THE ECONOMIC TIMES

AI for Marketing Leaders

Masterclass

COMPLETE CLASS NOTES · DAY 2

Sessions by Rishi Jain and Varun Mayya

Notes by Smriti Sharan

Part 1 · The idea underneath everything: folders

The day opened with a small point that turns out to explain most of what followed. Claude Cowork, Claude Code and Antigravity all work on a folder.

What it is

A session is pointed at a folder on your machine. Anything sitting in that folder becomes available as context. References, brand guidelines, past scripts, photos, spreadsheets, all of it.

Why it matters

You stop pasting context into chat. You put it on disk once, and every run after that already knows it.

The same prompt produces different work depending on what is sitting in the folder

@mentions: pointing at a specific file

Inside a prompt, @filename tags a file in the folder and the agent reads it as context.

The bulk ads agent. Note @rishijain.JPG referenced mid-prompt as the visual anchor for every creative.

The prompt in that screenshot is doing four separate jobs at once. It sets the art direction (bold, pattern interrupt, a mix of cartoon and hyper realistic and manga), then uses @rishijain.JPG to lock the face into every output, then adds styling instruction, then asks for variety across the batch.

Watch out @mention only resolves if the file is actually in the folder. Wrong name or a file sitting in Downloads and the agent either errors or quietly ignores it.

Antigravity previews images inside the app, so you can confirm the reference is right before spending a single credit

CLAUDE.md: the file the agent reads every time

If you find yourself repeating the same instructions every session, write them once into a file called CLAUDE.md in the folder. The agent loads it at the start of every session automatically.

The official documentation for the .claude directory, shown in class

• Keep it under 200 lines. Longer files still load, but the agent follows them less reliably.
• If something only matters for one type of task, move it into a skill instead of the main file.
• List the commands you run most often so the agent does not have to be told each time.
• Run /memory to open and edit it from inside a session.

Part 2 · API and MCP: the idea the day rests on

MCPs are less reliable than APIs. API is your best friend. That is directionally right, and it is worth understanding why rather than taking it on faith, because the reason tells you when the opposite is true.

What each one is

An API is a published contract with a service. You call a specific address, you send specific fields, you get specific fields back. LinkedIn publishes one. Google publishes one. It is versioned, documented, and it does not change under you without notice.

MCP is a layer that sits between your agent and that API. It exists so the agent can discover what tools are available and call them in plain language, without anyone hand writing an adapter for each service.

Same destination, one extra translator

Why direct API wins for anything automated

The translator is the cost. It takes time, it consumes tokens describing what it can do, and it can misread intent. None of that matters when you are asking one question and reading the answer. All of it matters when the job runs 500 times overnight.

The decision comes down to one question

The numbers behind this Token efficiency: direct API calls run about 33% leaner than the same work routed through MCP. Throughput: a batch job checking 500 tools takes roughly 50 seconds direct and roughly 25 minutes through MCP. Adoption: MCP is not niche. It passed 97 million monthly SDK downloads by March 2026, and Anthropic donated the spec to the Agentic AI Foundation under the Linux Foundation in December 2025. It is an open standard, not a vendor lock in. The strongest setups use both. MCP to explore and discover, direct API on the hot path. Combined approaches have been measured at 96.3% task pass rates with 63% fewer tokens than prompting alone.

Note 1 · Go to the platform, not the middleman

This was one of the two moments in the day that changed how you think about the stack, so it gets its own section and more depth than the class had time for.

The instinct when you need data from LinkedIn or YouTube or Instagram is to reach for a scraping service. Apify, Phantombuster, whichever aggregator is trending. The better move is to go to the platform’s own API.

Two routes to the same data, and why one keeps working

Why the direct route holds up

The contract is published

A platform API is documented and versioned. When it changes, there is a deprecation notice and a date. A scraper reads the page a human sees, so a redesign nobody told you about breaks it silently on a Tuesday.

The data comes back named

The API hands you structured JSON with typed fields. The scraper hands you HTML that something has to interpret. That interpretation step is exactly where things get guessed.

One less bill and one less dependency

The aggregator charges you on top of everything else, and it is one more service that can go down or change its own pricing. Your token to the platform costs nothing.

The four doors worth learning

These four cover almost everything a content and marketing pipeline needs. The class walked through two of them live.

