The upcoming Anthropic IPO promises large payouts, but not equally for everyone. Investment funds like Menlo Ventures and Sequoia expect billions. However, according to Anthropic’s S-1, none of them own 5% of the shares, which is the threshold for inclusion in the list of major shareholders. This is highly atypical for IT IPOs, where leading venture investors usually hold double-digit stakes. Why did this happen, and who will be the main beneficiary?
Why did VC funds get crumbs?
The reason is simple: Anthropic’s size. With a private valuation of $965 billion, 1% of shares is worth almost $10 billion. For example, Menlo Ventures’ position of $14 billion accounts for only about 1.5% of the company, according to Bloomberg.
Venture funds did not receive large stakes because Anthropic actively sold shares to strategic partners. This changed the traditional ownership structure, where VC funds usually hold dominant positions.
Amazon: The main beneficiary of the IPO
Amazon is likely to be the biggest financial beneficiary of the IPO. In 2023, Amazon pledged to invest up to $4 billion in Anthropic, and then injected billions more. Most of the investment was structured as convertible bonds – debt that converts into equity upon IPO or sale.
According to the S-1, as of June 30, Amazon held $6.2 billion in convertible bonds.
As Anthropic’s valuation grew, the value of these bonds increased significantly.
In its own reports, Amazon valued its bonds at $97.9 billion as of June 30.
Accounting games and real cash
The increase in the value of Amazon’s convertible bonds forced Anthropic to reflect this in its financial statements. In the first half of 2026, the company recorded $72.4 billion in the line item “change in fair value of convertible bonds, derivative liabilities, and SAFEs.”
This is a non-cash item; no real money left the company. Amazon’s bonds can be converted into 1.84 billion shares, but there is a limit of 12% of the “partially diluted” share count. Amazon can waive this limit.
Additional Amazon investments
Bonds are not all. In the second quarter, Amazon additionally invested $10 billion in non-privileged preferred shares, investing $5 billion in each of Anthropic’s last two funding rounds.
The total amount of Amazon’s investments, including billions in bonds from 2023 to 2025, reaches approximately billions. This makes Amazon a key player in Anthropic’s future ownership structure.
Not just shares: AWS and Google
In addition to the increase in share value, Amazon will receive significant revenue from Anthropic, which has committed to spending at least $110 billion on AWS over the next 10 years. Google is also a winner:
Google owned about 14% of Anthropic early last year.
The company invested another $10 billion in April.
Google’s stake is now estimated at 5.6% after dilution.
Google has also secured at least $114 billion in cloud contracts with Anthropic until 2033.
Conclusions for investors
“While venture capitalists prepare for their payouts, the S-1 clearly shows: the biggest winners of the Anthropic IPO will be tech giants like Amazon, who supported a potential competitor at an early stage and are now selling it most of its computing power. The S-1 even warns that some of these cloud providers are creating competing AI models and their interests do not always align with Anthropic’s.”
This is a classic case where strategic partners skim all the cream, leaving VC funds with only crumbs. To evaluate profitable CPM and cheap hypothesis testing, it is important to look at the whole picture, not just the stated investments.
Frequently Asked Questions
What CPM should be considered profitable with such investments?
When evaluating advertising campaigns, a profitable CPM is one that provides a positive ROI. In the context of the Anthropic IPO, CPM is not a direct indicator, but it reflects the cost of attracting attention to a startup that is already valued at hundreds of billions. For us, arbitragers, unit economics are important, not hype.
Is it worth investing in AI startups at an early stage?
As the Anthropic example shows, early investments can bring huge profits, but only if you are Amazon or Google. For ordinary investors, the risks are significantly higher. It is important to analyze not only the potential, but also the deal structure, and who the strategic partner is.
What is cheaper: banner advertising or influencers for promoting an AI product?
This depends on the target audience and campaign goals. Price comparison shows that banner advertising can be cheaper for 1 million impressions, but influencers often provide higher engagement. For AI products, where expert authority is important, influencers can be more effective, despite a higher CPM. Always calculate ROI.
Conclusion
Anthropic IPO is a prime example of how major players like Amazon and Google use strategic investments to generate colossal profits, significantly exceeding the returns of traditional venture funds. This is a lesson for anyone looking for budget alternatives to seeding and striving for maximum ROI. Always count the money, not the promises. Want to know how to get the most out of every ruble? Read our cases on ROI optimization.
Despite multi-billion dollar lawsuits and litigation, Meta demonstrates impressive revenue growth, especially in the third quarter. How does the company manage to stay afloat and even thrive amidst such financial challenges? The answer lies in strategic investments in artificial intelligence, which is becoming a key driver of Meta’s advertising revenue growth, even when new products like Muse are not yet profitable. Let’s delve into the details of this phenomenal success and learn how Meta uses AI to optimize its advertising platforms.
Meta: 26% revenue growth despite Muse’s unprofitability and AI’s influence
Meta has had a challenging year, marked by multi-billion dollar lawsuits and litigation. One of them, related to the harmful effects of social networks on children, led to an $18 billion settlement. Another lawsuit, concerning data privacy violations, resulted in an order to pay the state of New Mexico $25 billion. This decision was made hours before Meta’s third-quarter report, and the company has already stated its intention to appeal it.
Despite these legal battles, the third quarter was strong for Meta. Revenue from the family of apps grew by 26% year-over-year, reaching $64.2 billion. Of this amount, $63.1 billion came from advertising revenue. The number of ad impressions also increased by 10%.
Muse: Popularity without Profit
Weeks after Muse’s launch in early September, Meta introduced Muse for small businesses. This is a “new set of skills and connectors” within the Muse app that helps companies find customers, analyze ad effectiveness, optimize it, and develop campaigns, as Meta CFO Susan Li told investors.
From a consumer perspective, the launch of Muse was “one of the most successful launches in the company’s history,” according to Meta CEO Mark Zuckerberg. This claim may be debatable given the privacy concerns associated with early versions of the model. However, as with other Meta platforms, privacy concerns have not hindered widespread adoption.
Zuckerberg noted that all these agents are “very computationally intensive.” Meta expects cost efficiency to improve 10x over the next year, which “will allow it to become profitable once revenue lines scale,” according to Zuckerberg. In other words: Muse is popular for now, but not making money.
If Meta’s ad revenue growth cannot be attributed to Muse, then what is the reason? The answer is artificial intelligence, which has long been central to Meta’s overall ad platform strategy.
Earlier this year, Meta launched Generative Recommender — an LLM-based tool that analyzes ad content and user preferences to determine the most relevant ads for a specific person. Target audience segments have seen an almost 2% increase in conversions since the tool’s launch in Facebook feeds, Meta CFO Susan Li said.
Meta also expanded GEM, its generative ad recommendation model, to include more user touchpoints, Li said. This means Meta can now unify recommendation data from various platforms in one place. This capability has led to:
A 1.7% increase in conversions for external campaigns.
A 9.6% increase in in-app conversions last quarter.
“AI investments have been benefiting Meta’s advertising business for many years. As LLMs become the foundation of more and more of the AI infrastructure, Meta will gain more insights into people’s preferences and interests, both in organic content and recommendations,” — Susan Li, Meta CFO.
The goal is not only to fuel Meta Advantage+’s ad optimization engine but also Meta’s creative personalization features.
Netflix: Revenue Growth and Advertising Challenges
Netflix’s revenue grew 12% year-over-year to $12.9 billion in the third quarter, meeting forecasts. However, the advertising business is slightly behind expectations, which led Netflix to adjust its forecast for the remainder of the year.
Frequently Asked Questions
What was Meta’s total revenue growth in the third quarter?
How much did Meta’s total revenue grow in Q3?
Meta’s total revenue grew by 26% year-over-year in the third quarter, reaching $64.2 billion.
How much ad revenue did Meta generate in Q3?
Meta’s ad revenue was $63.1 billion in the third quarter.
Is Muse profitable for Meta?
No, Muse is not currently profitable. Mark Zuckerberg noted that Muse agents are “very computationally intensive,” but Meta expects a 10x improvement in cost efficiency next year.
