ChatGPT Thinking Animation: Searching and Reasoning
The most interesting part of an AI conversation video is often not the answer. It is the pause before it.
A prompt lands, and for two or three seconds the interface says something is happening — a pulsing dot, a spinner labelled "Searching the web", a line that says "Reasoning". That gap is where the viewer leans in. It is also the part most fake ChatGPT videos skip entirely, cutting straight from prompt to fully formed answer, which is exactly why those clips feel wrong without the viewer being able to say why.
This guide covers the states before and around the response in the ChatGPT template: the thinking pulse, the labelled activity indicators, error states, the chrome around the conversation, and the timing values that make a sequence read as real. For the text-streaming layer specifically, see the streaming text effect guide.
First, what this actually is
Worth stating plainly, because it shapes how the format should be used: MockClip does not run a language model. It is a deterministic renderer. You write the prompt, you write the response, and the template animates the interface around them — including the activity states a real assistant would show while working.
For video work that is a feature rather than a limitation. A scripted exchange renders identically every time, so you can iterate on pacing without a model producing a different answer on each take, and you can demonstrate a specific workflow without waiting on a live response. It does mean the honesty burden sits with you: if you are demoing a product, the response you write should reflect what the product genuinely does. The same applies to the labels — see the matching rule further down.
The two waiting states
There are two distinct ways the template fills the gap before a response, and choosing between them is the main creative decision in this format.
The thinking pulse. A single filled circle, about 16 pixels across, scaling smoothly between 1.0 and 1.3 on a two-second loop. It carries no text. It appears during the delay before an assistant message when that message has no activity attached. On the dark theme it is white; on the light theme it is black, which is worth knowing if you are matching a brand palette.
The activity indicator. A spinner that turns once per second, with a 14-pixel text label beside it. It replaces the pulse and tells the viewer specifically what is supposedly happening: searching, reasoning, reading a file.
The rule of thumb: use the pulse when the wait is pacing, and an activity indicator when the wait is content. If the point of the video is that the assistant went and looked something up, the label is the point. If the assistant is answering from general knowledge, a labelled indicator would be a lie about what happened, and the pulse is the honest choice.
The activity types and their labels
Seven types are available. Each has a default label, and each label can be overridden with your own text:
| Type | Default label | Reads as |
|---|---|---|
searching | Searching the web | Going out for current information. The most legible to a general audience |
browsing | Browsing | Reading a specific page rather than running a query. Useful when the script names a source |
reasoning | Reasoning | Working through a problem rather than retrieving. Suits multi-step or analytical prompts |
analyzing_image | Analyzing image | Processing an image the user supplied |
generating_image | Creating image | Producing an image. Pairs with the template's inline images, below |
reading_file | Reading file | Working through an uploaded document |
custom | (none — you supply it) | Anything the list above does not cover |
The override field matters more than it sounds. "Searching the web" and "Searching 4 sources" imply different things about thoroughness, and the difference is one field.
The custom type is the one worth remembering. If your video is about a specific tool or workflow — checking a calendar, querying a database, running a test suite — a custom label represents it directly instead of forcing it into the nearest built-in category.
Browser-based. No install.
Open the ChatGPT templateTiming the wait
Four values control pacing, and all four are set per message rather than globally.
| Field | What it controls | Default | Range |
|---|---|---|---|
| Delay before | Seconds before the message begins. For an assistant message with no activity, this is how long the pulse runs | 0 in the editor; the MCP tool uses 1 for assistant messages and 0.5 for user ones | 0 upwards, in 0.1 steps; the MCP tool caps it at 30 |
| Action duration | How long the activity indicator holds | The message's delay, or 1 s | From 0.5 s, in 0.5 steps |
| Stream speed | Characters per second for an assistant reply | 40 in the editor, 45 through MCP | 1–200 in the editor |
| Typing speed | Characters per second for a user message typed on screen | 15 | 1–50 |
A pattern that works for a single-exchange clip: a short delay of under a second after the prompt, an activity indicator for three seconds, then a moderate streaming speed for the response. That is around six to eight seconds before the answer completes — long enough to feel like work happened, short enough for a vertical feed.