Platform		Where you get the key	What it lets you do	The catch
LinkedIn		developers.linkedin.com	Post, read, manage on your behalf	Needs a Company Page, most products need approval
YouTube		Google Cloud Console	Search, metadata, captions, analytics	No downloading video, that breaks the terms
Meta		developers.facebook.com	Publish to Instagram and Facebook	Instagram publishing runs through the Meta portal
Google Ads		Google Cloud plus OAuth	Campaigns, reports, ad data	OAuth client secret shows once and never again
The YouTube question, answered properly You can search videos, pull metadata, read comments, and fetch captions and transcripts through the YouTube Data API. You cannot download video files. That is a terms of service line, not a technical one. People use third party tools for it, but do not put that inside a production agent. The useful version for you: pull a transcript, have the agent summarise or reframe it for your own content. Fully allowed, and it is the highest leverage thing the YouTube API gives a creator.
The pattern to internalise Create an app. Get a token. Paste it into .env. Same four steps for LinkedIn, Meta, Google Ads and Google AI Studio. Once you have done it twice, the third one takes ten minutes.

Part 3 · The three places an agent can live

Same prompt, same folder, three different environments. What changes is who approves actions and where the thing actually runs.

Pick by what you are trying to do, not by which is newest

Cowork

You trigger it, you watch it, one task at a time. No approval prompts to click through. Best for exploring and for one off work.

Claude Code

Built for repeatable work. It asks permission before commands, loops over many items, and on Desktop it can be scheduled. This is where you build something you intend to run again.

Antigravity

Google’s agent first development platform, free in public preview for individuals, running on macOS, Windows and Linux. The thing that makes it interesting is model optionality: Gemini, Claude and OpenAI models in the same session, so you can compare mid task. It also has browser in the loop control, which is what powered the scraping demos later in the day.

The same reels scriptwriter agent running in Claude Code, with an Antigravity panel alongside it on Gemini 3.1 Pro

Permission modes: the dial between control and speed

Both Claude Code and Antigravity ship the same idea with the same five settings. Class showed the menu twice, so here is what each one actually does.

The five modes, and where each one belongs

Claude Code asking before running a shell command. This is Manual mode doing its job.

The mode switcher. Manual, Accept edits, Plan, Auto, Bypass permissions.

Auto mode in Antigravity: self-approves anything that passes a safety check, pauses on anything risky

Watch out Bypass permissions plus a .env holding live API keys is how people accidentally post to a real audience. Auto is the sensible default for daily work. Save Bypass for sandboxes.

AskUserQuestion: when the agent stops and asks

Twice in the day the agent paused mid run and asked a question rather than guessing. Both times the reason was the same: the folder was empty, so it had nothing to reason from.

The agent listed the folder, found it empty, and asked which blogs should count as sources

Look at the bash output in that screenshot: total 0. The folder had nothing in it. This is the folder point from Part 1 showing up as a real consequence.

The LinkedIn agent asking the question that actually matters: draft for review, or publish live

Draft only versus auto publish

• Draft only. Agent scrapes, writes, saves to a file. You review and post. Nothing bad ever goes live unattended.
• Auto publish. Goes straight to your feed through the API. Fully hands off, and one bad scrape is public before you see it.

Start with draft only. Move to auto publish when the agent has produced thirty posts you would have approved anyway.

Where it runs when you are not watching

The class covered this with a mind map

The five rungs, from you pressing go to fully headless

useful details on scheduling

• Each task carries its own permission mode, set when you create it.
• Task prompts live on disk at ~/.claude/scheduled-tasks/<name>/SKILL.md. Edit the file directly and the next run picks it up.
• You can manage them conversationally. “Pause my dependency audit task” works from any session.
• The CLI’s /loop is different and lighter. It is session scoped, runs up to about three days, and dies when you exit.
• A worktree toggle gives each run its own isolated Git worktree so scheduled runs do not collide with your manual work.

Modal.com, and when you actually need it

Modal pricing: per second billing on GPUs, nothing charged while idle

Modal is serverless compute. You deploy your code and it runs on their machines. Billing is per second of actual compute with no idle charges, which is the right shape for spiky work like a nightly batch.

The Starter plan gives $30 of compute credits every month at no cost, which is roughly 50 hours on a T4 or about a dozen hours on an A100. Free tier, not unlimited. Two things worth knowing: you get $30 only after binding a payment method (without a card it is closer to $5), and real bills run above the headline per second rates because of regional and non-preemption multipliers. Set a workspace spending limit the moment you sign up.

The honest decision Laptop open and you can trigger it: Cowork or Claude Code. Nothing else needed. Scheduled, laptop shut, light work: a Claude Code Desktop remote routine. Already included. GPU heavy, custom code, or genuine production: Modal or similar. This is where the $30 free tier gets used up fast.

