How to go viral on TikTok: copy what already works in your niche.

Your niche has already run the test for you.

Most new accounts start with what they want to say: a product, a tip, an idea they like. Then they wonder why nobody watches. Meanwhile, in every niche, other people's videos have already been shown to real viewers, and the views and likes tell you which ones worked.

Research turns that into a plan. You learn which openings make people stay, how long the winners run, who is talking and how they are filmed. Then you copy the format, not the video. A format is the shape: the kind of hook, the turn, the length, the setting. It is not anyone's words, face or voice.

This matters even more for an AI character. You can give a character any face, voice and story you like. Research tells you which story the niche already rewards, so you build the character to fit the videos that work, not the other way round.

How to find what goes viral in your niche, step by step.

Twelve steps, from the first search to the third video. Each one takes minutes, not days.

1. Step 1

   Pick the words your niche searches with.

   Write down four or five ways people search for your niche: the plain term, the slang and the code words. Creators in some niches use code words because the plain term returns almost nothing in TikTok search. Search every phrasing, not just one.

2. Step 2

   Pull the top videos.

   For each phrasing, sort by most liked and keep only the last three months. Older hits show what worked then, not now. Save the strongest 20 to 30 videos from TikTok and Instagram Reels.

3. Step 3

   Throw out what does not fit.

   Keep only videos where one person talks to camera about your niche, because that is the kind of video your character will make. Drop memes, news clips, brand ads and anything off topic. A short clean list beats a long noisy one.

4. Step 4

   Put every video in one table.

   One row per video: link, views, likes, like rate (likes divided by views), length in seconds, date, the first spoken line and a one-line summary. The first line is usually in the auto-captions. Writing it down is what turns scrolling into research.

   Column	What goes in it
   Link	So you, or anyone checking, can open the video again.
   Views, likes	As shown on the video, on the day you looked.
   Like rate	Likes divided by views. It shows how much people liked a video, not just how far it travelled.
   Length, date	Length in seconds, and the day it was posted.
   First line	The first thing said, from the captions. Do not paste it into your scripts.
   Summary	One line in your own words.
   Hook shape, creator type	Your labels, added in steps 5 and 7.

5. Step 5

   Find the hook formulas.

   Read the first lines side by side and group them by shape, not topic: a confession, a numbered list, a direct warning, a hot take. A shape that shows up in three or more strong videos is a formula worth copying.

6. Step 6

   Check the lengths and the settings.

   Sort by like rate and look at length. Note where each video was filmed, how many takes it uses and whether a product is on screen. In our research the short videos earned the best like rates, and almost all of the strong ones were one selfie take with no product in sight.

7. Step 7

   Name the creator types.

   Group the creators by the role they play on camera, such as the expert, the insider or the sceptic. Note which type gets the most views and which gets the best like rate. These are the parts your character can play.

8. Step 8

   Choose the part your character plays.

   Pick a type with a high ceiling that you can play honestly. An AI character has no real credentials, so skip the expert role if the niche needs one. In our run the insider with a sense of humour had the highest ceiling and the best like rate, and needed no credentials.

9. Step 9

   Write the character brief.

   In one or two sentences, say who the character is, how they got into the niche, how they talk and what they will never do. The "never" line matters most in health and money niches: no medical advice, no promised results.

10. Step 10

    Turn the top formats into your own scripts.

    Take your three strongest formats. For each, keep the shape: the kind of hook, the turn, the length. Then write every word new, in your character's voice. Copying a format is fine. Copying someone's words, face or voice is not.

    A good test: if the creator watched your video, they would recognise the format but not a single line. Use your own character, too. Never use a real person's face or voice.

11. Step 11

    Make the videos.

    Film each script as one or two selfie takes, the way the winners were filmed, and keep it short. In AI Cast Studio you create the character once, pick a voice, then start each video from your idea or your own words.

    Label the videos as AI when you post. Our guide to creating an AI influencer covers labels, faces that stay the same and voices.

12. Step 12

    Measure, then do it again.

