Agents Now Sign In and Press Your Buttons
- Agents can work in signed-in websites while a user steps away.
- They act through labels, buttons and form fields they can identify.
- Unlabeled fields, image-only buttons and picture-based steps can stop an agent.
- I measure AI visibility with repeatable questions and checks, then retest changes.
- One engagement showed a directional rise in recommendation wins, not proof of a general result.
OpenAI's August 25, 2026 release notes say ChatGPT agents can work in signed-in websites. A user can hand off a task and step away while the agent keeps working.
That changes who uses a company's login pages, forms and account screens. The site has to tell software what each control does, just as it needs to make the task clear to a person.
I build web systems and study how AI tools describe companies. For site owners, the practical question is whether an agent can identify the next step without guessing.
The takeaway
Build account pages with clear labels and actions that software can identify. If an agent cannot complete a user's task, that user may take it to a site where it can.
What changes when an agent signs in to my site?
An agent can continue a task on a signed-in website after the user steps away. OpenAI's release notes also add webhooks, which let an agent respond when something changes in another service. A site task can therefore sit within a larger process that reacts to events elsewhere.
The agent operates a page through its structure: labels, buttons and form fields it can identify by name. A useful design test is whether each control's name tells the agent what action it can take, even when the user is not there to clarify.
What makes a form hard for an agent to use?
An unlabeled field, a button that is only an image, or a step that depends on reading a picture can stop an agent. The problem is that the page has not given the agent a name or instruction it can identify. A person may infer the intended action from the image or surrounding context. Software can fail when it has to guess.
When an agent cannot finish a user's task on your site, the user may move to a site where it can. This can make a form or account screen a point where a customer leaves the process, even if the rest of the site is clear.
How can I tell whether AI tools recommend my company?
I measure AI visibility by asking a locked library of a category's real buying questions of ChatGPT, Claude, Perplexity, Gemini and Grok on a fixed schedule. I check official domains, so a company with the same name does not count as a win. I also ask the model why it chose what it chose.
I keep each run's records in one format. That lets me compare week 11 with week 1 without changing how I count results. The method tracks whether a company appears in answers to buying questions over time. For a business improving its site for agents, it measures a separate issue: whether AI tools recommend the company, not whether an agent can complete a form.
A recommendation win rate measures how often the company is recommended in those tests. The questions and domain checks make the result easier to interpret, while the model's explanation gives me something to inspect when an answer changes.
What did one AI visibility engagement show?
I built the platform Trinzik used for the engagement. Over 11 weeks, Soapbox's AI recommendation win rate rose from 6.2% to 36.2%. This is directional evidence, not a controlled study.
On another program, I published into the gaps and the win rate stayed flat. That result is why I retest instead of declaring victory after a change. A change to a company's web presence does not guarantee a change in AI recommendations.
I also published my rubric for scoring a site's readability by agents. Any firm can use it to score itself and get the same answer. That gives a site team a way to check its own pages before and after making changes.
Common questions
Do I need to redesign my whole website for agents?
Begin with a page where a user needs an agent to complete a task, then inspect the labels, buttons and fields the agent must identify. A failed task points to a specific page or control to review. A readability score does not guarantee that every agent will complete every task.
How should I choose tasks to test?
Choose a task with a clear outcome, such as completing a form, and follow each step the agent must take. Record the point where it stops or cannot identify a control. That gives your team a concrete page issue to investigate, separate from whether AI tools recommend your company.
Should an agent be allowed to take any action after it signs in?
A sign-in gives an agent access to an account, so decide which actions need a person's approval before handing over work. A readable page helps an agent act; it does not decide whether a particular action should be handed off.
If you want help with AI visibility or scoring your site's readability for agents, talk with me.