Part 4 · The agents, built live

Six agents in one session. They look different on the surface and they are all the same underneath, which is the useful thing to notice.

Six demos, one skeleton. Learn it once and the rest is swapping the plug.

Agent 1 · Reels scriptwriter

The opening prompt: scrape AI news from official blogs, produce five or six 30 second reels scripts

The instruction that made the room sit up was the one appended at the end: make sure you write the scripts in rishi’s voice.

Two different things called voice

• Written voice. Sentence rhythm, how a hook opens, which words never appear. That is what this prompt means. It is a style instruction.
• Audio voice. A cloned or synthesised speaking voice. That is ElevenLabs territory and it sits at the end of the pipeline, after a script exists.

The full system prompt: role, framework, tone, and a quality guard

Four layers stacked into two sentences, which is why it works:

Layer	What it says	What it does
Role	Act like a copywriter with 10 years of experience	Gives it a professional identity instead of a default assistant voice
Framework	PAS framework only	Every script gets the same spine, so output is predictable and reusable
Tone	Super conversational, like an influencer	Fights the formal corporate drift models default to
Quality guard	Make sure it doesn’t … (cut off)	The anti slop rule. This is the line doing the most work.

Agent 2 · LinkedIn daily post agent

This one got the full treatment: API setup from scratch, then the agent building its own project structure. It is the template for every other connector in the day.

Step 1 · The developer portal

linkedin.com/developers/apps. Note the handwritten annotation: a LinkedIn Company Page is a prerequisite.

Step 2 · Create the app

App name, the Company Page it attaches to, a privacy policy URL, and a logo

Step 3 · Request the products you need

The Products tab. Share on LinkedIn is granted by default. Everything else needs asking.

Step 4 · Client ID and secret

The Auth tab. Client ID, generate a client secret, and the OAuth scopes your app will hold.

Step 5 · The redirect URL

http://localhost:8080/callback. This is where the OAuth handshake completes on your machine.

Step 6 · Hand it to the agent

The prompt, and the CLAUDE.md tip surfacing in the same frame

The prompt asked for the whole thing at once: scrape AI news, write a conversational post, generate a relevant image with his face on it using Nano Banana Pro, create the .env file to hold both keys, publish to LinkedIn. One instruction, whole pipeline.

What the agent built by itself

One prompt in. Scripts, posts folder, .env, .env.example, .gitignore and a 78 line README out.

Worth reading the comment the agent wrote inside the .env file itself: this file is gitignored, never share it or paste it. It also wrote the .gitignore in the same run.

The .env file in the folder tree, where every credential lives

The best trick of the day: screenshot the docs

Rather than let the agent invent a project layout, paste a screenshot of the official .claude directory structure straight from the documentation into the prompt.

The result: agents, commands, rules, skills, workflows, plus assets, config, drafts, logs, published and src

Why this is worth stealing The agent will happily invent a structure. It will be fine, and it will be different every time. Give it the documented best practice as an image and you get production grade organisation on the first run, consistently. Generalise it: any time there is an official way to do something, screenshot it into the prompt rather than describing it.

Agent 3 · Instagram carousel agent

Instagram publishing runs through the Meta developer portal, because Meta owns Instagram. Same four steps as LinkedIn.

developers.facebook.com. Fourteen apps on this account, which tells you how routine this becomes.

Create App, then the same token dance

The folder structure that made it work

The carousel agent mid run, with a reference carousel open in the preview panel

Look at the folder tree on the left. It is not one flat directory:

• queue/ drafts waiting to go out
• published/ what already went live
• References/Design 2/ ten screenshots of other people’s carousels
• Rishi’s Pictures/ his own photos, for the face in each creative

One prompt triggered the whole chain and the agent narrated it: scan today’s verified AI news, fact check, write, design, then publish to Instagram.

How to use reference images without copying

The prompt is explicit: study the structure, do not copy it, and the output must not look like that person’s work

The instruction that makes references legitimate The prompt asks the agent to learn where the headline sits, where the subheadline goes, how emojis are placed and where the face lands. Then it says the copy must be unique and the result must not resemble that person’s carousel at all. That is the difference between reference and plagiarism, and it is a distinction worth stating explicitly in your own prompts.

Agent 4 · Google Ads agent

The route: Google Cloud Console, enable the Ads API, create OAuth credentials

Google Ads API enabled. A REST API for managing accounts, campaigns and reports.

