Predictions: 01 to 15
What will be, will be.
This is a little late, but I’ve been holding off on predictions because I don’t like making them, so these are more directional than outlandish. I’d like to get back to this list by the end of 2026 and review what I got right, and what I didn’t. If you agree or disagree with something, or have something to add, do email. I’ll include your predictions in the next predictions post. This one covers posts till Jan 20th.
How to read these: The insights are numbered, and the number of the last prediction in the post is on the featured image, so it’s easy to locate when you’re scanning posts. You might want to read the original article the prediction is drawn from. Here goes nothing:
1. Agentic traffic on the web will exceed human traffic before the end of 2027.
At present, the bot traffic on the web is about 35-40%, at least on Cloudflare, even though a majority of that traffic is for scraping the web, either for training, or for RAG models. Human initiated bot traffic is probably low, but it will be a struggle to differentiate between bot traffic that is human initiated and that which is for scraping. At some point in time in the next two years, the overall bot traffic will exceed 50%. This will lead to a rework of how businesses align their websites and apps to work with AI.
Based on: AI and the quiet rewiring of the open Internet; AI Agents, and why Meta acquired Manus; What happens when AI buys or sells for you.
2. We will see the advent of an agentsphere, which is an agent-centric web. Major websites and services will publish machine-only access layers, such as AI-only endpoints and agent-specific discovery files ( (e.g. an agents.txt or equivalent, similar to a robots.txt) that are not meant to be used or viewed by humans at all.
This happens because MCP and agentic execution invert the interface assumption: services are no longer primarily called by people navigating UI, but by AI agents executing tasks programmatically. Once agents become the dominant sources of traffic, maintaining human-facing flows for those interactions becomes unnecessary overhead, including human-facing navigation, HTML structure, and UX conventions, which are inefficient and brittle for machine execution. Agents need clear data to function: permissions, callable actions, data schemas, and constraints, none of which are reliably expressed through pages meant for humans. Bots have navigated the web before. Now they’re also navigating on behalf of humans, and some autonomously. We will
Based on: AI and the quiet rewiring of the open Internet; AI Agents, and why Meta acquired Manus; What happens when AI buys or sells for you.
3. AI will become default for search and planning
For most tasks that involve search or planning, users will shift completely to starting with AI, including for (restaurant and product review and search, travel planning), in ChatGPT or AI Mode. This shifts the entry point for the Internet completely.
Based on: The Opportunity Trap of the ChatGPT App Store; AI and the quiet rewiring of the open Internet
4. A major consumer price-comparison service will shut down or publicly discontinue its consumer-facing product, citing AI buyer agents as the reason it is no longer viable.
Price aggregators exist to reduce search costs for humans, but AI buyer agents eliminate those costs by performing continuous, direct price discovery across merchants without intermediary interfaces. Once agents crawl, compare, wait, and transact autonomously, aggregators are bypassed rather than consulted. Why go to a website to find something when your agent can find it for you? Comparison sites might be the first casualty of the disappearance of the human decision step.
Based on: What happens when AI buys or sells for you; AI and the quiet rewiring of the open Internet
5. AI chat based health apps will introduce a persistent, longitudinal medical memory feature that integrates with hospital systems, as a primary health record
At present, medical records are spread across different sources, including hospitals, printed prescriptions, fitness devices, and PDFs of test reports. The data is disconnected, incomplete in itself, and medical practitioners and users see benefit in a single source of truth. This memory will serve as an alternate Electronic Health Record, something with governments have tried to build, but haven’t been able to justify because it feels like a privacy violation for citizens without offering them enough value in return. A durable memory that can work across multiple systems, and act as an EHR that you can chat with, works for users, as a persistent personal health system.
Based on: The product challenges that ChatGPT Health will have to navigate
6. ChatGPT and Google will introduce an explicit bidding system for app invocation
App invocation is currently stochastic and opaque, while creating real business impact, and raising bias and self-preferencing risk. The inversion is not better ranking, but decision externalisation: shifting the final selection for paid invocation to a bidding system for specific inferences (travel, shopping, product reviews). Being chosen is valuable enough for someone to pay for it, and gatekeepers extract a price for letting apps get through to a user.
Based on: The Opportunity Trap of the ChatGPT App Store; How to beat the opportunity trap of the ChatGPT App Store.
7. A major social media platform like Instagram or X will enable AI to publish content directly within user feeds.
AI is already present as a tool or responder in X, but still usually requires an explicit trigger (tagging, asking, replying). Within a year, a platform like Instagram or X will start injecting AI-generated content, such as “you might want to know”, or “did you know” content directly into the feed as first-party posts, incorporating either existing social media content, or content repurposed from the open web. Social platforms already optimise for velocity, volume, and engagement, while human content creation is slow, scarce, and unpredictable. Right now platforms personalise feeds. Next they will personalise content as first-party participants inside social platforms.