For longer content, vary it. If your video has three exchanges, do not give all three the same three-second search. Make one instant, one slow, one labelled differently. Uniform pacing across exchanges is what makes a multi-turn clip feel synthetic, and it is the single most common reason a clip looks fake without the viewer being able to point at why.
Reasoning as its own beat
The reasoning indicator deserves separate treatment because it changes what a viewer expects from the answer that follows.
A searching label promises retrieved facts. A reasoning label promises a conclusion — and a viewer who has watched a reasoning indicator for three seconds expects the response to contain judgement, not a list. If you label the wait as reasoning and then deliver a bare definition, the mismatch registers as disappointing even though the viewer will not articulate why.
This pairing — indicator promises, response delivers — is the most useful discipline in the format. Match the label to the shape of the answer.
The user's half of the beat
The wait before a reply is only half the rhythm. The other half is the prompt arriving, and the template animates that too.
Turn on typing for a user message and the text is typed out character by character at its own typing speed instead of appearing whole. The on-screen keyboard has three modes: auto slides it up while the user types and away afterwards, always keeps it up for the whole clip, and never hides it entirely. Auto is the default and the most lifelike; always suits a clip that is entirely about composing a message; never is right when the keyboard would crowd a square frame.
This matters for the thinking beat because of what it does to the pacing. A prompt that is typed out buys you a second or two of motion before any waiting state is needed, which means a shorter delay reads as natural. A prompt that appears instantly puts the whole burden on the pulse or the indicator.
Error states
Any assistant message can carry an error string. It replaces the reply with a rounded box tinted red, a warning icon at its left and red text — not a bubble, a clearly failed response.
This is a narrow feature with a few strong uses. Content about rate limits, failed requests or context-length problems needs to show the failure, and a screenshot cannot show the timing — the wait, then the failure. Troubleshooting tutorials benefit for the same reason. It is also the honest way to depict a limitation: if your video is about what an assistant cannot do, an error state says so directly.
A sequence that works well: prompt, activity indicator for a few seconds, then the error. The wait makes the failure land harder, because the viewer has already invested attention in an answer that never arrives.
Images and the generation states
An assistant message can carry an inline image, with a position of before or after the text and its own reveal duration (default 2 seconds, from 0.5 up). Paired with the generating_image activity, you get the full arc.
The reveal is specific and worth knowing, because it is what sells the beat: while the image is "generating" it is drawn heavily blurred, with the blur breathing between roughly 18 and 22 pixels and the brightness between 0.7 and 0.9 on a two-second cycle. Then a half-second transition takes it to full sharpness and full brightness. Nothing is generated — it is your uploaded image under a blur that pulses — but the arc reads exactly like generation.
Position matters for pacing. An image before the text means the visual lands first and the text explains it. After means the text builds expectation and the image pays it off. For short-form video, after usually holds attention better, because the wait for the image is doing work. The ChatGPT image generation animation guide covers the effect in more detail.
The chrome around the conversation
The activity states sit inside an interface, and a handful of clip-level settings change what that interface implies before a single message appears.
The header title. Free text, up to 40 characters, drawn at the top with a small chevron beside it — "ChatGPT" by default, and the editor suggests "ChatGPT Plus" as an alternative. Leave it empty for no title at all. It is a title, not a model picker: there is no working model switcher and no real model behind the clip, so keep whatever you write consistent with the story you are telling rather than implying a capability.
The header banner. In place of the title you can show a banner: Get Plus, Memory full, or custom text of your own. Visibility is its own setting — always, only while the conversation is empty, or never. The title is hidden while a banner shows. This is a cheap way to frame a clip about limits, upsells or long-running chats, and it is the kind of detail that makes a mockup read as a real session rather than a clean demo.