Agent 5 · Blog writer and publisher

The skill stack: blog writer, image creation, publishing agent

Step	Skill	What it needs
Write the post	Blog Writer skill	A keyword and your tone reference
Create the image	Image Creation skill	KIE or Google AI Studio key
Publish it	Publishing agent	WordPress token in .env

Keyword in, SEO structured post out, featured image generated, published to WordPress. No manual step in the middle.

Agent 6 · Faceless video channel

The four prerequisites for an emotional-story channel with no face on camera

What the four branches mean

1. A consistent character. The hardest part. The channel’s whole identity rests on the same character appearing in every video.
2. A scriptwriter. Emotional storytelling has a specific rhythm. Reference scripts in the folder, pointed at with @mention.
3. A video API. Seedance, Kling or Veo. Which key opens which is covered in the next part.
4. An Instagram token in .env. Same setup as the carousel agent.

Watch out on character consistency If the character prompt or seed drifts, the face and style change between videos and the channel stops feeling like one thing. Save the exact character prompt in the folder and reuse it verbatim on every single run. Do not retype it.

Part 5 · .env, the one habit that matters

The class called .env the most protected part of your entire agent setup. That is not an exaggeration.

How the three files work together

What it is

A plain text file in your project folder holding every credential. Your agent reads from it. It never appears in your code.

Why it matters

If a key is hardcoded in a script and you push that script to GitHub, it is public. Someone can post as you, spend your API credits, read your inbox. Adding .env to .gitignore means Git skips the file every time.

How to set it up

1. Create a file called .env in the project folder.
2. Add your keys, one per line: LINKEDIN_TOKEN=xxx
3. Create a file called .gitignore and put .env in it.
4. Push whatever you like. The keys stay on your machine.

.env.example, the file nobody explains The agent created this alongside .env and it is genuinely useful. It holds the variable names with no values. Safe to commit, safe to share, and it tells anyone else exactly which keys they need to supply. Yours stay private.

Part 6 · Models, keys and what things cost

Getting the Google AI Studio key

The API Keys page in Google AI Studio. Left sidebar has Usage, Rate Limit, Spend and Billing, all worth checking before you start generating.

The loaded view. Existing keys, the project each belongs to, and the billing tier. Create API key sits top right.

Sign in, create the key, paste it into .env. That single key opens Imagen, Nano Banana, Nano Banana Pro and Veo.

Set a spend cap before your first call That sidebar has Rate Limit and Spend pages for a reason. An agent in a loop with a bug can burn through a lot of image generations before you notice. Set a monthly cap on day one, the same way you would on Modal.

Image generation, what it actually costs

The pricing slide from class, in rupees and dollars

Cheapest to most expensive, with what each tier is for

The lever most people miss Batch API is a flat 50% off across every model above, if the output can wait up to 24 hours. Bulk ad creative can always wait. A nightly scheduled agent producing tomorrow’s images is exactly the case Batch was designed for.

Manual tools versus API, and why it stops being close

Google Flow is the manual route: a flat monthly fee, and you sit there generating and downloading one at a time. The API is pay per use and the agent does the sitting.

	Google Flow (manual)	API
Cost shape	Flat monthly, around ₹1,950	Per image, from about ₹2
Who does the work	You, click by click	The agent, unattended
Realistic volume	Ten or twenty in a sitting	Hundreds, scheduled overnight
Fits a pipeline	No	Yes

The crossover is early. If you need more than about twenty images a month, or you need them without being present, the API is both cheaper and faster.

The KIE moment

This was the second thing in the day that genuinely reframed the stack. One API key, hundreds of models, at a real discount.

kie.ai. Video, image, music and LLM models behind one API.

The pricing page. Note the discount column against official rates.

What it is

KIE is a unified API gateway. Instead of separate developer accounts, separate keys and separate bills at Google, OpenAI and ByteDance, you top up one credit wallet and call any of them through one consistent endpoint.

Why it matters for you specifically

You were paying attention here because Higgsfield is expensive and you want Seedance videos generated programmatically. KIE is the answer to exactly that, and the maths is not marginal.