    Post, give each video a few days, then add your own rows to the table: views, likes, like rate. Keep the formats that beat your average and drop the ones that don't. Run the research again every month or so, because what works in a niche changes.

What we found when we ran this on one niche.

We ran this exact process for a new character in the peptide niche: a corner of health and wellness content. The goal was a character whose videos copy what already performs there, not product demos. We give no health advice here, and neither does the character.

The research

An AI agent searched TikTok and Instagram by keyword through a data API. It used several phrasings of the niche, sorted by most liked, and kept the last three months. That took about 18 calls and cost about 3.5 cents in total. It kept only videos where a person talks to camera and dropped memes, news clips, brand ads and off-topic results. For each video it pulled views, likes, length, date, the first spoken line from the auto-captions, and a one-line summary.

Twelve strong videos were left. The top one had about 880,000 views and ran 14 seconds. The others had between 150,000 and 630,000 views.

The six hook formulas

What the numbers said

• Short won on like rate. The best like rates, 6 to 10%, were on videos of 14 to 66 seconds. Long explainers of 70 to 145 seconds still got views, but their like rates sat around 1.5%.
• One take, no product. Almost every strong video was one continuous selfie take with no product on screen.
• Humour, safety and news beat how-to. How-to videos had far fewer views.
• Words matter. Creators used code words, because some plain terms return nothing in TikTok search. Searching only the obvious term would miss them.

The three creator types

What we built from it

One character, modelled on the insider with dry humour: a hairstylist who fell into the niche. She is not a doctor and gives no medical advice. Then three videos, each copying one proven format with new words:

Each was one or two selfie takes, and each filmed in about four to five minutes. No product on screen, and every one under 30 seconds.

How to do it by hand in 30 minutes.

You do not need an agent or a data tool to start. A phone, a spreadsheet and half an hour will get you a first version. Here is how we would split the time.

1. Minutes 0 to 5: the words. List four or five ways to search your niche. Type the plain term into TikTok and look at the search suggestions and the captions of the top results for the words insiders use.
2. Minutes 5 to 15: the videos. Search each phrasing. Where the app offers filters, sort by most liked and pick a recent date range. Save any talk-to-camera video that is well above the others. Do the same on Instagram Reels, where you have fewer filters, so scroll and compare view counts.
3. Minutes 15 to 25: the table. One row per video, using the columns in step 4. Stop at 15 or 20 rows. Work out the like rate for each.
4. Minutes 25 to 30: the pattern. Read the first lines in a row. Circle the shapes that repeat. Sort by like rate and note the lengths at the top. Pick three formats and one creator type.

That is enough to write a character brief and a script for each of three formats. The by-hand version is slower to repeat and easier to fool with one lucky video, which is where an agent helps.

How to do content research with AI agents.

An AI agent is a model like Claude that can use tools on its own: run a search, read the result, run the next one. For this job it needs one tool, a social data API that can search TikTok and Instagram by keyword and return each video's numbers and captions. Several services sell this by the call. Our whole run cost about 3.5 cents.

Give it a tight brief

A loose brief ("find me viral peptide content") gets you a loose answer: famous videos, off-topic hits, numbers it half-remembers. A tight brief says exactly what to search, what to keep, what to drop, what to return and what it must never do. The prompt below is the one to start from.

• What to search: every phrasing, sorted by most liked, last three months.
• What to keep: one person talking to camera, on topic.
• What to return: one row per video, with the link, so you can check every row.
• What it must not do: estimate numbers, or describe pictures it never saw.
• A budget: a cap on calls, and a report of what it spent.

The research prompt

Copy it, fill in the parts in square brackets, and give it to an agent that has a social data tool.

You are researching short videos for a new character in the [NICHE] niche.
Goal: find the talk-to-camera videos that already perform best, so we can copy their formats, never their words.