Incognito. Switches the interface to the temporary-conversation state, including the privacy notice on the welcome screen and an input placeholder that reads "Temporary conversation" instead of "Ask ChatGPT". Useful when the framing is about privacy, sensitive prompts or throwaway sessions.
The welcome screen. Show or hide it, set your own welcome text, toggle the quick-action buttons beneath it, and optionally keep it visible rather than clearing it when the conversation starts. For a cold open, starting on the welcome screen and having the prompt land gives you half a second of recognisable context before anything happens.
Theme. Light and dark. Dark is the more common choice for short-form because it holds contrast against captions; light reads as more ordinary and workaday, which suits tutorial content. Remember the thinking pulse flips colour with the theme.
Sound. Typing has a sound that is baked into the export and can be switched off — worth doing if you are layering a trending audio over the clip later.
None of these are load-bearing alone. Together they set the frame that the activity indicators play out inside, and leaving all of them at default is a missed half-second of storytelling. The interface realism guide goes deeper on which details viewers actually notice.
Building the sequence from a prompt
If you have MockClip connected to an AI client, create_chatgpt_conversation takes the whole clip in one call: up to twenty messages, each with its delay, streaming speed, typing flag, error string, activity indicator and image, plus the header title, banner, incognito flag, theme, welcome-screen toggles and keyboard mode. Asking for "a prompt about flight delays, a three-second searching indicator, then a four-line reply streaming at 30 characters a second" gets you a link that opens exactly that. How to connect MockClip to Claude and ChatGPT covers the setup; the same framing fields let you ask for a square clip or a Full HD render in the same sentence.
Where this format fits
- Product and workflow demos. Script the exact exchange your product enables. Because the render is deterministic, you can iterate on wording and pacing without a live model producing different output each take. See ChatGPT conversation video for marketing demos.
- Prompting tutorials. Show a prompt and the shape of a good response. The activity indicator communicates that the assistant did work, which is often the teaching point.
- Explainers about how AI tools work. The activity states are the most legible way to show that an assistant retrieves, reads and reasons rather than simply knowing.
- Troubleshooting and limitation content. Error states, covered above.
- Cold opens. A prompt and a three-second search indicator is a complete hook in under five seconds, with the answer withheld until after your intro.
Common mistakes
Skipping the wait entirely. Cutting from prompt to complete answer is the most common error in fake AI conversation videos. It removes the beat that makes the exchange feel like an exchange.
Indicators too short to read. Under a second, the label is decorative. You know what it says; the viewer does not.
Labels that do not match the answer. A reasoning indicator followed by a one-line retrieved fact, or a searching indicator followed by an opinion. The mismatch registers even when the viewer cannot name it.
Uniform pacing across every exchange. Identical delays and streaming speeds across a multi-turn clip flatten it. Vary at least one value per exchange.
Overusing activity indicators. If every message is labelled, none of them is a beat. Use the plain pulse for the neutral waits so the labelled ones stand out.
Stacking long waits. Three exchanges with three-second indicators each is nine seconds of spinner in a clip people will give you fifteen. Budget the waits against the whole clip, not message by message.
Writing responses that oversell. Since you author both sides, it is easy to write an assistant response that claims more than your product does. Keep it accurate.
Quick start
- Open the ChatGPT template
- Write the user prompt as the first message and turn its typing animation on
- Add the assistant response and attach an activity to it
- Pick the type —
searching,reasoning, orcustomwith your own label - Set the activity duration to around three seconds
- Set a streaming speed that lets the response be read at a natural pace
- For a failure sequence, add an error string instead of a response
- Set the header title or a banner, pick the theme, then preview and export
Export tiers and watermark options are listed on the pricing page.
