> ## Content Index
> Fetch the complete content index at: https://bielawski.pro/llms.txt
> Use this file to discover other available public pages before exploring further.

# Press review 15.11.2026: terms for agents, token pricing and webinar show rate
- URL: https://bielawski.pro/en/press-review-15-11-2026-terms-for-agents-token-pricing-and-webinar-show-rate/
- Published: 2026-11-15T06:07:56.000Z
- Updated: 2026-11-15T06:07:56.000Z
- Description: A US publisher lost 37 percent of its Google visits in a year while its AI impressions ran ten times higher than its clicks. An airline shut out AI agents with its terms of service, not with robots.txt.
- Author: Bartosz Bielawski
- Tags: Prasówka, #en

## SEO and search in AI

### An airline lets Google index its site but makes AI agents ask for permission

United Airlines updated its terms of service on 11 November. Search engines may index the public pages, but any AI agent that searches, books or buys on united.com needs the airline's written permission. Bookings made by an agent are cancelled without notice.

The definition of an agent is broad: AI assistants, bots, scrapers, browser extensions and agent features built into browsers or apps. It also covers cases where a human supervises the agent's work.

As of 14 November, United's robots.txt did not name a single crawler or AI agent. The whole decision lives in the terms, not in the file that bots read. Amazon blocked Meta's Muse agent under its own Conditions of Use, and eBay added shopping agents and LLM bots to the list of tools that require consent.

What to do: check whether your own terms mention agents at all. If you run a store and want traffic from shopping agents, make sure your terms and your CDN rules do not close doors you meant to keep open.

**Sources:**

- Search Engine Journal, 2026-11-14: [United Lets Search Engines Index Its Site, Makes AI Agents Ask First](https://www.searchenginejournal.com/united-ai-agents-permission-search-indexing/597418/?ref=bielawski.pro)

### ChatGPT brings less than one percent of the sessions Google does - check your own numbers before comparing channels

Marie Haynes analysed 12 months of GA4 data across 100 sites and more than 145 million sessions. ChatGPT delivered 0.62% as many sessions as Google organic, and 0.73% when all LLMs were included.

A benchmark by the agency The SEO Works across 95 sites showed a similar picture: organic search delivered 39 times the traffic of AI assistants. Testing 632 cases across 12 sites, Kevin Indig found no reliable link between AI visibility and later growth in branded searches.

The picture changes when you look at traffic as a whole. In Similarweb's analysis of 10,000 US domains, organic visits fell 13.2% between August 2024 and August 2026, and publishers lost about 41.5% of their organic traffic. Growth in paid traffic offset only about a third of those losses, while email grew 16% and was the most stable channel.

What to do: stop comparing yourself with global benchmarks and track your own share of assistant sessions and conversions. Tag the AI channel separately in GA4, or it will end up under Direct and you will conclude that traffic fell for no reason.

**Sources:**

- Search Engine Journal, 2026-11-12: [The Traffic Pool Is Shrinking, And Which Channel Fills It Matters Less Than We Think](https://www.searchenginejournal.com/the-traffic-pool-is-shrinking-and-which-channel-fills-it-matters-less-than-we-think/596346/?ref=bielawski.pro)

### A booking agent picks a sponsored listing nine times more often when it works for the platform

The McGill study covered 100 hotel offers, 5,000 agent sessions and four models. Agents searched more thoroughly than people did and never once declined to book, and a single offer took 78.2% of all bookings.

Position on the results page had a weak and unexpected effect on the choice - the middle of the page, not the bottom, was inspected least. Agents that were told they worked for a booking platform chose sponsored listings 25% of the time instead of 2.8%, with all other attributes unchanged.

The authors' conclusion is practical: in agentic search, what is visible on the results page matters more than where you sit on the list. Clearly stated price, availability and terms decide which offer an agent opens and which one it books.

What to do: if you sell through platforms, check how your offer looks within the results themselves - whether an agent can see the price, availability and refund terms without opening the page. Without that data, position counts for less and less.

**Sources:**

- Search Engine Journal, 2026-11-12: [The Traffic Pool Is Shrinking, And Which Channel Fills It Matters Less Than We Think](https://www.searchenginejournal.com/the-traffic-pool-is-shrinking-and-which-channel-fills-it-matters-less-than-we-think/596346/?ref=bielawski.pro)

### You can claim a business profile in ChatGPT with a DNS text record and a QR code - no app needed

OpenAI launched the beta of Businesses in ChatGPT on 10 November. You claim a business in the Business Hub: you verify the domain with a DNS TXT record and your identity through Persona - via a QR code or a secure link on your phone, with no app to install.