Twenty 8-second clips a month, three ways to pay for it

The verified numbers behind that diagram

• KIE. No subscription. Credits do not expire and failed generations are not charged. Discounts run 20% to 86% below official rates, typically 30% to 50%. Veo 3 Fast lands around $0.30 for an 8 second clip.
• Official Google. Veo 3.1 Fast is about $0.15 per second, so roughly $1.20 for 8 seconds. Lite is cheaper at about $0.05 per second. New Google Cloud accounts get $300 in credit, which covers roughly 250 Fast clips.
• Higgsfield. Plus is $34 a month for 1,000 credits. Kling works out around $0.61 to $1.03 per usable clip after iteration. Premium models like Sora 2 and Veo 3.1 cost 40 to 70 credits each, so the Plus plan yields only three to eight usable premium videos a month.

What you give up by going cheap KIE is a reseller, not a direct partner. It has removed models before when a provider objected, and it suspended Seedance access for a stretch in March 2026 before it came back. Latency is uneven. There are credible reports of an eight minute queue for a three second clip on a busy afternoon. Nothing on KIE is exclusive. Every model there is reachable through official APIs or another aggregator, which is your fallback if it disappears. Independent reviews land around 7.8 out of 10, with the cost savings described as legitimate and the experience compared to a budget airline. You get there, do not ask when.

The split I would actually run Drafts, volume, thumbnails, tests: KIE. The savings are real and nothing is at stake. Anything carrying your face or your brand: first party API. Pay the higher price for predictable latency and no dependency risk. Before you commit: run one real Hinglish script through the whole chain and listen to it. Code switching between Hindi and English tech terms is still where synthesis gets shaky.

The video models worth knowing, as of now

Model	What it leads on	Reach it through
Seedance 2.0	Top ranked with audio. Takes up to 9 images, 3 clips and 3 audio inputs per generation. Up to 4K.	KIE and other aggregators
Kling 3.0	Native 4K, 60fps, 15 second clips, multilingual lip sync	KIE
Veo 3.1	Only model with true synchronised dialogue. True 4K. Scene extension up to 20 chained clips.	Google AI Studio
Sora 2	About $0.10 per second base, $0.30 for Pro at 720p	OpenAI directly, or KIE

Part 7 · Running a whole business on one loop

Second session, taught by Varun Mayya. Fourteen years building, six businesses, bootstrapped, and a media network doing over a billion monthly views. His one line of hard earned advice: for the first ten years he built what he found interesting, not what customers wanted, and what customers want is always surprising.

The framing worth keeping is the loop. The agent is not a tool bolted on the side of a business. It is the connective tissue between every part of it.

The slide: THINK, SIZE, BUILD, SELL, PUBLISH, OPERATE

The loop, with what the live demo actually produced at each stage

Humans and AI have different jobs

The split that underpins the whole era of ideas argument

Why this is an argument about scarcity, not a motivational slide

The demo, stage by stage

THINK · what is even worth building

The idea was a CRM for Indian real estate brokers. Instead of building, the first move was research. The agent came back with the regulatory landscape (RERA, the DPDP Act, state specific compliance) and then scoped an MVP.

Eight features in, everything else deliberately deferred

The positioning line it produced is the part worth studying: forward us your enquiries and inventory, and every morning we tell your team whom to contact, what to show them, and which deals are slipping. That is a promise, not a feature list, and it connects directly to broker revenue.

The habit to copy Most people start at BUILD. This started at THINK, and that is why the product had a defined buyer and a price before it had a single feature. The research step costs five minutes and saves weeks.

SIZE · what to charge and how many will buy

Five minutes sixteen seconds of research, turned into a pricing ladder

The reasoning: seven competitors scraped, and a defensible position argued between them

BUILD · the actual product

BrokerDesk. Logo, tagline, and a working pipeline populated with real Indian property data.

Name, logo, tagline, four stage kanban, deals from DLF, Sobha, Brigade, Godrej, Lodha and Kalpataru. From one prompt. This is not a wireframe, it is something you can put in front of a broker on a call.

SELL · finding and reaching buyers

A pitch email the agent researched and wrote by itself

The agent found Larisa Realtech’s YouTube channel and website, worked out their scale (15,000 clients, 200 projects, Gurgaon focus), and wrote a pitch aimed at that specific operation. It cited its own personalisation sources at the bottom: a LinkedIn profile, a YouTube channel, a company page. All public.

The browser is the scraper

Whatever is open in your Chrome tab, the agent can read

Note what the agent said before doing anything: I’ll only collect the information, no calls, messages or contact requests will be made. It drew its own line, out loud.

What this unlocks, and the line to hold Unlocks: NoBroker, Amazon, IndiaMart, LinkedIn, anything you are logged into and browsing becomes a data source. Point it at a page, ask for structured output, schedule it weekly. public data only, and read only. Scraping what a logged in human can see is one thing. Automating contact, bulk messaging, or harvesting private data is another, and it will get accounts banned regardless of how the prompt is worded.