Which AI tools contributed to Meta’s ad revenue growth?
Growth was driven by the Generative Recommender, which increased conversions by 2%, and the expansion of GEM, which led to a 1.7% increase in conversions for external campaigns and a 9.6% increase for in-app campaigns.
How much did Netflix’s revenue grow in Q3?
Netflix’s revenue grew by 12% year-over-year, reaching $12.9 billion.
Conclusion
So, Meta is demonstrating strong revenue growth, primarily due to investments in AI, which significantly improve advertising performance. Products like Muse are currently unprofitable but have potential. While Netflix faces challenges in its advertising business, Meta continues to prove that favorable CPM and cheap hypothesis testing with AI optimization are a reality. Want to know how much money each ruble spent will bring? Learn how AI is changing the unit economics of advertising campaigns.
Hello, fellow arbitragers! My name is Timur, and I am responsible for the marketplace at Cloud.ru. We provide a platform for AI services that integrate into clients’ cloud infrastructure. My task is to ensure pricing transparency. Currently, many evaluate AI by token cost, but this is fundamentally wrong. The client pays not for tokens, but for a specific result. In this article, we will analyze how to correctly calculate the unit economics of an <a href="/ru/blog/massovyy-posting-video-4-strategii-povtoreniya-zadach-ii-agentami/" title="Mass video posting: 4 strategies for repeating tasks by AI agents">AI agent, based on Cloud.ru data, and determine the real cost of a useful response to understand what CPM is profitable.
“A vendor on the marketplace must set a price for their agent, the client must understand it, and the platform must make it possible to compare prices on the storefront. As long as all three parties measure cost in tokens, none of this works.”
Evaluating the Cost of AI Services: Focus on Unit Economics
Unit of Measurement: Useful Response, Not Tokens
A token is a bad metric. The client buys not tokens, but a solution to their problem. A “closed dialogue” is also deceptive. Intercom, for example, distinguishes two types of solutions: confirmed (the client said “thank you”) and assumed (the client simply left).
Both are paid for, but there’s a nuance: if the client returns with the same question, the solution is annulled. An example from Lorikeet (195 cases, 38 vendors): out of 4200 dialogues closed by AI, 900 were closed with a link to a knowledge base article. Without them, the share of resolved issues drops from 70% to 55%.
The difference between “closed” and “resolved” is huge. Therefore, our unit of measurement is a useful response. This is a dialogue closed by AI without human intervention, where the client did not return with the same question within 72 hours and did not leave a negative rating. Everything that did not pass these filters goes into the numerator of costs, but not into the denominator of useful responses. Unsuccessful attempts are paid for by successful ones. This is a key point.
Model Case: Online Store Support
For calculations, let’s take a typical scenario: online store support. This is a mass case with a large volume of public data. The store receives 40,000 inquiries per month. 90% of them (36,000 dialogues) are routed to the AI agent.
Dialogues closed by the agent without an operator, let’s call them auto-closures – this is the metric that vendors show. Agent structure: RAG (knowledge base search + response generation) plus the “order status” tool. One dialogue accounts for three LLM calls:
Understand the request and choose a route.
Respond based on the found fragments.
Refer to the tool or compile a summary for the operator.
On average, this is 12,000 input tokens (system prompt, fragments, history) and 800 output tokens. There are two knowledge base queries per dialogue. The filter model checks the client’s reply and the agent’s response (1000 tokens for each of the three calls). Another 15% of tokens are allocated for re-generations.
Infrastructure and Costs
We use Cloud.ru services:
Knowledge Base: Evolution Managed RAG.
Agent: Evolution AI Agents.
Logs: Evolution Managed PostgreSQL.
Cost of a support operator in Moscow: 125,000 RUB/month (85,300 RUB salary according to hh.ru for 2025 + contributions and overheads). With 40 tickets per day, one ticket costs 149 RUB. Operator minute – 15.5 RUB (125,000 RUB / 80% of 10,080 minutes). For chat in retail (60 requests/day), the ticket cost is 99 RUB.
Assumptions for Calculations
Key assumptions:
Share of auto-closures: median for public cases – 70%. Ozon: 27% (2020) -> 62% (2022). We take the middle – 50%.
Dialogues that did not pass the filter: client returned within 72 hours or gave a negative rating. Assumption: 6% (in the Lorikeet example, 21% of closed dialogues were links).
Cost of such a dialogue: 150 RUB compensation + 3 minutes of operator time (200 RUB).
Operator time for escalation: 1.5 minutes beyond a regular ticket.
Manual check: 3% of all dialogues for 3 minutes each.
Total infrastructure: 13,420 RUB/month. This is the cheapest line in the calculation.
Calculation of the Cost of a Useful Response
Cost of one dialogue
First, we calculate the cost of one dialogue using different models:
Qwen3-30B-A3B: 0.43 RUB
GigaChat-3-Pro: 1.32 RUB
Claude Haiku 4.5: 3.76 RUB
Claude Sonnet 4.6: 10.99 RUB
A 25-fold difference! If we stopped here, model selection would be critical.
Cost of a useful response
Out of 36,000 dialogues, with 50% auto-closures, 18,000 were closed. Of these, 6% were errors. This leaves 16,920 useful responses. Each useful response incurs not only token costs but also:
Tokens from failed attempts.
Operator minutes on escalations.
Cost of errors.
Verification.
Knowledge base.
Implementation.
The 25-fold difference in tokens translates to a 1.4-fold difference in useful response cost: from 51 RUB to 73 RUB. The expensive model (Claude Sonnet 4.6), which is 25 times more expensive in terms of tokens, turns out to be only 40% more expensive in terms of result cost. Why? Because all other costs are the same for all models.
Price structure: GigaChat-3-Pro (53 RUB)
For GigaChat-3-Pro (53 RUB per useful response):
Escalations and erroneous responses: 71% of the cost. These costs are usually not accounted for as “AI expenses” but are embedded in support salaries or compensation.
Tokens: 2.8 RUB (5%). This is the only line item everyone compares.
Cloud infrastructure: 0.80 RUB.
For Claude Sonnet 4.6, tokens already account for 32% of the price. Here, it makes sense to save on tokens. How?
Context reduction: reduce 12,000 input tokens to 6,000. The useful response on Sonnet becomes cheaper from 73 to 65 RUB.
Caching: reading from cache is 10 times cheaper. If half of the context is a system prompt, the price drops to 66 RUB.
On Qwen, these techniques save pennies.
Key levers for savings
1. Auto-closure rate
Transitioning from a pilot (35-50% auto-closures) to a production environment (65%) saves 15-19 RUB. Ten percentage points of auto-closures yield as much as switching to a cheaper model (on Sonnet, an increase from 50% to 60% saves 14 RUB, replacing Sonnet with Haiku saves 15 RUB). But the former works on any model, the latter – only once.
2. Error rate
Few measure the percentage of closed dialogues that failed the filter, but it is critical.
Percentage of failed filters
Cost of useful response
Expenses for them (RUB/month)
6%
X
Y
3%
X – 8 RUB
Y – Z
Reducing errors from 6% to 3% saves 8 RUB on each useful response. Switching from GigaChat-3-Pro to Qwen3-30B-A3B saves only 2 RUB. 15% errors can cost more than half of an operator’s salary if every dialogue where the client went silent is considered resolved.
3. Escalation quality
I allocated 1.5 minutes of operator time for escalation (reading summary, continuing). This works with a “warm” handover: the agent provides the operator with a concise history, order number, and list of actions. With a “cold” handover (the operator reads the entire dialogue where the bot answered incorrectly three times), 1.5 minutes turn into 3 or 6.
At 3 minutes, a useful response on GigaChat-3-Pro increases in price from 53 to 77 rubles, at 6 minutes – up to 127 rubles. This is already more expensive than a regional operator. If the agent identifies the client and pre-fills the ticket, the operator saves a minute or two, and this line goes into the red.
“Transmission quality turns out to be a separate budget item with the widest spread in the entire model, and it should be measured: the average processing time for escalated tickets should be calculated separately from regular ones.”
