Home Services Work About Blog FAQ Contact
← Back to blog

Why our own chatbot isn't AI (and why that's fine).

The chat bubble that sat in the corner of our site answered fast, sounded confident, and would tell you it was an AI chatbot if you asked. It wasn't one. It was 117 if-then rules, and building it taught us the exact thing we tell clients when they ask whether they need a chatbot or an agent.

Update, November 5, 2026: We've taken the bubble off our site, so this post now describes how it ran while it was live. The tradeoffs still hold.

Every page on genesisweb.org loaded a file called chatbot.js. Click the bubble, type a question, and something answered in under a second. No API call left the browser. No language model ever saw your message. It was a lookup table with a typing animation bolted on, and we built it that way on purpose.

What was actually running when you clicked the bubble

The whole thing was 117 question-and-answer pairs sorted into eight buckets — pricing, process, trust, navigation, and so on. When you typed something, the script tried three things in order. First, did your exact text match a stored question? Almost never. Second, did your text contain one of the stored phrases, or vice versa? Third, did enough individual words overlap that it could guess the closest entry anyway?

If none of that scored above zero, you got a fallback line pointing you to Cole's email. That's it. No memory of what you asked thirty seconds ago, no understanding of the sentence, just string comparison with a scoring system duct-taped on top.

There was even a fake delay — 600 to 1,000 milliseconds of random wait before the answer appeared, so it read as "thinking" instead of "instant database lookup." That line of code was the most honest thing in the file: it admitted the whole experience was theater layered over something much simpler.

What a rule-based bot is actually good at

Here's the part that surprised us a little. Of those 117 entries, 57 lived in the services category — what do you build, do you do SEO, can you make an app. That wasn't an accident. It's what people actually type into a box on an agency site, and a plain keyword match handles "do you build apps" just as well as a model would, for a fraction of the effort.

It also couldn't lie to you about a number. Every price in that file — the $500 AI audit, the $197 social add-on — was one we typed in ourselves. A model-backed version could, in theory, invent a different number on a bad day and say it with total confidence. Ours physically couldn't, because there was no generation step where a wrong answer could appear. It either found the line we wrote or it didn't answer at all.

Where it fell apart

Ask it "how much does this cost" and it nailed the answer. Ask it "what's this going to run me," and it had nothing — no overlapping words with any pricing entry, so it fell straight to the fallback. Same question, different words, completely different outcome. That's the tell. The matcher had no concept of paraphrase, synonym, or intent. It only knew whether your letters showed up near its letters.

Ask a keyword bot the same thing twice in different words, and you've found the edge of what it can do.

Negation broke it the same way. A question built around "don't," "not," or "won't" could still match the positive version of that entry, because the matcher was scoring word overlap, not reading the sentence. And it had no memory between messages — every question started cold, so a natural follow-up like "and how long does that take" meant nothing without the context a person would carry automatically.

Why we didn't just bolt on a language model

We could have swapped the matcher for an actual model call any time. We didn't, because the tradeoff ran the wrong way for what this widget needed to do. A model-backed version costs something per conversation, adds latency, and — the real issue — can answer confidently and be wrong. For a widget whose main job was quoting the audit price and explaining what we do, a bot that sometimes shrugs and hands you to a human beats one that sometimes makes up a number nobody approved.

We wrote up the fuller version of this tradeoff on our AI agent vs. chatbot comparison page — when off-the-shelf wins, when custom is worth the build cost, and why most businesses are right to start with the cheap option.

When the upgrade is actually worth it

The line isn't about sounding smarter. It's about whether the bot needs to touch something outside the conversation. Ours never checked a calendar, never pulled a project status, never remembered you from last Tuesday — it just answered the same forty or so questions over and over. The moment a chatbot needs to look something up in a live system, hold context across turns, or take an action instead of describing one, a keyword matcher stops being a design choice and starts being a liability.

That's the line between a chatbot and an agent, and it's worth getting right before you pay for either one. If you're not sure which side of it your business is on, tell us what you want the bot to actually do and we'll give you a straight answer, including "you don't need to build anything yet."

— Cole

Not sure if you need a chatbot or an agent?

30-minute discovery call. We'll tell you honestly which one fits — including when the free option is the right call.

Book a Discovery Call →