Two ideas from the session worth keeping

The era of ideas

Execution used to be the moat. Who could code, who could design, who could ship fast. AI now does execution. So the remaining scarce thing is knowing what to execute: spotting the right problem, having taste about the answer, making the judgment call.

The arbitrage window, and the honest counter argument

Two views, both defensible Varun’s position: the window is open. Most people have not adopted these tools yet, so whoever builds the habit now has a head start. By the time the rest catch up you already have the audience, the trust and the systems. The window is a time advantage and it closes. What survives after it closes is not output quality, because that levels out. It is niche depth, lived experience the model cannot invent, and a consistent point of view. Information is free now. Perspective is not.

On video editing, honestly

There is a video editing plugin and it works. It is not as good as a human editor. Use it for rough cuts, captions and repetitive batch work. Keep a human on pacing, emotional beats and hook timing.

Part 8 · What content is actually for

The most useful reframe in the content section: views are not the goal. The DM is the goal.

The chain between posting something and getting paid

Why followers matter, and why they are not the score

Followers are a credibility signal. When you reach out to a broker and their team looks you up, a real account with real engagement raises the reply rate. That is what the number is for. It is not the outcome.

The metric to actually watch A thousand followers who send ten DMs a month beats a hundred thousand who send none. Count conversations, not followers.

The content plan the class asked for

Starting point: a short form content plan for the BrokerDesk Instagram account

The full ask: daily Indian real estate news, one minute shorts, funny

The mechanism is simple and it is what your own skills already do. Agent scrapes the day’s news, writes the script, you shoot it, it goes out. The only human step is the one that has to be human.

Two rules stated firmly in class

Use your own face

Personal brand builds trust faster than a faceless brand account. A broker will DM a person, not a logo. Trust comes from a face because a face carries accountability.

Do not use an AI avatar

This was the strongest statement of the session AI avatars used to pass. They do not any more. Newer audiences spot them immediately and trust is gone in a second. The rule: let AI write the script, edit the cut, schedule the post, generate the thumbnail. But you are in front of the camera, in your own voice. That part does not get delegated.

Volume is the strategy

You cannot predict which video breaks out. You can only raise the number of chances. Three months of consistent output, and somewhere between month three and month six something lands.

Where AI actually changes the game Without AI, twenty videos a month is a realistic ceiling for one person doing research, scripting, editing and posting by hand. With the research, scripting, scheduling and publishing automated, the ceiling moves and you only spend time on ideas and delivery. Watch out: do not conclude nothing is working after three weeks. Algorithms take time to trust an account. Three months is the honest minimum before you judge the format.

What volume compounding looks like at the far end. The line on the slide is the point: when one experiment works, it pays for the next twenty.

The reason that slide belongs in a section about volume rather than a section about Varun is the mechanism behind it. Nobody picks the video that breaks out. You run enough experiments that one of them does, and the one that works funds the next round.

Programme mentioned in class

Varun pointed to genc.club as his own structured content creation programme, described as including 90 videos. Worth a look if you want the content system taught end to end rather than assembled from notes.

Part 9 · AEO: how people find you now

Day 1 mentioned AEO in passing. Day 2 gave it the practical treatment, and this is the section with the most immediately actionable material in the whole day.

The shift, in one sentence

SEO earns you a ranking. AEO earns you a citation inside the answer.

Two different games with two different win conditions

Why this is not hype

A page can hold position one on Google and still never be mentioned in the AI summary sitting above it, because citation logic is not ranking logic. Meanwhile the summary is eating the clicks.

The verified picture, as of mid 2026 Roughly 25% of Google searches now trigger an AI Overview, from a study of 21.9 million searches. Position one click through rate drops about 58% when an AI Overview is present. Ahrefs, 300,000 keywords. Visitors arriving via AI citations convert at around 4.4 times the rate of standard organic traffic, and spend about 68% longer on site. They arrive further along in their research. Only about 20% of organisations have started on AEO at all, while 70% believe it will reshape their strategy within three years. That gap is the opportunity.

Third party beats your own website by a wide margin Brands are roughly 6.5 times more likely to be cited by an AI engine through third party sources than through their own domain. That is an Airops finding, and it is the single strongest justification for everything the class recommended. Two more from SE Ranking that point the same way: domains with heavy Reddit and Quora brand mention volume are around 4 times more likely to be cited, and having profiles on review platforms gives roughly a 3 times higher chance of being picked as a source by ChatGPT. Which is why the advice was Instagram, YouTube, LinkedIn and Reddit before your own site. Your website is the hardest surface to get cited and the slowest to build.