Unlike Google Business Profile, the hub accepts online-only businesses. Entering a physical location led to a waitlist, while signing up through a website went through. Every brand and every location needs its own profile, because there is no central way to manage a parent company's brands yet.

Search Engine Journal claimed its profile and saw 353,356 mentions and 2,663 clicks to its site between 14 October and 10 November, with daily mentions up about 90% from the first day to the last. A mention counts only when ChatGPT references the business's website, not when the brand name alone appears.

What to do: claim your domain and write the description in your customers' language, because the model may use it to match services to questions. Saving a draft triggers an offer of 500 dollars in ad credit, but claiming a profile does not guarantee higher placement - it is one source of information about a business, not the overriding one.

**Sources:**

- Search Engine Journal, 2026-11-12: [How to Submit Your Website to ChatGPT Business Hub](https://www.searchenginejournal.com/submit-website-chatgpt/596785/?ref=bielawski.pro)
- OpenAI, 2026-11-10: [Introducing Businesses in ChatGPT (Business Hub + Business Agents)](https://openai.com/index/introducing-businesses-chatgpt/?ref=bielawski.pro)

### A publisher lost 37 percent of its Google visits in a year while its AI impressions were ten times its clicks

Ziff Davis reported that organic visits from Google fell about 37% year over year. Over the same period, its content collected 2.7 billion impressions in AI Overviews and AI Mode - almost ten times the clicks coming from Google Search.

That shows where the gap opens up: a great many people see the AI answer, but very few reach the site. A model does not have to skip you for your traffic to fall - it only has to answer in your place.

The same report notes that each platform defines its own AI visibility metric: a mention, an impression and a click mean different things in ChatGPT, in Maps and in AI Mode. Unless you know the date a model changed, numbers taken two weeks apart are not comparable.

What to do: measure your own impression-to-click ratio in AI reports and record the dates when models change, so you know when a comparison stops making sense. In your budget, treat visibility without a click as a separate result.

**Sources:**

- Search Engine Journal, 2026-11-14: [Core Update, ChatGPT Business Hub & AI Mode's New Model](https://www.searchenginejournal.com/seo-pulse-core-update-chatgpt-business-hub-ai-modes-new-model/597135/?ref=bielawski.pro)

## AI, agents and costs

### Meta and OpenAI agents account for almost 90 percent of agent traffic, and half of Muse requests arrive unsigned

HUMAN Security monitored AI agent traffic in October and found that agents from Meta and OpenAI made almost 90% of requests. Meta's Muse alone accounted for 59% and sent 3.7 times as many requests as ChatGPT Agent, the runner-up.

ChatGPT Agent's request count grew 44%, but its share fell from 28.5% to 15.9% because Muse grew faster. The Claude Chrome extension lost half its share, and Perplexity's Comet fell from 41% in August to 3.6% in October.

Most of that traffic goes to one place: 85% of requests hit product and search pages, and e-commerce consistently generated about 60% of agent traffic. The catch: between 7 and 9 November, Muse moved its Web Bot Auth signatures from Meta's key directory to muse.ai, and by 10 November half of its requests were unsigned.

What to do: if you block bots at the WAF, check that the rule does not block shopping agents along with scrapers. Signature checks built on the old key directory no longer work, and an agent like Muse has no name for robots.txt to match.

**Sources:**

- Search Engine Journal, 2026-11-14: [Meta's Muse Now Sends Most AI Agent Requests One Security Firm Sees](https://www.searchenginejournal.com/meta-muse-ai-agent-requests-human-security/597388/?ref=bielawski.pro)

### The same research costs 50 times less when the agent's search runs on a cheaper model

Fable 5.2 costs 8 dollars per million input tokens and 40 dollars per million output tokens, while Opus costs 4.20\. A typical job with 75,000 words in and 15,000 out costs about 1.60 dollars on Fable, 80 cents on Opus and 2 cents on Haiku.

The difference comes from what the agent does, not from the model's price. By default, the search subagent runs on your session's model, up to Opus. The same agent set to Haiku gives a similar answer for about 50 times less.

Effort level has a price too. Fable on medium scored 37.6% on the Frontier Coding benchmark at 2.48 dollars per task, and 45.3% on high at 6.72 dollars. Subagents inherit the effort level from the main session, so a single setting raises the cost of the whole job.

What to do: for repetitive work such as research or summaries, set the subagent to a cheaper model and medium effort. Calculate cost per task rather than per token, because orchestrating several subagents can cost five times more than a single pass.

**Sources:**

- AI Edge, 2026-11-14: [Anthropic Just Leaked How To Use Fable 5.2 For CHEAP! (insane hack)](https://www.youtube.com/watch?v=sGH7TXA-BeA&t=72s&ref=bielawski.pro)

### A screenshot of a dashboard is enough to get a working panel with live data

Claude Fable 5.2 can rebuild an entire dashboard page from a single screenshot. The live dashboard feature, now renamed Claude Data, lets you connect the rebuilt panel to numbers that update live.