Fighting for accuracy = burning tokens? If, to reduce errors from 6% to 3%, tokens per dialogue increase by 1.6 times, on GigaChat-3-Pro, a useful response still becomes cheaper from 53 to 46 rubles. On Sonnet, extra tokens are more expensive than the accuracy gained; it’s better to improve the knowledge base there.
Conclusions: what to consider a profitable CPM
Practical conclusion: in the working support circuit, the most expensive thing is not tokens, but the denominator. Everything that increases the proportion of useful responses (knowledge base quality, correct routing, test sets) pays off faster than any bargaining for the price of a million tokens.
Expensive models (e.g., Sonnet) pay off if they bring 10 points of auto-closures and 2 points of accuracy compared to cheaper ones. If the difference on your test set is less, you are overpaying. If it’s more, you are unnecessarily saving on tokens.
Comparison with a human: two caveats
1. Ticket price: at 149 rubles per ticket, even a pilot with 35% auto-closures on Sonnet (110 rubles per useful response) looks profitable. Parity with an operator is at 20-27% auto-closures. But if an operator closes 60 tickets a day (99 rubles/ticket), the same pilot on Sonnet (110 rubles) is already more expensive. A regional/outsourced operator (64 rubles/ticket) changes the picture: GigaChat-3-Pro (42 rubles) is cheaper, Sonnet (62 rubles) – parity. The question of “will it pay off” depends on your ticket cost and the proportion of auto-closures.
2. Volume: fixed costs (knowledge base, implementation, infrastructure) – 163,000 rubles/month. For 36,000 dialogues. For 6,000 dialogues, a useful response on GigaChat-3-Pro is 101 rubles, for 3,000 – 159 rubles (more expensive than an operator). Below 3,500 dialogues per month, the agent does not pay off. If fixed costs are higher (a full-time position for the knowledge base – 200,000 rubles/month, implementation – 1.5 million rubles), the break-even point shifts to 7,000 dialogues per month. The share of tokens in this case drops to 4%: the more realistically you calculate other costs, the less significant the price per million tokens appears.
What is not accounted for in the model (in both directions)
Against the agent: ticket system or chat platform license.
In favor of the agent: staff turnover (up to 50% per year), recruitment and training costs, slow work of newcomers, scaling for sales, payment for night/weekend shifts. These things are difficult to calculate without specific data, but businesses are often willing to pay for them, even if daily unit economics show parity.
Frequently Asked Questions
What is a “helpful answer” in the context of an AI agent?
A helpful answer is a dialogue closed by AI without human intervention, where the customer did not return with the same question within 72 hours and did not leave a negative rating. This is a key metric for evaluating the real effectiveness and cost of AI services.
We are looking for a profitable CPM where competitors are losing customers. Cancellation requests and complaints about competitor products are not junk, but a goldmine for PPC. These are high-intent leads ready to convert. One SaaS client case showed a CTR of 15% and a CPA 18% lower than in regular campaigns. And this was despite the query not containing the client’s product name.
People were not looking for a new provider, but a way out of a situation where the old one had become annoying. The decision to leave was made; the question remained, “where to?” This case confirms: the most valuable thing you can change in an account is not the campaign, but the offer.
Traditional Keyword Approach: Why Is It Expensive?
PPC specialists are used to collecting keywords: product terms, modifiers, comparisons, long tails. The logic is simple: bid on what people search for when they want to buy our product. The problem is that this only captures obvious demand.
“Anyone searching for ‘best CRM software’ is already in the market, and all competitors know it. The auction is overheated, CPC is through the roof. The click goes to whoever has more money or better optimization.”
Compare this with the query “cancel [competitor] contract.” Few people bid on it, and many even exclude it. The user is not comparing products; they are solving a different problem, and this makes them the most receptive buyer.
They have already passed the most difficult part of the sales cycle: they have decided that the current solution is not working. This is not an exception; it is a pattern. The highest-intent queries often do not contain product terms but express frustration, a desire to leave, and unresolved problems that your offer can solve.
This data is not about the features people want, but about the problems that are currently bothering them. A feature request is a plan for the next quarter. A complaint is ad copy for today.
Why Do PPC Teams Ignore This Data?
Often, this data resides in other departments: support, product, CX. The PPC team receives a list of keywords and a landing page, and they are told to “make the numbers.” Integrating data from CX, product, and sales into paid search is precisely the M-shaped skill needed for a new approach to PPC. Data silos are expensive.
“Every competitor customer complaint is a proposition for your ad. Every reason for cancellation is an objection that your landing page must address before the call to conversion.”
The data exists, but it doesn’t reach those who can use it to improve campaigns.
How to get started with “cancellation” queries
By default, “cancel” is often added to negative keywords. From the perspective of relevance to a regular landing page, this is logical. From the perspective of intent, it’s a mistake. You are excluding someone who is actively leaving a competitor and has not yet chosen a replacement.
This approach is successfully applied in local SEO to optimize Google Business Profile and website content. The same complaint data works great for PPC campaigns.
The offer solves everything, not just the keyword
Simply finding smart keywords and bidding on them is not enough. If the landing page is the same as for regular clicks, there will be no conversions. Someone looking to “cancel contract [competitor]” doesn’t need a list of your features. They need to know three things:
The transition will be easy.
You understand why they are leaving.
The specific pain that led them to search will not follow them to you.
Ideally, offer a full switching service, data import, or onboarding assistance. The offer is the variable. Not the bid, not the match type, not even the ad copy (although it is critical here, as it must immediately signal that you understand the user’s problem).
An ongoing process, not a one-time event
This is not a one-time insight, but a recurring workflow. Complaints and churn behavior change over time, especially in competitive markets where product updates shift pain points every quarter.
Framework for implementation:
Competitor review analysis: Examine one- and two-star reviews on dominant platforms. Look for recurring complaint themes.
Your sales data: The reasons why potential clients don’t buy often coincide with the reasons why clients leave competitors.
Search query reports: Filter by competitor names, “cancel”, “alternative”, “switch from”, “problems with”. Often these queries organically appear in broad match and are mindlessly added to negative keywords.
Action algorithm:
Collect 1-2-star competitor reviews.
Identify 3-5 recurring complaint themes.
Map each complaint theme to a search query format. For example, “too expensive” → pricing queries; “terrible support” → migration/switching queries; “missing feature X” → comparison queries for that feature.
Check your search query reports. If they exist and are excluded, that’s the first fix. If not, create a campaign to test them.
For each query cluster, check the landing page. It must address a specific complaint. If the page only talks about your features, ignoring the user’s pain, the click is wasted.
Test specialized landing pages against general ones. The page should not be complicated. It should acknowledge the problem, show how you solve it, and make the next step as simple as possible.
PPC teams optimize their campaigns, keywords, and landing pages. This is logical; it’s what they control. But the biggest opportunity may not be in your account. It’s in the frustration your competitor is generating right now: price changes, product gaps, support failures, contract terms. Each such frustration creates a seeker. And this seeker is not looking for a product. They are looking for a way out.
Question for a PPC specialist:
Instead of “What should I bid on?” ask: “What problem has my competitor just created for their customers, and how quickly can I become the solution?”
Automation amplifies any offer. If the offer doesn’t match the intent, you’re scaling a miss. The keyword led to an auction. The offer closed the deal. Your competitor built an exit. Your task is to be what’s at the end of that exit.
Frequently Asked Questions
What are “churn” queries and why are they important?
“Churn” queries are search phrases expressing dissatisfaction with a current provider or a desire to switch (e.g., “cancel [competitor] subscription,” “problems with [competitor’s product]”). They are important because they indicate a user with a high degree of intent, who has already decided to leave and is looking for an alternative, making them highly receptive to your offer.
How can I find such keywords if they are not in standard tools?
They can be found by analyzing competitor customer complaints on review sites (G2, Trustpilot), social media, support forums, and churn survey data. It’s also useful to review search query reports in your campaigns, filtering them by competitor names and words related to cancellation or problems.
Do I need a special landing page for “churn” queries?