SEARCH
- Search TikTok and Instagram by keyword with the data tool you have.
- Use each of these phrasings: [PHRASE 1], [PHRASE 2], [PHRASE 3], [PHRASE 4]. Add any code words creators in this niche use instead of the plain term, and tell me which you added.
- Sort by most liked. Only videos posted in the last three months.
- Stop at about 20 tool calls. Tell me how many calls you made and what they cost.

KEEP
- Only videos where one person talks to camera about [NICHE].
DROP
- Memes with no one talking, news clips, brand ads and anything off topic. Count what you dropped and why.

FOR EACH VIDEO YOU KEEP, RETURN ONE ROW
link | platform | views | likes | like rate (likes ÷ views) | length in seconds | date posted | first spoken line, from the auto-captions | one-line summary | hook shape (your label) | creator type (your label)

RULES
- Every number must come from the tool's data. If a field is missing, write "missing". Never estimate or round up.
- You only have captions and numbers, not the picture. Do not describe what a video looks like unless the data says so.
- If a video has no captions, write "no captions" for the first line. Do not guess it.

THEN GIVE ME
1. The table, sorted by views.
2. Each hook shape that appears more than once, and how many videos use it.
3. The average like rate for videos under 60 seconds and for videos over 60 seconds.
4. The creator types you see, and which has the best like rate.
5. The topics that did well and the topics that did badly.

The character and scripts prompt

When you have checked the research, paste it into a new chat with this prompt. It turns the table into a character brief and a script for each of three formats.

Here is my research on the [NICHE] niche:
[PASTE THE TABLE AND THE SUMMARY]

1. Pick the creator type with the best like rate that does not need real credentials. Explain the pick in two sentences.
2. Write a character brief in one or two sentences for a fictional person of that type: who they are, how they got into the niche, how they talk, and what they will never do (for example, give medical or financial advice).
3. Write one script for each of the three strongest hook shapes in the table, each copying its shape:
   - 10 to 30 seconds read aloud.
   - One person talking to camera in one continuous selfie take. No product on screen.
   - Keep the shape: the same kind of hook, the same turn, about the same length.
   - Use all new words. Do not reuse any line, joke or phrase from the research.
   - No health claims and no promised results.
4. Under each script, name the shape it copies and the link of the video it is modelled on, so I can check it.

How to check the agent's work

1. Open three links at random. Do the views, likes and length match what the agent wrote?
2. Redo one like rate by hand. Likes divided by views. If it is wrong once, check them all.
3. Read the dropped list. Did it throw out anything that belonged, or keep a meme it should have dropped?
4. Check the count. If it says a hook shape appears five times, find the five rows.
5. Read the scripts against the first lines. If any line is close to a real video's words, rewrite it.

Where agents go wrong

• Inventing numbers. When a field is missing, a model may fill the gap with a plausible number. The "write missing, never estimate" rule, and your spot checks, catch this.
• Counting off-topic results. A keyword search returns anything that matches the word. An agent that does not filter will count a viral video about something else as proof.
• Describing what it never saw. With captions and numbers only, the agent does not know what a video looks like. If it tells you about the lighting or the outfit, it is guessing.
• Copying too closely. Asked for a script "like" a video, a model may hand back the same joke in different clothes. Ask for new words, and check.

From scripts to videos, with Claude

Once you have a brief and your scripts, the making is the quick part. Describe the character in AI Cast Studio (creating them is 46 credits, and you pick a voice), then start each video from your idea or your own words. If you work in Claude, the AI Cast Studio plugin lets Claude create the character and film for you, and it says the price before spending anything.

Mistakes that waste the research.

• Searching one word. You miss the code words, and with them the creators who use them.
• Looking at all-time hits. Keep to the last three months.
• Judging by views alone. Long videos can pull views that people did not like. Look at the like rate too.
• Copying words instead of formats. Same shape, new lines, always.
• Playing a part you cannot back. An AI character is not a doctor, a lawyer or a financial adviser, and should never act like one.
• Making product demos. In our research almost none of the strong videos had a product on screen.
• Making it long. If the winners are 14 to 66 seconds, a two-minute explainer is a different bet.
• Researching once. What works in a niche changes. Run it again every month or so.