Related templates and guides
- ChatGPT template — the editor used in this guide
- How to create a fake ChatGPT conversation video — the format pillar
- ChatGPT streaming text animation effect — the text-streaming layer
- ChatGPT image generation animation — the blur-to-reveal image effect
- ChatGPT video interface realism — which interface details viewers notice
- ChatGPT conversation video for marketing demos — the product-demo application
- AI assistant demo videos for client products — which template to use for a client's demo
- Connect MockClip to Claude and ChatGPT — generating these clips from a prompt
- Why fake chat videos look fake — the pacing mistakes in full
- All MockClip templates — the full format index
- MockClip vs CapCut — templates versus manual editing
- Pricing — export tiers
Frequently Asked Questions
What is the ChatGPT thinking animation?
It is the state between a prompt being sent and the response starting to stream. In the MockClip template it renders as a single filled circle about 16 pixels across that scales smoothly between 1.0 and 1.3 on a two-second loop — white on the dark theme, black on the light one. It shows during the delay before an assistant message that has no activity indicator attached.
What is the difference between the thinking pulse and an activity indicator?
The pulsing circle is the generic waiting state and carries no text. An activity indicator replaces it with a spinner that turns once per second next to a label such as Searching the web or Reasoning. Use the pulse when the wait is only pacing, and an indicator when you want the viewer to know what is supposedly happening.
Which activity states can I show?
Seven: searching, reasoning, analyzing image, generating image, reading file, browsing and custom. Their default labels are Searching the web, Reasoning, Analyzing image, Creating image, Reading file and Browsing; custom has no default and takes your own text. Any type's label can be overridden, so Searching 4 sources is a one-field change.
Does MockClip actually search the web or run a model?
No. MockClip is a deterministic renderer. It animates the interface states of an AI conversation using text you supply, and the header title is free text rather than a working model picker. Nothing is generated, no search is performed and no model is called — you write both sides of the exchange.
How long should an activity indicator run?
Two to four seconds reads as realistic for searching or reasoning. Under a second the viewer cannot finish reading the label, so the beat is wasted. Beyond about five seconds a short-form audience scrolls unless something else is on screen. The editor's floor is 0.5 seconds and the MCP tool accepts delays up to thirty.
Can I show an error response?
Yes. Any assistant message can carry an error string, which replaces the reply with a rounded box tinted red, a warning icon and red text. It is the honest way to depict a rate limit, a failed request or a limitation, and it is far more legible than a screenshot because the viewer sees the wait and then the failure.
Can I control how fast the response types out?
Yes, per message. Streaming speed is in characters per second — the editor defaults to 40 and accepts 1 to 200 — so a one-line answer can appear briskly while a longer explanation streams slowly enough to read. User messages have their own typing speed, defaulting to 15 characters per second.
Is this suitable for product demo videos?
Yes, and it is one of the more common uses. Because you author both the prompt and the response, you can script an exact demo without waiting on a live model or re-recording when the output varies. The honesty burden sits with you: the reply you write should reflect what your product genuinely does.
Related Articles
How to Create a Fake ChatGPT Conversation Video
The complete 2026 guide to creating realistic fake ChatGPT conversation videos for TikTok, YouTube Shorts, and Reels — no coding, no editing, free.
How to Recreate the ChatGPT Streaming Text Effect in Videos
Add the iconic ChatGPT word-by-word streaming text animation to your videos. No coding or screen recording needed.
Animate the ChatGPT Image Generation Effect (Blur Pulse to Reveal)
Recreate the ChatGPT AI image generation animation — blur pulse while generating, then sharp reveal — in video mockups with no screen recording.
ChatGPT-Style Chat Videos for Product Marketing Demos
Make a ChatGPT-style chat video for your product demo or ad: MockClip's AI Assistant Chat streams scripted answers under your own product's name and colour.
How to Connect MockClip to Claude and ChatGPT via MCP
Connect MockClip to Claude Desktop, ChatGPT or any MCP client with a free API key, then turn one prompt into a chat mockup video you export in the browser.