Yes, a specialized landing page is critically important. It should not just list your advantages, but directly address the problems and pains that caused the user to seek an exit from a competitor. The landing page should convince the user that the transition will be easy, you understand their situation, and you can solve their specific problem.
What CPM should be considered profitable when working with such queries?
CPM will be profitable if the customer acquisition cost (CPA) for these queries is significantly lower than for general campaigns. In the example provided, CPA was 18% lower. The main criterion is your unit economics and ROI. If these queries bring in customers cheaper and with a higher LTV, then the CPM that allows this is profitable.
Is this a one-time tactic or an ongoing process?
This is an ongoing process. Competitors’ complaints and reasons for customer churn can change over time, so it is necessary to regularly monitor this data and adapt your campaigns and offers. This allows you to keep your offers relevant and continuously attract high-intent leads.
Talks of millions from vibecoding are everywhere: a million in 17 days, a million by October, a million from one prompt. I decided to investigate: how many projects have truly crossed the threshold of $100,000 in monthly revenue or $1,000,000 in total income (including exit value). The picture turned out to be much more modest than expected. If we exclude projects whose product itself is an AI solution, very few startups based specifically on vibecoding remain.
In this analysis, I do not conflate ARR (Annual Recurring Revenue) and actual revenue. ARR is a projection of the current month multiplied by twelve, not money earned. Where a source indicates ARR, I note it and recalculate the monthly revenue separately. The order of projects in the table below is determined not by revenue size, but by the reliability of financial data confirmation.
Projects that have crossed the $100,000/month or $1 million total income threshold
The list includes projects with public reports of revenue from $100,000 per month or cumulative income from $1,000,000.
Medvi: $401 million in the first year
Telemedicine specializing in GLP-1 weight loss drugs. Founder Matthew Gallagher launched the project in September 2024 with an investment of about $20,000. According to publications, the first year brought $401 million in revenue, about 250,000 clients, and $65 million in profit with a 16% margin. Staff: two people (Gallagher himself and his brother Elliot), the rest are contractors.
Vibecoding: Code, website, and agents were written using Claude, Grok, and ChatGPT. Visuals for advertising — Midjourney and Runway. Client calls — ElevenLabs.
Business essence: The main part of the business — doctors, pharmacy, prescription writing, and legal framework — is outsourced through CareValidate, OpenLoop Health, and LegalZoom. Gallagher focused on assembly and traffic.
Base44: sold for $80 million
An application builder from text description, created by Maor Shlomo. Launched in early 2025. In May, the company showed $189,000 in monthly profit. In June, Wix announced its acquisition for $80 million in cash. At the time of the deal, the project had 250,000 users. Shlomo was the sole owner, but at the time of sale, eight people worked for the company, who received $25 million from the deal amount as retention.
Vibecoding: Shlomo deliberately built the codebase so that most of it was generated by models, and used prompt routing between different models. This is one of the purest examples where the product itself is vibecoding and built using the same method.
Confirmation: Financial performance is confirmed by the deal document, not a tweet.
Peec AI: $4M ARR (up to $833,000/month)
A Berlin-based company that tracks how a brand is represented in AI search (ChatGPT, Perplexity, Gemini). Founder Marius Meiners. The first prototype was built on v0 from Vercel in about a day and a half. The product was then developed in a traditional way by a team. In November 2025, the company announced a Series A of $21 million, with a total of $29 million raised. At that time, Peec AI had 1300+ connected companies and ARR of about $4 million (approximately $333,000 per month). Later, ARR grew to $10 million (about $833,000 per month).
Vibecoding: Used to quickly create a prototype and test demand, but the main product was built traditionally.
Wave AI: $7M ARR
An application for recording, transcribing, and retelling audio, focused on casual conversations, not meetings. Author Josh Moorer, former Uber GM for New York, not a programmer. He built the first working prototype in Swift in a day using ChatGPT. The project reached $7 million ARR with $3 million in annual profit and about 4000 paid subscribers.
Vibecoding: Prototype created with AI.
Feature: High costs for accessing the best models to ensure transcription quality, which is the key product.
Projects that did not reach the threshold, but are often mentioned
These projects often appear in selections, but their financial performance does not meet the specified threshold.
Cal AI: $2M monthly revenue, but the label is questionable
An application for counting calories from a photo of food. Authors Zach Yadegari and Henry Langmack. Launched in May 2024. By July 2025 – 8.3 million downloads and, according to CNBC Make It, about $1.4 million in gross monthly profit, and $2 million in monthly revenue is also mentioned. It passes in terms of money, but not in terms of the vibecoding label. TechCrunch revealed that Yadegari wrote the code himself, starting programming at the age of ten. Anthropic and OpenAI models are only used for image recognition within the product.
fly.pieter.com: $1M ARR, but peak $87,000/month
A browser-based flight simulator by Pieter Levels. Basic prototype in 30 minutes. $1M ARR was announced in February 2025, 17 days after release. Peak monthly revenue was around $87,000. Monetization: paid aircraft models and items for $29.99. Levels’ method: “one prompt – one layer.” $1M ARR is an annual projection, not a million earned. The project did not reach the $100,000 monthly revenue threshold.
Stanley: $1M ARR, but not revenue
An AI content tool created by Vitaly Dodonov and Jon Hu in 14 days. The LinkedIn version was released in June 2025 and showed $200,000 ARR after six weeks. The team passed the $1M ARR mark by October 2025. “14-day app” refers to the first working cut, not the full product. The main value was added after launch. Again, this is ARR, not actual revenue.
Lumoo: €700,000 ARR, about €58,000/month
A platform for fashion and retail brands by Henrik Skagerlind Fast and Peter Törngren. Offers photorealistic visualizations and virtual try-on. According to the Lovable case, in nine months the product reached €700,000 ARR, connecting more than fifteen brands. The goal is €1M ARR by November 2025. Monthly revenue is about €58,000, which is below the threshold. The project’s success is due to the founders’ industry connections, not a generated interface.
Plinq: $456,000 ARR, about $38,000/month
A background and safety check app by Sabine Matos. Built entirely on Lovable. 6,000 ARR was reported, R3 million raised, and over 10,000 users. 6,000 ARR is about $500 a month, and that’s in the best interpretation. The case is interesting because the app gathered an audience before it made money. Investments came after users appeared, not because the first version was generated by a tool.
General Conclusions and Key Success Factors
Analysis shows that project success is not limited to cheap, AI-generated code. There are several common traits among successful startups.
Team Size at Launch
Four out of five successful cases started with one or two people. This is not so much a credit to vibecoding as it is an indicator that a product can be built faster than a team can be formed.
Infrastructure on Others’ Shoulders
Successful projects actively used third-party services and APIs:
Medvi outsourced doctors, pharmacies, and lawyers.
Wave AI used third-party APIs for transcription.
Cal AI — for recognition and product database.
Lumoo integrated into existing brand workflows.
Only what was not on the market had to be built from scratch.
Money doesn’t come from cheap code
No product won solely due to cheap or generated code. Money came from other factors:
Cal AI — through influencers.
Stanley — through distribution via LinkedIn.
Medvi — through contractors with licenses.
Wave AI — through expensive inference (model computations).
Plinq — through an audience that appeared before revenue.
Code generation shortens the path to launch but is not the main source of income.
Multimodel Verification Method
Almost all teams in the selection use the trick: run one task through several models and compare the results. This is how Peec AI checked demand, Base44 routed prompts, and Levels assembled game layers, changing the model for each task. This allows for quick hypothesis testing and finding optimal solutions. You can test this approach in an evening: take your current prompt and run it through three different models to see the discrepancies and choose the best option.
Frequently Asked Questions
What CPM should be considered profitable for hypothesis testing?
For hypothesis testing, not only CPM is important, but also the speed of data acquisition. If AI tools allow for quick creative generation and campaign launches, then even with a relatively high CPM, hypothesis testing is cheap due to reduced time and labor costs. A profitable CPM is one that provides sufficient data volume for decision-making with minimal launch costs. For example, if the price per one finished clip, generated by AI, allows launching 100 variants for the same budget as 10 variants from a human, it is profitable.
