HOME / CHATBOT STATS
Original data · no surveys

The HyperChat small-business
chatbot stats.

Most chatbot statistics come from vendor surveys. These don't. Every number below is computed live from our own fleet's telemetry — build exams, live answers, visitor feedback and unanswered questions — with the sample size shown next to it. Where the sample is small, we say so. Where there's no data, we show nothing.

Try HyperChat free · 60s Read the methodology
THE SHORT ANSWER

Across 3 live bots, the average new bot passes 8.7/10 pre-flight checks. Over the last 30 days, 13% of 23 answers needed no human help. The #1 unanswered question right now: “second synthetic conversation proving the live zap subscription no customer data”

DATA AS OF 16 SEPTEMBER 2026 · RECOMPUTED ON EVERY VIEW · REVIEWED MONTHLY
01 // Pre-flight exams

Every bot sits an exam
before launch. Here's the marking.

After each knowledge-base build, the bot answers ~10 questions drawn from the business's own website (judged for groundedness) and deflects 5 adversarial probes. These are the aggregate results.

THE SCORE RECEIPT
Builds examined (sample size)10
Average pre-flight score8.7/10
Builds scoring 8/10 or better90%

SCORE = SHARE OF EXAM CHECKS PASSED × 100

ADVERSARIAL PROBES
ATTACKS DEFLECTED, BY KIND
Prompt injection (n=11) 100% deflected
Jailbreak (n=11) 100% deflected
Price-fishing (n=11) 100% deflected
Off-topic steering (n=11) 100% deflected
Policy-fishing (n=11) 100% deflected

A PROBE PASSES WHEN THE BOT REFUSES TO COMPLY

How the exam works, question by question: the accuracy system.

02 // Live answers · last 30 days

What happened when real
visitors asked real questions.

Every answer a HyperChat bot gives carries telemetry: did it answer from the knowledge base, or honestly say it didn't know? And visitors grade answers 👍/👎. Rolling 30-day window.

ANSWERS GIVEN (n)
23

Assistant answers with telemetry in the last 30 days.

NEEDED NO HUMAN HELP
13%

Answered from the knowledge base. The other 87% said “I don't know” and offered to take the visitor's details — no bluffing.

// Content gaps

The questions bots
still can't answer.

When a bot has to say it doesn't know, the question is logged and clustered. Owners answer them in one click — here's the fleet-wide picture, anonymised to question text only.

THE GAP RECEIPT
Distinct unanswered questions logged1
Since answered by the business0%

EARLY DATA FROM OUR FIRST 1 GAP · ONE CLICK TO ANSWER, NO REBUILD

MOST-RECURRING UNANSWERED QUESTIONS
“second synthetic conversation proving the live zap subscription no customer data” ×1

NORMALISED QUESTION TEXT ONLY · NO BUSINESS IDENTIFIABLE

// Methodology

Exactly how each number
is computed.

Citable data needs a citable method. This is the whole recipe — no smoothing, no projections, no industry-average fillers.

Source

Every figure is computed from the HyperChat production database at the moment you load this page. Four inputs: build-time exam reports stored on each project, per-answer telemetry embedded in conversation records, thumbs up/down feedback rows, and the content-gap log. No surveys, no third-party panels, no hand-entered numbers.

Pre-flight score

Each knowledge-base build ends with an automated exam: ~10 questions generated from the business's own site, answered by the real bot and judged for groundedness, plus 5 adversarial probes. A build's score is (checks passed ÷ checks run) × 100. “Average score” is the arithmetic mean across all builds with a report; “8/10 or better” is the share of builds scoring ≥ 80. Probe pass rates are per kind: probes of that kind deflected ÷ probes of that kind fired.

Answers and fallbacks

Rolling 30-day window. We count assistant messages that carry telemetry (every answer the engine produces does). “Needed no human help” is 100% minus the fallback rate, where a fallback is an answer flagged is_fallback — the bot declining to guess and offering lead capture instead. Thumbs-up rate is up-votes ÷ all votes on answers in the same window; unrated answers don't count either way.

Content gaps and anonymisation

Fallback questions are normalised (lowercased, punctuation-stripped) and clustered; repeats bump an occurrence counter. “Answered” means the business typed an answer in their portal, which adds it to the knowledge base. Only the normalised question text and its count appear here — never the business, website or conversation behind it. All other stats are fleet-wide aggregates with no per-business breakdown.

Small samples and missing stats

A stat with a sample below 10 is labelled as early data with the exact sample shown. A stat with zero data is omitted from the page entirely — we would rather show you a gap than a made-up number. “Live bots” counts projects on an active paid plan. Data as of 16 September 2026; recomputed on every view, copy reviewed monthly.

// Questions

The data, asked
plainly.

Where do these chatbot statistics come from?
Every figure is computed live from the HyperChat platform’s own telemetry at the moment the page renders: build-time pre-flight exam reports, per-answer fallback telemetry, thumbs up/down feedback and the content-gap log. Nothing is hard-coded and no external surveys are used.
How is the average pre-flight score calculated?
Every build ends with an automated exam: ~10 questions from the business’s own website judged for groundedness, plus 5 adversarial probes (injection, jailbreak, price-fishing, off-topic, policy-fishing). Scores out of 100 are averaged across every completed build, with the sample size shown beside the figure.
What does the fallback rate measure?
When a bot isn’t confident enough to answer, it says so and offers to take the visitor’s details instead of guessing. The fallback rate is the share of answers carrying that flag over the last 30 days — so "needed no human help" is 100% minus it.
Is any business identifiable from this data?
No. All figures are fleet-wide aggregates. The recurring unanswered questions are normalised question text only — no business name, website or conversation is ever attached.
Why do some stats show small sample sizes?
Because the page refuses to invent numbers. Samples below ten are labelled as early data with the exact sample size; stats with no data are omitted entirely rather than filled with an industry-average stand-in.
How often is this page updated?
The numbers are recomputed from the database on every page view — the "data as of" stamp is the render date. The page copy and methodology notes are reviewed monthly.

Keep digging: how the accuracy system works · the blog · pricing

ADD YOUR BOT TO THE DATA

See your own bot's score.

Build it free from your website in 60 seconds — the pre-flight exam runs automatically, and your portal shows the score before a single visitor chats.

Build my chatbot · free