About a month after ChatGPT launched in November 2022, it passed a million users. Two months later it was the fastest-growing consumer app anyone had bothered to measure. The interesting part isn’t the growth curve, though. It’s that years on, a lot of people still use the ChatGPT chatbot the way they’d use a search box: type a question, skim the answer, close the tab.
That leaves most of the tool untouched. A good session looks less like searching and more like talking to a colleague who doesn’t mind your half-formed thinking. A bad session looks like a very confident intern who would rather invent a citation than admit they don’t know.
What the ChatGPT chatbot actually is
Strip the branding away and there are two pieces. There’s the model, which predicts text, and there’s the interface wrapped around it, which is where most of the day-to-day value sits. OpenAI has swapped the underlying model several times, and the free tier runs a lighter version than the paid ones.
The constant is the chat structure: a running thread, each turn building on the last. That thread matters more than newcomers expect. A model with no memory of your previous message is a search engine with better manners. One that can see twenty turns of context can be told “cut it to 200 words, keep the third paragraph, stop using the word ‘robust'” and will actually do it.
Where it earns its keep
First drafts of almost anything
A blank page is the hardest part of most writing, and this is the one job where the tool is close to unambiguously useful. Ask for a rough structure, three competing openings, or a version written for a skeptical reader, and you get something to react to. Reacting is much faster than generating. The output still needs your fingerprints on it, but the first twenty minutes of a task tend to collapse into four or five.
Explaining things at the level you ask for
“Explain how a heat pump works, assume I know basic thermodynamics but nothing about refrigerant cycles” gets a genuinely different answer from “explain heat pumps.” That calibration is the whole trick. It’s also why the chatbot is more useful than a textbook for filling specific gaps: you can keep pushing on the one sentence that didn’t land, and it won’t get impatient.
Grinding through tedious text
Paste in a 40-page vendor contract and ask for every clause involving termination notice. Paste in six months of meeting notes and ask what decisions were actually made, with dates. Paste in a spreadsheet export and ask it to reformat as a Markdown table. None of this is glamorous. All of it is the sort of work that eats an afternoon.
Thinking out loud
There’s a quieter use that regulars swear by: describing a problem badly, on purpose, and letting the replies force you to sharpen it. Explaining what’s actually blocking you to something that asks follow-up questions often surfaces the answer before the model says anything useful.
The failure modes worth knowing about
Fake citations are the famous one. Ask for sources on a narrow topic and you’ll get references with plausible authors, plausible journals and plausible publication years that simply do not exist. This isn’t malice, it’s the model doing what it was trained to do: producing text that looks like the right answer.
Other habits to watch for:
- Sycophancy. Push back on a correct answer and it will often fold. Ask “are you sure?” three times and you can talk it out of being right.
- Confident arithmetic. It will usually get the reasoning right and the numbers wrong. Ask it to show the working, then check the working.
- Stale facts. Anything time-sensitive, from sports results to software version numbers, needs verifying elsewhere.
- Over-polished prose. Left alone, it defaults to a rhythm of short punchy sentence, short punchy sentence, tidy summary. Read three paragraphs of it and you can hear the seams.
If the specific gaps you keep hitting are about a particular kind of work, it may be worth looking at AI chatbot alternatives for the jobs ChatGPT handles badly, since different tools have genuinely different strengths rather than just different logos.
Free, Plus and Pro: what the money buys
The free tier runs a smaller model with usage caps, and most casual users never hit them. The $20-a-month Plus plan gets you the stronger models, higher limits, image generation, file uploads, voice mode and access to Projects, which let you keep a set of files and instructions attached to a workspace. The $200 Pro tier exists for heavy users who want the top-end reasoning models without throttling.
Honest guidance: if you’re using it fewer than three or four times a week, stay free. If you’re pasting documents into it daily or relying on it for coding, Plus pays for itself in the first week. The jump to Pro only makes sense if you’re regularly hitting limits that visibly interrupt your work.
Habits that separate good sessions from frustrating ones
Most complaints about the ChatGPT chatbot trace back to a vague first message. A few things that consistently help:
- Say who the output is for. “A finance director who will skim this on a phone” changes the result more than any prompt trick.
- Give one example of good output. A single sample paragraph outperforms three sentences describing what you want.
- Ask for questions first. “Before you answer, list what you’d need to know” catches a lot of misunderstanding early.
- Iterate in the same thread. Starting fresh each time throws away context you already paid for.
- Use Projects or custom instructions for recurring work. Setting your role, tone and formatting preferences once saves repeating yourself forever.
The app versus the browser tab
Most people meet ChatGPT in a browser and never leave. The desktop app is a different experience in ways that matter: a global keyboard shortcut that summons it over whatever you’re doing, drag-and-drop file handling, and no tab to lose. There’s a decent case that the desktop app beats the browser tab you never close, largely because the friction of opening a tab is exactly enough to stop you using it for small tasks.
On a phone the trade-offs shift again. Voice mode is genuinely good for thinking out loud on a walk, and the mobile apps handle photos of whiteboards or handwritten notes well. Screen space is the constraint, so long documents are painful. What works on mobile is a narrower set of tasks than on desktop, and it’s worth being deliberate about which ones.
How it stacks up against everything else
ChatGPT is no longer the only serious option, and the gap has narrowed in both directions. Claude tends to be better with long documents and more careful prose. Gemini’s integration with Google’s ecosystem is a real advantage if you live in Docs and Gmail. Open-source models have closed enough ground that the free tiers are genuinely usable, which is a change from a year or two ago.
For a side-by-side breakdown of strengths rather than marketing claims, this honest ChatGPT comparison with the other major chatbots is a reasonable place to start. If cost is the driver, the case for where the free open-source HuggingChat option actually wins is narrower than the hype suggests, but it’s real.
Reading reviews without getting fooled
Star ratings are close to useless for chatbots. A one-star review usually means someone hit a rate limit at a bad moment; a five-star review usually means someone was impressed in week one. Neither tells you whether the tool will hold up on your actual work in month four.
What does tell you something: whether the reviewer names specific tasks the tool failed at, whether they distinguish between the free and paid tiers, and whether they mention the model version they tested, since answers from a year-old model tell you almost nothing about today’s. AI chatbot reviews hide more than the star ratings show, and the useful signal is nearly always in the complaints.
One last thing worth doing: keep a short list of the tasks where the ChatGPT chatbot has actually saved you time in the past month, and a shorter list of the ones where it wasted it. That personal scorecard will tell you more about whether to pay for it, switch tools, or change how you prompt than any benchmark or launch announcement ever will.