What’s cheaper: banner advertising or bloggers?
A direct price comparison is difficult, as the result is different. Banner advertising (how much does 1 million impressions via banner cost) can be cheaper in terms of CPM, for example, CPM cheaper than 50 rubles per 1000 impressions. But bloggers provide native integration and audience trust. For a budget alternative to blogger seeding, you can consider a mass-posting package price which will be lower, but the reach may be less targeted. Cheaper than influencers can only be your own traffic or very targeted advertising.
Is it worth investing in AI for content creation?
If the goal is to reduce content production costs and increase speed, then is AI worth it — definitely yes. For example, the cost of generating 100 AI videos will be many times lower than creating them manually. The cost of uniqueizing one video with AI also significantly reduces costs. However, it is important to remember that price vs result is a key factor. AI speeds up the process, but a quality result still requires human control and refinement.
How to estimate the cost of AI video production?
The price per minute of AI video with an avatar can vary. It is important to consider not only generation, but also how much it costs to cut 1 hour of stream into short clips using AI. This allows for a lot of content from one source. The cost of publishing 500 videos, if they are generated or cut by AI, will be significantly lower than manual work. What is cheaper: banner or target — this is a question of effectiveness, not just price, and AI can increase the effectiveness of any channel through personalization and scale.
Conclusion
Analysis shows that “millions on vibecoding” is more of a myth than a reality, if we consider the net income from generated code. The success of projects lies in the ability to quickly assemble an MVP, use third-party infrastructure, and scale the business through effective distribution and customer work, and not just through cheap code. The unit economics of these projects are based on a comprehensive approach. Before believing in another success story, always demand statistics screenshots and ask the main question: how much money will each ruble spent bring. For those who are ready to experiment, we recommend trying different models on the same task. Use promo code CTRLAI for a 20% discount on any plan at SYNTX.AI.
CPM is profitable only when it converts into money. With the advent of AI Overviews (AIO) in Google Search Console, traditional position metrics are distorted, creating a false impression of visibility. This requires a rethinking of the approach to evaluating SEO effectiveness, focusing on real visits and conversions, rather than on numbers that are worthless.
Block Flattening: how Google distorts data
Google Search Console uses an approach called “block flattening” to display AI Overviews. This means that the entire AIO block, which often appears in the first position, is counted as a single position, and all links within it receive the same “top” position in reports.
“Tracking traditional positions for AI features is hard to make useful,” said John Mueller of Google Search.
This leads to even links hidden behind a “Show more” button receiving the same position 1 as those immediately visible. As a result, Search Console data may show high visibility, even though the user may not have noticed your site.
Why “average position” no longer works
The problem of “average position” existed even before AIO, as it is a simple mathematical average. Search is personalized: results vary depending on geolocation, device, and query. Mixing all these factors into a single average number does not reflect the real picture.
Your page might be in position 1 in AIO, but in position 19 in regular results.
Search Console might show an “average position” of 10, hiding these extreme values.
How impressions mislead
Google’s rules for counting impressions are outdated and do not account for the interactivity of modern search results. An impression is recorded as soon as a result loads on the page, even if the user does not see it. This means that:
All links in AIO are counted as impressions by default, even if the user only read the first sentence.
Links hidden behind interactive elements (e.g., “Show more”) are not counted as impressions until the user clicks on them.
This creates a paradox: invisible links get “impressions,” while useful but hidden ones are ignored.
Studies show that AIO significantly changes user behavior and affects whether CPM is profitable or not. For example:
A Seer Interactive study showed that CTR for cited sites decreased by 61%.
An SEJ field study found that AIO cut clicks on regular organic results by approximately 38%.
This happens because users get the information they need directly from AIO and do not go to third-party sites. This is especially true for informational queries, where AIO appear in 21% of cases.
Instead of chasing ephemeral positions, an arbitrageur should focus on real metrics. Testing hypotheses cheaply is possible if you understand exactly what you are measuring. If you see your brand mentioned in AIO – that’s already good. Next, you need to look at:
Actual website visits: how many users actually came to your resource.
Lead generation: how many inquiries or contacts you received.
Conversions: how many sales or target actions were performed.
These indicators are not distorted by “block flattening” and provide a clear understanding of how much money each ruble spent will bring. To control the real display of the page, use tools that simulate live viewing, rather than relying on averaged Google data.
Frequently Asked Questions
What is “block flattening” in Search Console?
“Block flattening” is a method where Google Search Console combines the entire AI Overview (AIO) block into one position, usually the first. All links within this block receive the same “top” position in reports, regardless of their actual location or visibility to the user.
Why has the average position in Search Console become less reliable?
The average position has always been an averaged metric that does not account for search personalization (geolocation, device). With the advent of AI Overviews, it has become even less reliable, as the same page can be in the first position in AIO and in the twentieth in regular results, leading to a distorted average value.
How do AI Overviews affect click-through rates?
AI Overviews can significantly reduce the number of clicks on organic results. Users often get all the necessary information directly from AIO and do not go to third-party sites, which leads to a drop in CTR for many resources.
What metrics should be used instead of positions in Search Console?
Instead of positions, you should focus on real indicators: the number of actual website visits, lead generation, conversions, and whether your brand is mentioned in AI Overviews. These metrics provide a more accurate picture of the real effectiveness of your SEO.
Conclusion
Old metrics don’t work. Price vs. result — that’s what matters. Forget about the illusions of “average position” and “impressions” that don’t bring cash. Focus on real visits, leads, and sales. Only then can you understand if your CPM is profitable and if AI is worth the effort. Analyze unit economics, not reports that Google itself admits are inaccurate. Switch to a new strategy for evaluating effectiveness, based on actual data, not on numbers that are worthless.
AI in FinTech: Who Leads the Search and What Does It Cost?
A study of 68,334 responses from five different AI platforms revealed which FinTech brands dominate the artificial intelligence sphere. If you want to understand how CPM is profitable in the new search reality, this information is critical.
Traditional SEO methods do not guarantee visibility in AI, which means many companies are missing out on the opportunity to get traffic where consumers are looking for answers to financial questions. We analyzed 250 queries across 10 categories of consumer FinTech products to determine the real picture.
According to Plaid, over 200 million people ask questions about personal finance on ChatGPT alone every month. Add Google AI Overviews and Claude to that, and the volume of AI-generated solutions becomes colossal.
A LendingTree survey showed that almost half (49%) of chatbot users stated that AI influenced at least one of their financial decisions, most often related to budgeting, taxes, or opening/closing accounts and loans.
“Success in traditional search does not mean automatic visibility in AI. You can dominate Google for key queries and still not appear in ChatGPT’s answers to similar questions. The size of this gap is striking,” experts note.
Out of 68,334 responses, 7,317 unique brands were mentioned. The maximum AI Visibility Score (a 100-point scale that considers frequency of appearance, position of mention, and number of categories) was 65.6.
Among pure FinTech brands, none exceeded the 65 mark. This indicates that the market is fragmented, and newcomers have a chance.
Methodology: How We Calculated the AI Visibility Score
10 categories: selected for industry relevance.
Mention Position: determined by the first appearance in the response text.
Sentiment: evaluated by the surrounding context of keywords. These are text proxies, not manual labeling.
AI Visibility Leaders: Who and Why?
Apple leads with 65.6 points. Although not a pure FinTech brand (Apple Pay, Apple Card), it sets the bar. Among pure FinTech companies, SoFi leads with 64.6 points.
The reason is simple: SoFi is the only brand that appeared in all 10 FinTech categories, including banking, lending, investments, and money transfers. Mention frequency accounts for 50% of the total score, and category coverage accounts for 20%. No other FinTech brand came close to SoFi in this metric.
Next are:
PayPal: 59 points (3rd place).
Charles Schwab: 53.6 points.
Chime: 53.0 points.
Revolut: 52.6 points.
Fidelity and Wise performed well despite fewer mentions, due to their early appearance in responses (on average, 3rd position). This shows that a cheap hypothesis test on the impact of mention position on conversion is well justified.