Based on: When AI enters the conversation
8. OpenAI will launch a first-party social network or social graph product integrated with ChatGPT.
This happens because AI systems that act, recommend, transact, and personalise at scale need relationship context that can’t be reliably reconstructed from prompts or isolated sessions. The social graph as Meta’s core defensive moat as AI absorbs the web and apps. OpenAI lacks this advantage. Without a native graph of relationships, shared history, and group context, ChatGPT is structurally weaker for socially-situated recommendations, trust calibration, and AI participation inside conversations. OpenAI already has usernamesThe observable break is when OpenAI ships persistent user-to-user connections (feeds, groups, shared spaces, or interaction history) that are owned by OpenAI and natively accessible in ChatGPT, rather than depending on external platforms.
Based on: When AI enters the conversation; AI and the quiet rewiring of the open Internet
9. Major consumer apps will move critical features behind authentication walls that deliberately break inside AI interfaces.
This resolves the failure where apps lose control, context, and leverage when invoked as tools inside ChatGPT-style orchestration layers. When discovery and invocation are controlled by AI, apps cannot rely on branding, onboarding, or exclusive attention. The concrete response is not abstract resistance but product design: high-value features (deep filters, personalisation, editing, history, premium outputs) will require users to exit the AI interface and authenticate in the native app or website. Users will hit “sign in to continue” or “view full results in app” walls from AI flows, not as a bug but as an intentional boundary.
Based on: The Opportunity Trap of the ChatGPT App Store; How to beat the opportunity trap of the ChatGPT App Store.
10. A major online platform will launch a clearly labeled, paid or access-controlled space that explicitly excludes AI-agents and AI-generated content by design.
Once AI-generated content is abundant, fast, and engagement-optimised, human-originated interaction loses default visibility and salience. The inversion is not a gradual preference shift but a discrete product decision to create a space defined by the absence of AI participation, and agents pulling content for users: participation is price you pay for access. This crosses a novel line by making “human-only” a premium or gated feature.
Based on: When AI enters the conversation; AI and the quiet rewiring of the open Internet
11. Wallets will make a comeback, and agentic wallets will become a norm
There’s a clear tension between giving an agent a credit card, and the need to enable autonomous purchases to complete transactions. As ecommerce purchases by AI Agents increase, and agentic commerce protocols go live, mechanisms will have to emerge in order to limit how much agents can spend on ecommerce. Wallets are a construct where there is limited access to currency. We’re seeing this in the usage of crypto wallets and stablecoins for transactions, and agentic wallets for fiat based transactions will emerge in order to enable agentic commerce while simultaneously ring-fencing risk.
Based on: AI Agents, and why Meta acquired Manus; What happens when AI buys or sells for you.
12. All Browsers will introduce an Agent Mode
Browsers have spent 15 years hardening against exactly what agents need to do, by breaking cross-site tracking, persistent authentication, enabling CAPTCHA in some cases, requiring user action for auto-fill. AI agents are mimicing user actions when they don’t need to, because browsers can enable them to track preferences and context across websites, stay authenticated, auto-fill forms. Agents are spending tokens overcoming friction that can be avoided. Browsers will enable this, once there is explicit user consent at mode level, not per-action level.
Based on: AI and the quiet rewiring of the open Internet; AI Agents, and why Meta acquired Manus; What happens when AI buys or sells for you.
13. Payment rails will be reworked to bring in agent-centric fees
Agents make mistakes, and at times make autonomous purchases. Human intervention is after-the-fact - they authorised the agent but not the transaction - and can lead to higher chargeback costs. Payments currently cant distinguish between an authorized agent performing an unauthorized action, an agent acting within delegated authority the user forgot about, or an agent getting compromised and performing a transaction. This means more disputes, longer investigations, ambiguity in liability, and higher false positives as fraud detection struggles to cope. This means higher cost of fraud and resolution, which has to be passed on to someone, which is going to be the user.
Based on: What happens when AI buys or sells for you
14. Cloudflare will differentiate between human-initiated agent traffic and scraping bots, allowing websites more choice for regulating traffic.
At present, Cloudflare’s approach is binary: either bots are allowed or not. They’re unable to differentiate between agents act on behalf of users, and those that are just scraping. Sites can’t block all agent traffic (users expect their agents to work) but also can’t allow all bot traffic (scrapers remain a problem). They’ll develop a structural mechanism for a three-tier access: human browsing, verified human-initiated agents, blocked scrapers, allowing websites more choice in enabling access, including delegation credentials.
Based on: AI and the quiet rewiring of the open Internet; AI Agents, and why Meta acquired Manus
15. AI chat based health apps will introduce hard blocking of recommendations, tied to specific classes of health advice after a documented case of an advice leading to user harm.
This resolves a concrete failure state: a recommendation is issued, followed, and later associated with a worsened health outcome that becomes widely referenced. Once such an incident exists, continuing to offer similar guidance without friction becomes indefensible. Rather than relying on disclaimers or softer language, the system response will be categorical: certain recommendation paths (for example, exercise load or medication-related) will be blocked unless predefined conditions are met. Merely recommending that the user consult a physician won’t cut it.
Based on: The product challenges that ChatGPT Health will have to navigate