Businesses with a short shelf life · as of 15 Oct 2026

Window of opportunity

The big AI players are taking the market segment by segment. There is a repeatable pattern, the median window is a little over two years, and the first bankruptcies have arrived. Below: who disappeared, which areas were taken over, how long it takes to close a gap, who is holding out — and who may be next.

Where the traffic is going

The leaders are eating the rest of the market

Share of web traffic of generative AI tools (Similarweb, websites only, no apps). Over the year ChatGPT gave up 15 percentage points, but Gemini and Claude took them. Perplexity, the only independent player at the top, fell from 2.5% to 0.9%.

Case study

Character.AI: from a $1B valuation to bankruptcy (Chapter 11)

On 12 October 2026 Character Technologies filed for Chapter 11 in a Delaware court. It is the first unicorn of the ChatGPT era to go bankrupt, and the clearest example of a window that closed.

Numbers from the filing and reports

Chapter 11 · 12 Oct 2026
28 → 12MMAU: peak in mid-2024 and the figure in the filing
120 minaverage daily time per user
$23.6Mrevenue in fiscal 2025 ($7.4M a year earlier)
$17.6Maverage monthly cash burn in FY2025 ($8.2M in Aug 2026)
194 → 151Mwebsite visits: Jan 2026 and Sep 2026 (Similarweb)
$20MDisney's offer for a non-exclusive licence (vs $2.7B from Google in 2024)
Timeline
  1. Nov 2021Noam Shazeer and Daniel De Freitas (ex-Google, Shazeer a co-author of the Transformer) found the company. Own models and a character platform.
  2. 2022–23Over 2 billion messages in 5 months, more than 2 hours a day on average. $150M at a valuation of about $1B (a16z).
  3. May 2023Mobile apps: 1.7 million installs in the first week.
  4. mid-2024Peak: about 28M MAU. Meta and xAI court a partnership.
  5. Aug 2024Google pays $2.7B for a licence; the founders and the model team go back to Google. The company stops training its own models.
  6. Oct 2024First lawsuit over a teenager's death. Then investigations by Texas and the FTC, and a warning from 44 attorneys general.
  7. Jun 2025Karandeep Anand (ex-Meta) becomes CEO. AI video, a social feed, micro-dramas.
  8. Nov 2025Open-ended chat banned for under-18s, age verification.
  9. Jan–May 2026Settlements with families (together with Google), lawsuits from Kentucky and Pennsylvania, full-screen ads and a paid currency, "Charms".
  10. Sep 2026Anand and part of the team move to Disney.
  11. 12 Oct 2026Chapter 11. Bids due 20 Nov, sale hearing 24 Nov. Without a buyer the platform will shut down.
Diagnosis: four forces at once
1 · The gap closedThe general model got a personality, memory and a voice

In 2022 ChatGPT was a dry assistant. Since 2024 it has memory, Advanced Voice and selectable personalities, and Grok has companions. "Talking to a character" stopped being something you couldn't get anywhere else.

2 · Brain drainThe Google licence took the people who built the models

After Aug 2024 the company used other people's models and competed with their owners. Investors got their money back; the product lost its technological edge.

3 · EconomicsTwo hours of chat a day is a GPU bill, not revenue

The more users loved the product, the more they cost. $23.6M of revenue against more than $200M of cash burned a year in 2025. Most didn't pay.

4 · LawThe most engaged users were teenagers

Lawsuits and the ban for minors cut off a large part of the core. The filing lists only 12M MAU, against about 20M reported earlier.

Lesson: engagement alone isn't enough when the feature lands in a free chat, every minute costs compute, and the team that held the advantage has been sold. Earlier I described Character.AI as "hollowed out". Today it's a formal collapse.

Under fire

Less traffic doesn't always mean less revenue — but it often does

Three companies with the same problem. Website traffic is falling from the peak for each of them, but only for one (Character.AI) did the money follow the traffic.

How long the window lasted

From creating a market to a giant taking it over

Each bar shows the time from the start of the category to the launch of the native feature or the collapse of the leader. Click a name to jump to its description.

the window in which the product made sensefeature built into ChatGPT / Claude / Gemini

Cases

Who plugged the gap and what's left of it

The pattern

Why some dropped out and others survived

They sold a missing feature, not an outcomeA description of a gap in the model, not a customer problem

"Chat with a PDF", "ChatGPT with internet", "talk to a character". When the gap disappeared, so did the reason to pay.