Wealthfront Example: Breadth Over Volume
Wealthfront, with 46 points, has many mentions but covers only 7 out of 10 categories. This gap in coverage cost it positions compared to brands with fewer mentions but broader category presence.
This demonstrates that CPM is profitable not only due to quantity but also due to the quality of coverage.
Category Leaders: Who Dominates the Niches?
Each of the 10 fintech categories has its own AI visibility landscape. The category leader is often not the leader of the overall index. Brand concentration varies greatly.
Buy Now, Pay Later (BNPL): the most consolidated. Klarna (91.8%), Affirm (83.1%), Afterpay (80%). Klarna is almost always mentioned first.
Money Transfers:Wise leads with 84.7% of mentions, maintaining stability across all five AI platforms. Remitly (66.8%) and Western Union (55.2%) lag behind. Wise ranks first in 57.7% of responses where it appears. This is an excellent example of a challenger brand outperforming a veteran.
Investments:Fidelity (74.5%) and Charles Schwab (73.8%) are co-leaders, AI does not favor one over the other. Robinhood (56.3%) is in third place. Smaller brands with strong content comparing features compete with giants.
Robo-advisors:Betterment (84.2%) and Wealthfront (75.7%) are a clear pair. Betterment is most often mentioned first.
Banking Services:SoFi (59%) and Chime (54.2%) lead. Capital One (40.3%) and Ally Bank (40.2%) share third place.
Cryptocurrencies:Coinbase dominates with 72.9% of mentions. Kraken (64.9%) and Binance (52.5%) follow.
Budgeting:YNAB leads with 72.8% of mentions and is almost always the first budgeting tool named by AI. Monarch Money (55.5%) is second.
Lending: the most fragmented category. SoFi (43.7%) leads, but the gap from Upstart (36.5%) and LightStream (33.1%) is minimal. Here, a cheap hypothesis test can yield quick results.
Business Software:Intuit leads with 58.3% thanks to QuickBooks and TurboTax. Xero (41.2%) is second.
Payments:PayPal (81.5%) is the most mentioned brand, but rarely first (average 4.46 position). Venmo (64.2%), Cash App (48.1%), and Zelle (46.4%) follow. Apple Pay, despite Apple’s overall high score, ranks fifth in this category.
Tone of Mentions and Influence of AI Platforms
Less than 2% of fintech brand mentions have a negative connotation. AI responses are most often neutral or positive. Leaders in positive tone are business finance tools: FreshBooks, QuickBooks, Wave (over 75% positive mentions).
This is because AI recommends them in the context of problem-solving, using phrases like “QuickBooks simplifies…”.
Wise (69% positive) and Capital One (68.6% positive) stand out with a high level of favorable descriptions (“reliable”, “transparent”, “well-established”). AI takes these formulations from third-party sources, such as lists of best products in financial publications.
On the other hand, Cash App and Wells Fargo have the highest share of negative mentions (1.7% and 1.5% respectively), related to fees and issues arising in the context of complaints.
Strategic Conclusion: Tone is shaped by the language of the sources AI uses and the context of the query. If your brand appears more often in the context of complaints, this will be reflected in the AI’s description. If it appears in instructions and reviews, this will also have an impact. That is, the tone reflects a mix of your sources, not just reputation.
Differences Between AI Platforms
The five AI systems differ:
Number of brands in the response.
Preferences for specific brands.
Sources of information.
ChatGPT more often mentions PayPal (25.3%) and SoFi (20.5%). Claude Sonnet 4, in contrast, is more selective. NerdWallet showed an extreme difference: 11.5% in Google AI Overviews versus 1.3% in Gemini.
Gemini provides the most dense answers (average 9.8 brands), while Perplexity and Claude provide only 6. Fewer brands in the answer mean less competition, which is important for budget alternative to seeding.
Reddit is mentioned in 12% of ChatGPT responses and 7.8% of Google AI Overviews, but less than 0.2% of Perplexity and Claude. This indicates that AI platforms have different data sources, which affects price comparison of promotion.
Frequently Asked Questions
What is the AI Visibility Score?
This is a 100-point metric that measures how often a brand appears in AI responses, how early it is mentioned, and in how many categories it is present. Essentially, it is an indicator of how much money will be generated for every ruble spent in the AI channel.
What is a good CPM in AI search?
Since AI search is only developing, there is no direct CPM yet. However, if a brand consistently appears in the first positions of AI answers, this is equivalent to a very low CPM, as the reach is organic and highly convertible.
Focusing on category coverage and early mention is key to a CPM cheaper than 50 rubles in terms of real contacts.
Is it worth investing in AI search optimization?
Yes, absolutely. AI is becoming the main channel for finding financial information. Companies that ignore this risk losing a significant portion of potential customers.
This is not a question of “how much does it cost to cut 1 hour of stream”, but a question of whether AI is worth it as a channel for attracting traffic. The answer: yes, it is, and the sooner the better.
What is cheaper: banner or target in AI search?
AI search is inherently closer to organic traffic. If your brand is mentioned by AI, it’s free, unlike paid banners or targeted advertising.
Investing in AI optimization is a long-term strategy for obtaining “free” traffic, which ultimately turns out to be cheaper than influencers and traditional advertising.
Conclusion
An analysis of 68,334 AI platform responses showed: the fintech market in AI search is fragmented, and companies have a chance to take leading positions. Apple and SoFi are leading, but their scores (maximum 65.6) show that the ceiling is far off.
This creates huge opportunities for testing hypotheses cheaply and for companies that focus on the breadth of category coverage and content optimization for AI. It should not be forgotten that a favorable CPM in AI search is achieved not through direct placement, but through deep integration into the AI information field.
The Whatnot platform has announced a new commission structure that could significantly impact sellers’ unit economics. The essence is simple: the more you sell, the less you pay. This is not charity, but a clear calculation aimed at retaining and motivating large players. Understanding what a profitable CPM is becomes key for sellers on this platform.
New Whatnot rates: counting the cash
Now, sellers who achieve monthly sales of $15,000 receive a commission reduction. The rate will continue to drop if sales exceed $250,000. For top sellers, Whatnot claims a possible commission reduction to 3%. This is presented as “the lowest commissions in the industry.”
“Our sellers made this possible. The businesses you’ve built have brought more buyers, products, and communities to Whatnot,” Whatnot said in a blog post. — “As the marketplace expands, we’ve found new ways to support what sellers are building here. Lower commissions, built into how Whatnot works, rather than offered as a temporary promotion, are the next step.”
How does it work in practice?
Sales from $15,000 per month: commission reduction.
Sales from $250,000 per month: further reduction.
Top sellers: commission up to 3%.
A profitable CPM for large players is obvious. The higher the turnover, the lower the percentage of deductions, which directly increases profitability. This is a classic example where a platform stimulates growth by offering more favorable conditions for those who already generate significant revenue.
Whatnot: facts and figures
Whatnot positions itself as a pioneer in e-commerce, combining second-hand sales and live streaming. Facts:
In 2025, the platform generated $8 billion in Gross Merchandise Value (GMV).
Attracts 650,000 new users weekly.
Raised $545 million in a Series G round at a valuation of $20 billion.
These figures show that the platform has a solid margin of safety and can afford to invest in seller loyalty by reducing commissions. The goal is not altruism, but strategic volume growth and retention of market leaders. This is a kind of budget alternative to seeding for the platform, motivating sellers to independently increase traffic and sales.
Seller motivation: Arkollab case study
As an example, Whatnot cites the case of seller Arkollab, who participated in the pilot program of the new rates and doubled their monthly sales. This confirms that commission reductions work as a strong incentive. For the platform, this is a way to provide a cheap hypothesis test and confirm the effectiveness of the new model.
Not just for whales: a chance for “small” players
Although the new terms primarily benefit large sellers, Whatnot states that almost 20% of sellers who achieved $15,000 in sales over four weeks were selling less than $5,000 during the same period six months ago. This indicates growth potential for everyone. If your CPM is cheaper than 50 rubles and you show growth, then this could be your path to more favorable conditions.
