What It Takes to Adapt to the Agentic Web
Companies can make their information easier for AI systems to understand today. The harder part is planning for what comes next: tomorrow’s agents may not all interact with the web in the same way.
We’re familiar with an internet built for people: we visit websites, read information, and decide what to do next. The agentic web changes who does that work. Increasingly, AI agents will navigate websites and interact with services on our behalf, often behind the scenes. Businesses will need to think about how they serve both human customers and the software acting for them.
That makes the agentic web difficult to plan for. Search platforms are starting to report when sites appear in AI answers. Meanwhile, agents are beginning to interact with websites and commerce systems, including helping customers complete purchases. The capabilities are at different stages, and the rules are still taking shape.
It is tempting to treat all of this as one big optimization problem. But being discovered and giving an agent permission to act are two different jobs. They carry different risks, and they need different measures of success.
Take a hotel. Being discovered means an AI assistant can find it and accurately explain its rooms, prices, and parking policy. Letting an agent act means supporting a connection to the booking system so it can check availability and reserve a room. The hotel controls access to that system and which actions it supports; the customer authorizes the booking and payment. Accurate information helps with the first job. The second also needs working integrations, clear limits, and a way to handle mistakes.
Last week, I went to Adobe’s AI Forum in New York. It offered a useful view of the decisions businesses are already facing.
1. Start with the customer’s question
SEO, or search engine optimization, helps search engines find and understand your website so people can discover it through relevant search results.
AEO, or answer engine optimization, structures your content so a system can extract and display it as a direct answer—for example, in a featured snippet answering whether a hotel offers overnight parking.
GEO, or generative engine optimization, focuses on getting your information or brand included in responses composed by generative AI. An assistant comparing hotels might combine your parking policy with prices, locations, and reviews from other sources.
The key difference is extracting an answer versus using information to compose one. The approaches share foundations, but they target different ways of presenting information.
Google’s guidance still emphasizes familiar SEO fundamentals: crawlable pages, accessible information, and useful content. Appearing in AI Overviews or AI Mode does not require special AI files or schema. That applies to Google Search; other platforms may work differently.
Measuring visibility is more complicated. Adobe’s Brand Visibility tools distinguish between prompts a company chooses to monitor and broader market visibility based on Semrush’s prompt database. Neither captures every customer question or AI platform.
Bing’s AI Performance reporting shows when a site is cited in some AI answers, along with example queries. Those counts do not establish ranking, prominence, or whether an answer helped a customer or led to a sale.
Start with the customer’s actual questions. Pull a small set from sales conversations, support requests, and customer interviews. Check whether your website answers them clearly and how your business appears when you ask relevant AI tools those questions. Compare those observations with the search, referral, and conversion data you already track.
AI-generated questions can help expand that list. But without customer input, it is easy to measure how well we answer questions we invented.
2. Being found is only the beginning
Once an AI system helps a customer find your business, can it also help them get something done?
A company can build its own AI assistant or support customers using outside agents. These require different preparations. A retailer’s shopping assistant, for example, does not automatically let a customer’s agent place an order.
Adobe’s Brand Concierge illustrates the first approach. In its announced partnership with DICK’S Sporting Goods, Adobe described conversational experiences and digital coaching inside the retailer’s mobile app. The announcement outlines the intended experience; it does not establish whether it improved sales or customer outcomes.
Start with the task a customer needs to complete: comparing products, checking whether an item meets their needs, finding a service, or making a purchase. Then decide how an agent could help.
Be clear about what the agent can do on its own. Answering questions and suggesting options may need fewer controls than placing orders, changing account details, or sharing personal information. Decide which actions require the customer’s approval and when a person needs to step in.
Those choices shape the customer experience—and whether people trust it enough to use it.
3. Give systems accurate information and clear limits
An AI system needs accurate information to describe your business. An agent also needs clear instructions and access to the systems required to complete a task.
Start with the basics:
- Make important pages crawlable, with product, service, and policy information available as readable text.
- Keep prices, availability, and policies current and consistent across your website, sales, and support teams.
- Use structured data—labels that help search engines understand a page—where it supports an established search feature. Google does not require special AI markup.
For some businesses, this means cleaning up a service catalog. For others, it means resolving conflicting information across systems. An agent cannot reliably book a room if the website and booking system disagree about availability.
Standards are emerging to make these interactions easier. WebMCP is an experimental approach that lets websites define specific actions agents can perform. UCP, the Universal Commerce Protocol, provides a shared way for businesses, platforms, and agents to support commerce. It already powers checkout in Google’s Universal Cart.
These standards help systems work together. The business still needs to decide:
- Which information should the agent rely on?
- Which systems and actions can it access?
- Which steps require customer approval?
- Who handles failures and corrections?
Answer those questions before connecting an agent to a system that can change a reservation, account, or purchase.
4. Plan for the work that follows
Automating a task changes the work around it. Someone still has to maintain the information, set the rules, review results, and handle mistakes. Those responsibilities need to be part of the plan.
Content production offers a useful example. At Adobe’s forum, reusable creative workflows connected AI models, editing tools, brand rules, and approvals. Adobe’s Firefly Graph lets teams build and reuse those processes. A workflow can carry approved settings across many outputs, but people still need to decide what gets approved, what needs updating, and which exceptions require attention.
The economics reflect that tradeoff. An Adobe and Incisiv report says organizations deploying generative AI for content reported 49% higher throughput and 30% higher quality-control costs, alongside 31% lower costs per asset and an 8% increase in conversion rate.
These are survey-reported results from vendor-sponsored research covering more than 3,400 senior executives. The public report does not provide enough methodological detail to treat them as a forecast for another business.
