $119 on Starter, and 36,000 replies cost $389 on Starter or $409 on Growth.

When a Free workspace reaches its 200 replies, conversations are handed to the team inbox with a friendly message instead of failing: a short notice tells the visitor that the team will reply right there in the chat. Every member of the workspace gets an email, and when nobody can reply in the chat, a card asks the visitor for an email address. On Starter and Growth, replies past the allowance are billed on the next invoice, and the meters on the billing page show usage during the month.

On replies alone, Starter plus overage stays $20 a month cheaper than Growth at every volume above 10,000 replies, and cheaper still below. Choose Growth for what else it includes: ten members instead of three (extra seats are $15 a month), no PepoChat branding on the widget, and priority support. The pricing page lists both plans.

How do you measure your real numbers after 30 days?

Count visits, conversations and AI replies for the same calendar month, then divide. On PepoChat, the three numbers come from your web analytics, the team inbox export and the billing page.

1. Pick one full calendar month after launch. The reply allowance and its meter run by calendar month.
2. Visits: in your web analytics, read the sessions for the pages that load the widget, for that month.
3. AI replies: before the month ends, note the AI replies used on Plans & Billing; the meter covers the current month.
4. Conversations: in the team inbox, open Export CSV, set From and To to the month, and count the rows. The export includes chats from the Test the agent drawer, so subtract any you ran that month. A row whose transcript has no user: line is a visitor who opened a chat and left without writing; count those separately as your drop-off.
5. Divide: conversation rate = conversations with a visitor message ÷ visits; replies per conversation = AI replies ÷ conversations with a visitor message.
6. Re-run the formula for your busiest expected month, then compare the result with your plan's allowance.

The team inbox documentation lists the export's columns. If a Free workspace reaches its cap mid-month, compute replies per conversation from the conversations before that date, because later ones were handed to your team without AI replies.

PepoChat does not count widget opens, so the chat-open rate and the message rate cannot be read separately from its dashboard, and the Analytics page shows satisfaction votes and topics, not volume. For an open rate, add your own chat button and track its clicks as an event in your analytics tool. For the metrics to watch once volume is known, see AI chatbot metrics that matter.

What to do next

• Put your monthly visits into the formula at 2 conversations per 100 visits and 2 replies each, and again at the busy case. If both fit under 200 replies, compare the free AI chatbot plans on what happens when the allowance runs out.
• Between 200 and 2,000 replies, compare paid tiers on their overage rate, not only the entry price.
• Start a free PepoChat workspace, measure one full month with the steps above, and choose the plan from your own numbers instead of a benchmark.