Frequently Asked Questions
What CPM should be considered profitable?
A profitable CPM is one that ensures positive unit economics. In Whatnot’s case, a commission reduction to 3% means that for the seller, every dollar of sales brings more net profit, making the CPM profitable at high volume.
Is AI worth it, and what’s cheaper: a banner or a blogger?
The question of the cost of uniqueizing one video or generating 100 AI videos is not directly related to Whatnot’s commissions, but it is important for a general understanding of content costs. In the context of Whatnot, where live content dominates, investments in producing quality streams may be more effective than mass production of AI videos for other platforms. What’s cheaper, a banner or a blogger? On Whatnot, the blogger is the seller, and their effectiveness directly affects the commission size. The commission reduction makes the “blogger” (seller) on the platform more profitable.
How will the new rates affect the cost of publishing 500 videos?
If we are talking about publishing 500 streams or short videos on Whatnot, then upon reaching high sales volumes, the cost of publishing 500 videos will indirectly decrease due to the reduction in the commission percentage from each sale. This makes the platform more attractive for those who are ready for intensive work.
Conclusion: cash matters
Whatnot’s new commission structure is not just about discounts, but a clear calculation for business scaling. For sellers, this is a direct incentive to increase volumes to get low commissions and increase net profit. For the platform, this is a way to retain and motivate key players, ensuring stable GMV growth. If you work on Whatnot, it’s time to recalculate your unit economics and understand how the new terms can increase your cash. Start streaming more to get your profitable CPM.
Generative AI promises a revolution, but the real cost of content production often remains hidden. We calculated how much it costs to create a 25-second commercial using neural networks, including script, images, video, voiceover, and music. Spoiler: the 25-second video cost 2,126 ₽, or 85 ₽ per second. The lion’s share of the budget, almost 90%, was “eaten up” by video generation.
How much does AI advertising cost: a 25-second video for 2,126 ₽
In a world where technology is developing at an incredible pace, artificial intelligence is becoming an integral part of many industries, including content creation. But what is the real price of this innovation? We conducted an experiment to find out how much it costs to create a 25-second commercial entirely with neural networks. The results may surprise you, because the cost of unique content for one video and profitable CPM for advertising materials are becoming key factors.
The experiment was conducted through BotHub using the following models:
We advertised a herbarium, without commercial implications. Important: only the marginal cost was taken into account — the cost of successful generations included in the final video. Actual costs may vary depending on prompting skills and perfectionism level, increasing the number of regenerations and, accordingly, the budget.
Methodology for calculating the cost of AI production
For an objective assessment, it is important to distinguish between two types of costs:
Cost of result generation: direct computational costs for content creation.
Cost of infrastructure access: costs for subscriptions or minimum credit packages for using services.
In the case of a single video, the second amount can significantly exceed the first. Of course, open-source models can be deployed locally, but this requires a powerful PC, which is not always available.
«Don’t replace professionals with browser tabs. AI content has nuances: rights, licenses, character stability, result control, and brand requirements. AI slop is easy to get, and you’ll get flamed for it».
We started by creating a concept. GPT-6 Astra acted as a creative producer and scriptwriter. The prompt was formulated so that the model immediately took into account the limitations of AI generation and formed prompts for each element.
Prompt to GPT-6 Astra:
«Role: You are an experienced creative producer and scriptwriter specializing in creating advertisements using neural networks. Task: Develop a detailed script for a 25-second commercial for a Herbarium (an album with a collection of dried plants/herbs). Important condition: All content (video, photo, audio) will be AI-generated. Therefore, your task is not just to provide a script with a description of the picture, but a script supplemented with prompts for each element. Atmosphere: warm, autumn, cozy».
The script development and adjustments took ₽102.09.
Step 2: Generating images for the ad
Nano Banana 2 was used to create eight base images that set the composition, atmosphere, and style. The total cost of this stage was ₽128.29, which is an average of ₽16 per image.
The images served as an intermediate state: they fixed the spatial structure of the frame, which simplifies the task for the video model. This allows for consistency between scenes, which is critical for a commercial.
Image-to-video: precise control over the result
Unlike text-to-video, where the model simultaneously interprets the description, chooses the composition, and adds movement, image-to-video fixes most variables. This gives greater control over the result and ensures stylistic unity across all frames, which reduces the cost of uniqueizing one video and provides a favorable CPM for advertising materials.
Step 3: Video generation: the most expensive part of AI advertising
Converting static images into video scenes is the most resource-intensive stage. Seven video scenes were generated using Kling v3 Video, preserving the original composition and adding natural dynamics.
The cost of one video scene was about ₽268, and the total cost of video generation was ₽1,878.17. This is 88% of the total budget. The high cost is due to the need to ensure temporal consistency: objects, camera, lighting, and scene geometry must be stable throughout the video.
Step 4: Ad voiceover with ElevenLabs: cheap and tasteful
ElevenLabs Flash v3 was used for the voiceover. This stage turned out to be one of the cheapest — only 6.04 ₽. TTS technologies significantly simplify the process compared to traditional voice actor recording, providing fast generation and low costs. This is an excellent option for cheap hypothesis testing and prototypes.
Step 5: Music for advertising in Suno: budget and limitations
The musical accompaniment was created using Suno v5.5. The model allows describing the mood and genre, rather than specific notes, which is convenient for advertising tasks. The cost of music generation was 11.82 ₽.
Although AI music saves budget, full control over the result is not guaranteed. However, for small projects, it is an effective alternative to stock tracks or ordering original music.
Total cost and comparison with direct subscriptions
The total cost of a 25-second advertising video created with AI was 2,126.41 ₽.
Tool
Cost
Script and prompts (GPT-6 Astra)
102.09 ₽
Base images (Nano Banana 2)
128.29 ₽
Video generation (Kling v3 Video)
1,878.17 ₽
Voiceover (ElevenLabs Flash v3)
6.04 ₽
Music (Suno v5.5)
11.82 ₽
TOTAL
2,126.41 ₽
If access to each service were paid for separately, minimum monthly subscriptions would be around 3,591 ₽ (at an exchange rate of $1 = 86.4 ₽):
ChatGPT (Plus): $20 (~1,729 ₽)
Nano Banana 2: ~$0.54 (~47 ₽)
Kling v3 Video: $6.99 (~604 ₽)
ElevenLabs Flash: $6 (~519 ₽)
Suno v5.5: $8 (~692 ₽)
These 3,591 ₽ are more of a starting capital than the cost of a single video, as tariffs usually include many more credits than required for one project. This highlights the benefit of using aggregators like BotHub for one-off tasks or cheap hypothesis testing.
Conclusions and prospects for AI production
AI production is not yet a seamless process. Each model has its own features and data formats. Cohesion between stages is ensured by an external workflow: prompts, references, manual editing. This distinguishes the demonstration of individual model capabilities from a full-fledged production system, where all elements must work in unison. The price per finished clip with AI can be significantly lower than traditional production, but requires competent pipeline management.
Overall, AI tools significantly lower the barrier to entry for video content creation, offering a budget-friendly alternative to seeding and traditional production studios. However, obtaining a high-quality result requires skills and an understanding of each model’s limitations. This makes CPM profitable for testing various advertising creatives.
Frequently asked questions
What CPM should be considered profitable for AI advertising?
A profitable CPM depends on the niche and campaign goals. If the cost of producing a video is low, as in our case, then even with an average CPM, the overall return on investment (ROI) can be high. The main advantage is the ability to quickly and cheaply test hypotheses.
What is cheaper: a banner, a blogger, or an AI video?
A direct comparison is incorrect, as these are different formats with different goals. However, AI video offers cheap hypothesis testing and can be significantly cheaper than producing a video with a blogger or even some types of banners, especially if complex animations are required. The cost of creating one AI video at 2,126 ₽ makes it competitive for mass content production.
Is it worth investing in AI for content production?