Still, they illustrate why production speed alone is an incomplete measure. More output can create more review work while lowering the cost of each asset. The question is whether the full process delivers better results for the money spent.
Include maintenance, review, exception handling, and corrections in the business case. Give those responsibilities an owner and a budget, then measure whether the workflow improves customer outcomes or simply produces more material to check.
5. Preserve the reason a customer chooses you
There is a risk in treating every new capability as a strategy. Making content easier to find, producing more of it, and connecting more tools can improve how a business operates. They do not explain why a customer should choose it.
That reason might be a product feature, unusually helpful service, expertise, or a relationship built over time.
As AI systems compare options for customers, businesses need to make those strengths understandable beyond their own websites. A claim about exceptional service needs supporting evidence. Specialized expertise needs to be demonstrated. Clear product information, customer reviews, and credible independent coverage can help communicate what sets a business apart.
If an AI system becomes the customer’s main interface, the business still needs to give that customer a reason to choose it.
What You Should Read
Cloudflare and OpenAI: Making AI discovery better for everyone
Cloudflare and OpenAI · October 15, 2026
The useful detail here is an early look at the infrastructure behind discovery. A pilot covering roughly 1% of Cloudflare’s network helped OpenAI find and index hundreds of millions of previously unknown pages. The companies also distinguish being discovered from what happens afterward. Businesses need to keep making that distinction: visibility is measurable sooner than business impact.
Amex introduces a framework to prevent AI agents from making unauthorized purchases
Sarah Perez, TechCrunch · October 13, 2026
American Express is proposing a way to capture a customer’s intent, constraints, and approval before an agent completes a purchase. It is a practical example of why the agentic web is a permissions problem as much as a discovery problem. Once software can act, businesses need a record of what the customer authorized, and a way to stop a transaction when the agent strays from it.
An intelligent new shopping cart
Suresh Ganapathy, Google · October 15, 2026
Google’s Universal Cart is now rolling out in the U.S. across Search and Gemini. It keeps products from multiple retailers in one place, tracks price and inventory changes, surfaces promotions, and supports checkout through Google Pay or the merchant’s own site. It offers a concrete view of an emerging customer journey, where discovery, comparison, memory, and transaction are starting to collapse into one interface.
Websites are closing their doors to AI agents
Emma Roth, The Verge · October 17, 2026
The agentic web is running into a practical barrier: many businesses do not want outside agents accessing their services. The article looks at companies blocking agents over security, licensing, customer-experience, and business-model concerns, while others are choosing partnerships instead. It brings the strategy question down to earth. Businesses have to decide where agents are welcome, what they may do, and what they must prove first.
Businesses confront the accountability gap in agentic AI
Julie Bogen, Axios · October 13, 2026
Business leaders from Box, Credo AI, and You.com discuss the controls companies need before agents can take on high-stakes work: clear access boundaries, human intervention points, and ownership of the consequences. A useful reminder that stronger models do not resolve the operating questions around privacy, compliance, and liability.
Who You Should Follow
Aleyda Solís
An international SEO and AI-search consultant whose frameworks, worksheets, and measurement guidance help make a noisy field more practical. Her recent work is useful for teams trying to connect technical readiness and citation visibility to business outcomes without treating GEO as magic.
Kevin Indig
The author of Growth Memo, a research-driven newsletter on organic growth, SEO, and AI search with more than 25,000 subscribers. Follow him for data-heavy analysis of how search behavior is changing and for his habit of testing claims that quickly become conventional wisdom.
Michael King
The founder and CEO of iPullRank. He posts frequently about AI search, relevance engineering, measurement, and the operating model behind GEO programs. His work is useful for separating what can be measured precisely from what AI-search tools only estimate.
Jack Galilee
A former Human Native co-founder now working at Cloudflare on the infrastructure and economics of the agentic web. His recent writing and posts offer a way to follow the standards, discovery systems, permissions, and payment rails still being assembled.
Crystal Carter
The head of AI Search and SEO Communications at Wix. She posts regularly about how websites, brands, and search teams can prepare for AI-mediated discovery, translating technical changes into practical guidance for marketers and business teams.
Where You Should Go
Philly Data & AI — October Happy Hour
Tuesday, October 20 · 6:00–8:00 PM · Con Murphy’s Irish Pub, Philadelphia
Hosted by Philly Data & AI. No speakers or sales pitches, just people working in data, AI, and startups. It is an easy, low-pressure way to meet local practitioners and compare notes on what is making it into production.
Airtable Philly Builder Crew: October Meetup
Tuesday, October 20 · Check the organizer’s registration page for the confirmed time · City Winery Philadelphia
Hosted by Airtable Partners & Community. The evening combines happy hour with a builder roundtable on Airtable problems, lessons learned, and newer features people want to explore.
15th Philly Hardware Meetup @ Drexel
Wednesday, October 21 · 5:30–8:00 PM · Bossone Research Enterprise Center, Drexel University
A strong crossover event. Matt Starfield will discuss methods, guardrails, and results from using agentic AI in day-to-day hardware engineering, alongside talks on diagnostics and climate hardware.
AI That Ships: The Operating Model and the Serving Stack
Thursday, October 29 · 6:00–8:00 PM, with networking until 8:30 PM · Bartley Hall, Room 1011, Villanova University
Two substantive talks: Mariano Mattei on the operating model required to move AI from pilot to production, and Vicki Boykis on ONNX and ONNX Runtime for model serving. The talks are followed by networking until 8:30 PM.
Philadelphia AI Tech & Finance Networking Event
Friday, October 30 · 7:00–9:00 PM · Tapster Philadelphia, Rittenhouse Square
Organized by SUM OF FIVE USA. An informal networking option for readers interested in meeting people across AI, technology, and finance rather than attending a formal talk.