The window was shorter than a funding roundAn 18-month runway for a 6-month niche

Browsing plugins lived for half a year, RAG wrappers for less than a year. A company with 18 months of cash was lost from the start.

Traffic leaves first, money laterPerplexity: −58% traffic, revenue grew

Perplexity lost about 60% of visits from the peak, yet revenue grew thanks to agents and business customers. Survival depends on whether the company manages to change course in time.

Selling the team kills the rest of the companyCharacter.AI, Inflection, Adept, Contextual AI

After a licence-and-hire deal investors get their money back, but the company is left without the people who created the advantage.

Those who went deeper survivedCustomer data, the customer's process, being the default choice

Gamma is profitable and works as a connector in Claude and ChatGPT. Harvey has the law firm's data, Cursor has the developer's workflow.

A small, profitable niche is also an outcomeChatPDF (~$440k ARR), PDF.ai (~$590k), Photo AI (~$130k MRR)

As a one-person business it works. As a startup aiming for a billion-dollar valuation — it doesn't.

Gaps that are closing now

  • Generating simple apps and websites: artifacts, Claude Design, apps in ChatGPT and Google AI Studio. Lovable's traffic fell 46% in Q3 2025, v0's by 64%.
  • Search with citations: Perplexity is already escaping into agents. Smaller "AI search" players have nowhere to go.
  • Meeting notes: record mode in ChatGPT, Gemini in Meet and Copilot in Teams versus separate note-taking bots.
  • Simple automations: AgentKit, Claude Cowork and MCP connectors versus "if X, then Y" workflows.
  • AI tutoring: study mode in ChatGPT (Jul 2025) versus "AI tutor" apps without their own content.
Counter-examples

Who is holding out — and how

Companies that had the same problem as the victims but survived or are growing. As of Oct 2026.

Canva — $4B ARR (2025) It went into the chats instead of fighting them

Claude Design and ChatGPT canvas generate layouts from a description — a threat to Canva. The response: a connector in ChatGPT and Claude, 26 million conversations via the ChatGPT app by Oct 2025, and the B2B segment grew to $500M ARR (+100% YoY). The CEO describes the company as "an AI platform that happens to have design tools". ARR grew from $2.8B at the end of 2024 to $4B at the end of 2025.

ElevenLabs — $350M ARR (2025) It went deeper: voice cloning, dubbing, voice agents for businesses

Since Sep 2024 Advanced Voice in ChatGPT and Gemini has done consumer speech synthesis for free. ElevenLabs isn't trying to win that segment: it went into video dubbing, voice agents for customer support and an enterprise API. ARR grew from $120M (end of 2024) to $350M (end of 2025) — +175% YoY. In May 2026 it announced passing $500M ARR.

Zapier — about $310M ARR (2023/24) Instead of dying from MCP, it became an MCP server

AI agents with MCP can replace simple "if X, then Y" workflows — a threat to Zapier's core. The response: its own MCP server giving agents access to 8,000+ apps and 30,000+ actions. Instead of fighting AI agents, Zapier became the action layer that agents call. Profitable since 2014, without VC.

Gamma — $100M ARR (Nov 2025) It went into the chats as a connector before decks in chat could kill it

ChatGPT agent has made decks since Jul 2025, Copilot in PowerPoint for a long time. Gamma responded by entering ChatGPT (Mar 2026) and Claude (Jan 2026) as an app. Website traffic is falling (−26% in the 3 months to Aug 2026), but ARR holds thanks to the connectors — it makes decks from within the chat, not instead of it.

In the crosshairs

Who may be next

Companies whose window is closing right now. Data as of Oct 2026.

Midjourney — about $500M ARR (2025) Profitable, but no API and no mobile app

No public API (the terms of service forbid automation), no mobile app (as of May 2026), access mainly via Discord and a web app from 2024. With $500M ARR from subscriptions alone the company is profitable — but without an API it won't get into any business product. When GPT-4o and Flux make images good enough for 90% of needs, a stylistic edge doesn't protect the revenue structure.

NerdWallet and SEO-dependent publishers $837M in NerdWallet revenue (2025), but organic traffic "pretty brutal"

CEO Tim Chen put it plainly: "consumers are increasingly turning to AI overviews and LLMs instead of traditional search". Credit cards: −24% YoY. SMB products: −12%. Ziff Davis: −18% revenue. NAHL Group (legal leads): organic traffic −25–60%. A company that sold answers to financial questions is losing those questions to a free chat.