Yes, it is. AI significantly reduces the time and cost of production, allowing for quick creation and testing of advertising materials. This is especially relevant for small and medium-sized businesses, as well as for arbitragers who value the cash and unit economics of each creative. However, it is necessary to consider that the price vs. result strongly depends on the skills of working with prompts and post-processing.
How much does it cost to cut 1 hour of stream using AI?
The cost of cutting an hour of stream using AI will depend on the services used and their rates for video processing. If it’s about creating short clips, it can be significantly cheaper than manual work and comparable to the price of generating one video, as the main costs will be for transcription and content analysis, not for generating video from scratch.
What is the price per minute of AI video with an avatar?
The price per minute of AI video with an avatar varies depending on the platform and the complexity of the avatar (2D, 3D, customization). In our experiment, 25 seconds of video cost 1,878 ₽, which is equivalent to approximately 4,500 ₽ per minute if only the video generation cost is extrapolated. However, avatar platforms have their own rates, which can be both higher and lower, depending on the features and quality provided.
AI advertising opens up new horizons for content creation, making it more accessible and faster. If you want to delve deeper into the world of AI and learn how it can transform your business, visit our blog. Discover new opportunities and start creating unique content today!
In the world of AI search, leaders are constantly changing, and the question of which platforms to track becomes critical for SEO specialists. Recent discussions have revolved around the Perplexity platform: should we continue monitoring it or focus exclusively on giants like ChatGPT, Gemini, and Claude? We will examine how changes in market share affect visibility assessment and how much money each spent ruble will bring to understand if CPM is profitable.
Is Perplexity really losing ground?
Ross Hudgens, CEO of Siege Media, claims that Perplexity’s market share is shrinking. Including it in general metrics alongside ChatGPT, Gemini, Claude, and Google AI products can distort the real picture of visibility. Hudgens urges: “Everyone should remove Perplexity from their LLM trackers today.”
However, as experience shows, hasty conclusions can be erroneous. In March 2002, at the Search Engine Strategies conference, the reduction of tracked search engines from 15 to 5 was discussed. Google was then excluded from the list. This turned out to be a serious oversight, as Google was rapidly gaining momentum. A similar situation could repeat itself today.
Data confirms Perplexity’s decline
According to StatCounter, in June 2026, Perplexity held 7.91% of AI chatbot referral share, almost on par with Gemini (7.94%).
By August 2026, Perplexity’s share had fallen to 4.31%, while Gemini grew to 10.9%.
“A small market share does not always mean small strategic importance. It’s like investing in stocks: you shouldn’t always look only at current capitalization, but also at growth potential.”
The market is consolidating, but not around a single leader
The latest Similarweb data for May 2026 shows the dominance of ChatGPT (53.9% of global web visits among seven major AI assistants). Gemini accounts for 27.9%, Claude – 9.2%, DeepSeek – 4.1%, Grok – 2.4%, and Perplexity and Copilot – 1.3% each.
Significant changes occurred not in the disappearance of Perplexity, but in the emergence of a strong second player. ChatGPT’s share of AI chatbot web traffic decreased from 76.4% a year ago to 52.7% in May 2026. Gemini grew from 9% to 27.3%, and Claude – from 1.6% to 8.9%.
Key players and their advantages
ChatGPT: over 900 million weekly active users (February), over 1 billion active users across all products (end of July). This is a huge direct consumer base.
Gemini: The Gemini app exceeded 1 billion monthly users (August). It has Google distribution through Search, Android, and the entire product ecosystem.
Claude: Has gained a strong position in the enterprise segment and among developers. In April, over 100,000 customers used Claude on Amazon Bedrock. Annual revenue exceeded $65 billion (August).
These are not three versions of the same business, but three different distribution advantages. The market looks like an emerging oligopoly, not a “winner takes all” situation.
What SEOs Should Track
Google AI Overviews and AI Mode should not be viewed as ordinary LLMs. They are much more important. In June, Google reported that AI Overviews cover more than 2.5 billion users per month, and AI Mode exceeded 1 billion monthly users. AI Mode queries have doubled every quarter since launch.
According to Similarweb, AI Overviews appeared in 43% of Google search queries in the US by May 2026, compared to 15% a year earlier. AI Mode visits grew from 126 million in June 2025 to 279 million in May 2026.
This means that Google AI is a separate layer in the world’s dominant search ecosystem. For B2B companies, Claude is also of immense importance, despite relatively low consumer traffic. PwC, TCS, and Cognizant are actively integrating Claude into their workflows, training tens of thousands of employees. Over 1,000 business customers spend more than $1 million annually on Claude.
Three-Tier AI Visibility Measurement System
First Tier: Platforms with scale and strategic importance. Track ChatGPT and Gemini separately. Add Claude for B2B, corporate, and professional audiences.
Second Tier: AI search embedded in existing ecosystems. Track Google AI Overviews and AI Mode separately from ChatGPT, Gemini, and Claude. The goal is to measure how AI changes the search journey. Microsoft Copilot can also be included in this tier if the company has a significant presence in Microsoft 365.
Third Tier: Emerging or specialized platforms. This includes Perplexity, Grok, DeepSeek. Do not ignore them, but do not give them equal weight. Monitor them for unusual visibility, referral traffic, citations, and growth.
Don’t Let Averages Mask Real Data
An aggregated LLM visibility metric can create a false sense of accuracy. For example, if a brand has 40% citation rate in ChatGPT, 35% in Gemini, 30% in Claude, and 90% in Perplexity, a simple average would be 48.75%. This number says little if Perplexity accounts for only a small fraction of the traffic important to the business.
Instead of asking, “What is our average LLM visibility?” you should ask, “Where do the people important to this business actually encounter our brand?”
Visibility: mentions, citations, referring URLs, queries that generate them.
Business Impact: linking AI referrals to engagement, leads, sales, and other conversions.
My Verdict on Perplexity
Ross Hudgens is right about the problem, but not entirely right about the solution. Don’t remove Perplexity from your LLM tracker, but reduce its weight. If Perplexity generates 1% of your AI traffic, giving it 25% or 33% weight in your visibility score is unjustified. However, if it accounts for more than 5% of referrals, as Hudgens himself suggests, the argument for it becomes stronger.
Perplexity’s trajectory is worth watching. Similarweb’s August 2026 analysis still describes Perplexity as an active player in the AI search market, noting its ad-free strategy and focus on subscriptions and enterprise deals. This shows why “small” and “insignificant” are not synonyms.
Frequently Asked Questions
What CPM should be considered profitable for AI platforms?
Profitable CPM in AI platforms depends on your target audience and ROI. If a platform generates high-quality leads or sales, even with a relatively high CPM, it can be profitable. The main thing is to consider not only the cost per impression but also the conversion rate.
Is it worth testing hypotheses cheaply on new AI platforms?
Yes, cheap hypothesis testing on new AI platforms is essential. This is how you can discover the next Google. You shouldn’t spend huge budgets on them, but monitoring and minimal testing will allow you not to miss a potential breakthrough.
What is cheaper: a banner or a blogger, when compared to AI platforms?
The question of what is cheaper, a banner or a blogger, is incorrect in the context of AI. AI platforms offer a different type of interaction. Price comparison should be done through efficiency metrics: how much does 1 million impressions via a banner cost versus the cost of generating 100 AI videos or the cost of publishing 500 videos via AI tools. Each channel has its own unit economics.
How to estimate the cost of uniqueizing one video using AI?
The cost of uniqueizing one video using AI depends on the complexity of the task and the tool used. It is important to evaluate not only direct costs but also the time that AI will save. The price per minute of AI video with an avatar also varies, but it often turns out to be more profitable than manual labor, especially for large volumes.
Conclusion
The AI search market is dynamic. Ignoring Perplexity or other niche platforms entirely means risking missing important signals. However, blindly giving them equal weight with giants distorts the picture. A mathematical approach to analysis, weighing platforms by their real business impact, and constant monitoring of new players are key to success in SEO and a profitable CPM strategy.
We count cash, not just the number of platforms in the tracker. Don’t overlook the outsiders who are gaining momentum. Perhaps among them lies the next market leader. Start analyzing your data today so you don’t miss new opportunities!