That changes the order of work on a report: instead of designing the layout from scratch, you show the model a screenshot of the panel the client already knows and get a working version of it. The panel is not static, so the numbers do not fall out of sync on the first refresh.

What to do: use it to prototype a client report - start with a screenshot from GA4 or Looker Studio, then swap in the real data source. Before anything goes out, check that the numbers in the panel match your export.

**Sources:**

- Skill Leap AI, 2026-11-14: [Claude Turned a Screenshot Into a Live Dashboard](https://www.youtube.com/watch?v=%5FZEsnRx5gZ0&t=17s&ref=bielawski.pro)

## Google Ads and paid media

### Local Services Ads now live in Google Ads, but the old reports did not make the move

Local Services Ads were moved into Google Ads, and the old screen disappeared. Clients who looked for it after the navigation change report that they cannot find the reports they were used to.

The billing model has not changed: you still pay for a call or a message rather than a click, and the ads still show in the same places in Search and Maps. What did not carry over is the data on impressions, clicks and spend.

What to do: save your historical LSA reports before the account is migrated, because you will not be able to rebuild them in the new interface. If you manage many local accounts, set a reminder to export them before Google closes access to the old view.

**Sources:**

- Missoula SEO Geek, 2026-11-14: [They Couldn't Find Their Local Services Ads #shorts #googleads #localseo](https://www.youtube.com/watch?v=hi-drNCZa9E&t=0s&ref=bielawski.pro)

### Ads get harder to scale because of proof you can fake, not because of creative

In local campaigns, the offer matters more than brand building - trust in a local business is naturally high. An ad works when it shows proof, and that is exactly why scaling is hard: the next piece of proof has to be earned, not written.

About 80% of effective ads rely on proof, specifically the kind that is hardest to fake - a concrete number, a real customer, a visible result. Adding friction to a form raises the cost per lead but improves quality when leads are plentiful and weak.

Scaling ROAS comes down to two routes: lowering the cost of acquiring a customer or raising customer value over time. On the first route, most of the gain comes from studying the first three seconds of your best ads and modelling the rest on them.

What to do: before raising your budget, check whether you have enough proof in stock for the next audiences. If not, work on customer value and form friction instead, because pouring budget into ads without proof only buys the same traffic at a higher price.

**Sources:**

- MoreMozi, 2026-11-14: ["Why Are My Ads Getting Harder to Scale?"](https://www.youtube.com/watch?v=USfh8GMrX88&t=397s&ref=bielawski.pro)

## Marketing, sales and business

### Three text messages, not a new offer, lifted webinar attendance from 18 to 24 percent

A portfolio company was doing 570,000 dollars a month and had been flat for six months. One webinar booked 58 sales calls, but attendance stood at 18%, against a target of 25% or more.

The reason was mundane: the webinar ran once a week. After the company added text reminders - 24 hours before, on the morning of the event and an hour before - attendance rose to 24%, a 33% increase.

Text messages get 95-99% open rates, so they reach people that email no longer does. The author points out that lifting attendance by a third is easier than lifting the close rate by the same amount, and that the first minutes of the presentation matter most.

What to do: if your webinars or demos have weak turnout, start with three text reminders and check attendance the following week. Leave changes to the offer for later, because the problem is often reaching people, not the content.

**Sources:**

- MoreMozi, 2026-11-15: ["Why Do Leads Go Cold Between Marketing and Sales?"](https://www.youtube.com/watch?v=IFvZPg71SlE&t=326s&ref=bielawski.pro)

### In three years most software users will come in through an agent, not through the interface

Michał Barczak argues that in about three years, 60-70% of users of any software will use it through MCP, that is, through their own AI agent, and only 30-40% will log in to the interface. Today, nobody designs a product that way.

In his view, it has never been easier to build working software and never harder to sell it. That is why a startup should put more budget into marketing and sales and less into development.

At the same time, defensible value shifts to systems that serve as the single source of truth, such as a CRM. The basics of marketing have not changed, and chasing the next model is secondary, because next week there will be a different one.

What to do: on new implementations, ask whether the client's data is reachable through an API and whether an agent can complete a task without clicking through a screen. It is the same work as integrations, only planned three years earlier than usual.

**Sources:**

- Michał Barczak, 2026-11-14: [Ludzie pytają AI zamiast Google. Co to znaczy dla Twojej firmy?](https://www.youtube.com/watch?v=ygq9WK4s2l8&t=699s&ref=bielawski.pro)