Where to put your effort, strongest first

Read it as a work order, not a leaderboard

YouTube is the strongest single channel

Google owns YouTube, so Google surfaces YouTube. A well optimised video ranks and gets cited even with no website behind it.

The title formula in practice: a hook, a pipe, then the money keyword

Reddit and Quora, with real numbers

The class said Reddit is king of community SEO. The data is stronger than that framing suggests.

Reddit’s actual position Reddit accounts for roughly 40% of AI citations across ChatGPT, Gemini and Claude, per a 2026 index built on more than 680 million citations. Google pays Reddit around $60 million a year for content licensing, which is why it shows up so heavily in AI Overviews. Citation is concentrated. The top 15 domains hold about 68% of all citations, which means getting into one of them matters more than spreading thin. And it moves fast. Citation shares have shifted by 50% in under a month after a single algorithm change. This is not a set and forget channel.

How to actually work Reddit and Quora

1. Search your money keyword on Google and look for Reddit or Quora results.
2. If a relevant question exists, write a genuinely useful detailed answer and mention your brand naturally.
3. If no question exists, you may be the only possible citation for that query. Post the question, then answer it.
4. Have the agent do the searching and draft the answer. You post it.

Watch out Reddit communities remove promotional answers fast and will ban accounts. The answer has to stand on its own as useful even if your brand name were deleted from it. If it does not, it is an ad and it will be treated as one.

LinkedIn articles rank on Google

A LinkedIn article with 20 reactions, a year old, holding page one above actual websites

LinkedIn’s domain authority does the work. Write an article with the money keyword in the title and it can rank with no SEO effort at all.

Instagram profiles rank too

No website in that result. An Instagram profile with 2.3K followers on page one for a commercial keyword.

The keyword formula, reused everywhere

Crack Your Instagram: name, bio, captions

One pattern, six placements

What a money keyword is

The exact phrase someone types when they are ready to buy or hire. Not what you find interesting. What they search.

The test If someone asked what you do and you had to answer in the words a stranger would use to find you, that answer is your money keyword. Put that exact phrase in your name field, your title and your bio.

The three placements on Instagram

1. Name field: Business Name, pipe, money keyword. Both halves get indexed.
2. Bio: a second variation of the same keyword. Another layer for the engines to read.
3. Captions: the keyword used naturally in the body, then five or six related keywords at the end.

A real caption: story first, then the keyword block at the end

The mind map version. Natural keyword in the caption, then a dump of five or six at the end.

Watch out Keyword stuffing mid sentence gets penalised twice. Instagram deprioritises spammy captions, and AI engines skip content that reads over optimised. Keep the body natural and put the block at the end.

Google Business Profile and Facebook

Both are quick wins with the same logic. Put the money keyword in the name and the description. GBP owns local and category intent, Facebook is ten minutes of work for one more indexed surface. Neither is glamorous, both get indexed.

Influencer collaborations are AEO, not just reach

When ten creators mention your brand alongside your keyword, that is ten independent citations across ten profiles. The engines read repeated mention from separate sources as a trust signal. This is the same mechanism as the Reddit and Quora finding, applied deliberately.

Semrush and finding the gaps

Semrush tells you which keywords have volume, where competitors rank, and where nobody has covered a topic well. That last one is the useful part: fill the gap and you may be the only citation available for that query.

Watch out on cost Semrush is a paid tool and the free tier is limited. Start with the free searches and Google’s own autocomplete and related searches. Move to paid only when you are running this weekly.

The practical AEO checklist

Beyond what the class covered, three structural moves separate cited pages from invisible ones, and they are worth adding:

• A direct answer block of 40 to 60 words at the top of every major section. Extractable, quotable, self contained.
• Question shaped headings, with FAQ schema markup where you control the page.
• Named entity density in the first 500 words. Mention the specific tools, places, people and products by name early.
• Refresh quarterly. AI surfaced URLs run about 26% fresher than traditional search results. Stale content drops out.

Your AEO weekend, in order 1. YouTube: fix titles to Hook | Money Keyword. Rewrite descriptions with keyword variations. 2. Instagram: name field, bio, and a caption template with the end block. 3. LinkedIn: publish one article with a money keyword title. Codex writes it, you edit and post. 4. Reddit and Quora: find three existing questions in your niche and answer them properly. 5. Google Business Profile: twenty minutes, then leave it. The website comes after all of that, not before.