Duolingo — $748M revenue (2024), Q1 2025 +38% YoY Growing, but the stock reacts to every GPT language demo

Study mode in ChatGPT (Jul 2025) and voice conversation with corrections do the same for free. 90% of users don't pay. During the GPT-5 demo in Aug 2025 Duolingo's share price fell noticeably — the market signals risk despite good current results. The threat is long-term, not immediate.

Meeting transcription — Otter.ai about $100M ARR The category is commoditising — Teams, Meet, Zoom and ChatGPT do it natively

Otter and Fireflies are growing, but their core — record, transcribe, summarise — is now a native feature in Copilot for Teams, Gemini in Meet, Zoom AI Companion and ChatGPT (Jun 2025). A separate app with a separate link does what will shortly be built into the tool you already use. Fireflies reached a $1B valuation in Jun 2025 — will they manage to go deeper in time?

Video generation — Runway, Pika, Luma Runway raised $237M at a $1.5B valuation. Sora launched in Feb 2025.

Runway was the first to show commercial AI video generation (Gen-1, Gen-2). A $1.5B valuation in 2023 looked reasonable. Then OpenAI released Sora (Feb 2025), Google released Veo 2 (Dec 2024), and China's Kling was free. In a year the category went from "nobody but Runway does this" to "every platform has its own model". Pika Labs ($80M), Luma AI ($43M) and Runway now compete in the same segment as Sora and Veo — that is, with companies that have their own data, their own models and their own distribution.

Content-based SEO tools — Surfer SEO, Jasper and similar Meta-irony: they sell a fix for a problem that AI is making unsolvable

Surfer SEO, Jasper and dozens of "write for SEO with AI" tools flourished because ranking in Google still worked. NerdWallet is losing 24% of credit card revenue to AI Overviews, and NAHL Group is losing 25–60% of legal leads — their customers are exactly the ones Surfer and Jasper helped rank. An SEO content tool whose customer loses traffic despite ranking will lose its reason to pay. The window is different — not "a feature built into the chat" but "the platform took the end result".

The long tail

Hobby categories — closed wholesale by a single launch

"There's An AI For That" lists 15,000+ tools. The most-saved list is dominated by niche image generators and simple editors. None is a big business, but together they are tens of thousands of active tools with millions of visits. GPT-4o (Mar 2025) closed most of these windows at once — not by building a competitor, but by adding image generation to the chat users already had open.

Categories that disappeared or are losing their reason to exist

  • Niche image generators: Minecraft skins, Pokémon-style characters (Fakemon), sports jerseys, lingerie designs, Ghibli style, tattoo designs, fantasy creatures. Each category had dozens of tools and hundreds of thousands of visits a month. GPT-4o makes the same images from a description in the chat — no separate site, no sign-up, no limit on styles.
  • Background removal and upscaling: remove.bg led for years. Today removing a background is one command in ChatGPT, Canva or even Apple Photos. Upscaling is built into Lightroom, Topaz and every photo tool. A separate service makes little sense when it's one mode in a tool you have open anyway.
  • Logo and branding generators for non-designers: Looka, Brandmark, Logo.com and hundreds of clones were built on the fact that Canva takes time and an agency costs money. Now it's one prompt in ChatGPT or Claude. They didn't displace professional design — they displaced the cheap template alternatives, themselves included.
  • Simple one-off content generators: product descriptions, sales emails, social posts, profile bios, app reviews. Each had one template and one function. They are now redundant next to a system instruction in any chat.
  • Online course builders without their own content: LearningStudioAI and similar generated a course outline from a topic. When ChatGPT does it in one prompt without a subscription, the value of a separate tool drops to zero — unless it offers hosting, certificates or a community.
  • Rewriting and paraphrasing tools: QuillBot was the leader with millions of users. Wordtune, Paraphraser.io, dozens of clones. Grammarly barely holds on as an AI platform — smaller paraphrasing-only tools stand no chance when Claude and GPT do it better in every browser window.

The pattern differs from the previous sections: here there was no single company to describe as the victim. There were 10–50 tools in each category, none with billion-dollar valuations, but together a large ecosystem of micro-businesses and indie developers that shrank after Mar 2025.

Sources

Where the data comes from

Traffic: Similarweb (estimates, websites only). Character.AI: bankruptcy filing via Bondoro, TechCrunch, CBS, Reuters. ChatGPT feature dates from OpenAI announcements. Revenue of small companies is declared or estimated by the media.

Prepared by Claude for Dawid, 15 October 2026. Complements the "AI takeover atlas".