Part 10 · Everything corrected, in one place

Seven things from the day that needed sharpening. Collected here so you can scan them without rereading the document.

What class said		The precise version
You need Modal.com to run agents with your laptop off		Claude Code Desktop offers remote routines that run on Anthropic-managed cloud with your laptop shut. Modal is for GPUs and custom code, not just scheduling.
Modal is free		$30 of compute credits per month on the Starter plan, and only $30 if you bind a card. Real bills run above headline rates due to regional and non-preemption multipliers. Set a spending limit on day one.
With an API, hallucination is zero		Zero for the data it fetched, because it reads rather than invents. Not zero for the prose written around that data. Trust the figures, check the framing.
Google’s Omni model, accessed via Google AI Studio		There is no Google model called Omni. It is Veo, currently Veo 3.1. Legacy Veo 3.0 endpoints shut down 30 June 2026.
Use KIE not Higgsfield because MCP is unstable		Right conclusion, wrong reason. Higgsfield has no public API to build a pipeline against. It is a subscription dashboard, so it cannot be automated regardless of protocol.
PAS framework, presented as part of the AI toolkit		PAS is Problem, Agitate, Solution, a copywriting framework from the 1960s. Nothing to do with AI. Which means it works in your own writing too, and AIDA and BAB can be specified the same way.
Fable is better but expensive, Sora is cheaper		Different categories. Claude Fable 5 is Anthropic’s reasoning model, announced 9 June 2026, and does not generate video. Sora 2 is a video model. Worth checking the recording for which comparison was meant.
Direct API is measurably faster and leaner. And brands are around 6.5 times more likely to be cited by AI through third party sources than their own domain.

Part 11 · Your two aha moments, turned into a plan

Two things in this day changed how the stack looks. Both are worth acting on this week rather than filing away.

Aha 1 · KIE for Seedance, instead of paying Higgsfield

The reason this landed is that you already knew Higgsfield was expensive and you wanted Seedance clips generated programmatically. KIE answers both at once.

1. Sign up at kie.ai and top up the minimum, which is $5. That is enough to test properly.
2. Put the key in .env as KIE_API_KEY. Confirm .gitignore already lists .env.
3. Generate one 8 second Seedance clip through the API, not the playground. You want to know the call works from code.
4. Time it. Note the actual latency, not the advertised number, because that is the real tradeoff you are accepting.
5. Run one real Hinglish script end to end and listen. Code switching between Hindi and English tech terms is still the weak spot in synthesis.

Where this plugs into what you already have Budget reality: twenty 8 second clips a month lands around $6 on KIE against $34 minimum on Higgsfield Plus, and Higgsfield credits expire in 90 days without rolling over.

Aha 2 · Platform APIs instead of aggregators

This is the more durable of the two, because it changes how you build everything from here rather than just what you pay.

The four doors, in the order that pays off fastest for you

1. YouTube Data API. Highest leverage first. Pull transcripts from any video, have the agent reframe them into your own content. Free, allowed, and it feeds your entire shorts pipeline. Set it up through Google Cloud Console.
2. LinkedIn API. Your main distribution channel. Start the Company Page and product approval now, because approval takes days and everything else waits on it.
3. Meta Graph API. Instagram publishing. Same drill, and worth doing once your carousel format is settled.
4. Google Ads API. Lowest priority unless you start running paid. Skip for now.

The rule to hold Before reaching for a scraper, check whether the platform publishes an API for what you need. It almost always does, it is almost always free with your own token, and it does not break when someone redesigns a page.

Day 2 Summary

The five things that actually matter

1. Context lives in a folder. Not in chat. Reference material on disk, pointed at with @mention, standing instructions in CLAUDE.md.
2. Go direct to the API. MCP for exploring, platform APIs for anything that runs unattended. And go to the platform, not an aggregator sitting in front of it.
3. Keys live in .env, always. With .gitignore. No exceptions, no screenshots.
4. Every agent has the same six parts. Folder, prompt, skill, key, trigger, destination. Six demos, one skeleton.
5. Distribution is a separate skill from building. Content is a funnel to DMs. AEO is how you get cited. Both are free and both compound.

The line from the day worth keeping

Execution is available to everyone now. What is scarce is knowing what to execute. Which is why AI does speed, scale, structure and repetition, and you do judgment, taste, trust and responsibility. That division is not going to move.

End of Day 2 Notes