Can an AI assistant see the page you are looking at?
Most AI assistants do not see your screen. They either receive a structured description of the page — sent by the site itself — or they know nothing about it at all. Understanding which mode you are in changes what questions are worth asking, and what answers you can trust.
What does it mean for an AI to 'see' a page?
There are two distinct mechanisms, and they are often confused. The first is screen-reading: the assistant has access to a screenshot or a live feed of your display, and it interprets what is visible as an image. The second is page-context injection: the site passes structured text — headings, prices, product descriptions, policy clauses — directly into the assistant's context window. The assistant reads words, not pixels.
A third situation is the most common of all: the assistant has neither. It was trained on data up to a certain date, it has no live connection to the page, and anything it says about that page is either a guess or a hallucination. Knowing which situation applies is the first thing to establish before you rely on an answer.
| Mode | What the assistant receives | Typical risk |
|---|---|---|
| Screen-reading | A rendered image of the visible viewport | Misreads small text, misses content below the fold |
| Page-context injection | Structured text chosen and sent by the site | Only sees what the site decides to share |
| No connection | Nothing — relies on training data alone | Confidently describes a page that may have changed or never existed |
Why is page-context injection safer than screen-reading?
When a site injects its own content, it controls exactly what the assistant sees. Sensitive fields — payment details, private account numbers, session tokens — are simply not included in the payload. A screen-reading assistant, by contrast, sees everything rendered on the display, including content in other tabs if the implementation is careless, and including anything you have typed into a form.
Page-context injection also produces more reliable answers. Structured text does not depend on font size, contrast ratio, or whether an element is partially obscured. The assistant reads the canonical version of the content, not a visual approximation of it. For a pricing table or a legal clause, that distinction matters.
- Context window
- The block of text an AI model can read at one time. Everything the assistant knows about your current situation must fit inside it.
- Grounding
- Tying an AI's answer to a specific, verifiable source rather than letting it draw on general training. Page-context injection is a form of grounding.
- Hallucination
- When a model generates plausible-sounding text that is not supported by any source it was actually given. The risk rises sharply when the assistant has no real page context.
What questions can you actually trust an AI to answer about a page?
If the assistant has been given structured page content, you can ask it to summarise, compare, or explain what is written there. You can ask it to find a specific clause, identify a price, or tell you whether a product matches a requirement — and the answer will be grounded in the actual text.
If the assistant has no page context, treat any answer about a specific page the same way you would treat a stranger's guess. It may be right by coincidence, but you cannot verify it without checking the page yourself. The safer habit is to copy the relevant passage and paste it into the chat, which turns the second situation into a version of the first.
- Ask the assistant what it can see before you ask it about the page.
- If it hesitates or gives a vague answer, assume it has no live context.
- Paste the relevant text yourself when accuracy matters.
- Do not ask an ungrounded assistant to confirm prices, availability, or legal terms.
Does the same logic apply to AI agents that take actions on pages?
Yes, and the stakes are higher. An agent that can click buttons, fill forms, or submit orders needs accurate page context to act correctly. If it is working from a screenshot, a misread element can trigger the wrong action. If it has no context at all, it may fabricate a form structure and attempt to interact with something that does not exist.
The more reliable pattern is an agent that receives structured tool output — a description of the page's interactive elements, their current state, and their labels — before it decides what to do. This is slower to build but produces actions that are auditable: you can read the description the agent used and verify it matches what the page actually contained.
GROX takes this approach when its Connectors act inside external tools: the agent works from structured data about the tool's state rather than from a visual interpretation of it. A simpler browser-automation script may be the better choice if you only need to repeat a fixed sequence of clicks on a page that never changes.
Common questions
Can an AI assistant read a page I have open in another tab?
Only if it has been given explicit access to your screen or browser. A chat assistant embedded in a website sees only what that site chooses to send it. An assistant with screen-reading permissions could potentially see other tabs, which is why reviewing those permissions carefully before granting them is worthwhile.
If I paste a URL into a chat, will the AI read that page?
Not automatically. Some assistants have a browsing tool that fetches and reads a URL when you share one; others will acknowledge the link but have no mechanism to retrieve it. If you are unsure, paste the relevant text directly rather than assuming the assistant has fetched the page.
Is screen-reading always less safe than page-context injection?
For privacy, generally yes — a screen reader sees everything visible, including content you did not intend to share. For completeness, the answer is more nuanced: a screen reader sees the rendered page as a user would, including dynamic content loaded after the initial request, whereas page-context injection only includes what the site explicitly packages. Neither method is perfect.
What should I do when an AI gives me specific details about a page I did not share with it?
Treat those details with scepticism. The assistant may be drawing on an old cached version, on training data that included the page, or it may be confabulating plausible-sounding content. Always verify specific claims — prices, availability, policy terms — against the live page before acting on them.
If you want an agent that acts on structured context rather than guessing from a screenshot, GROX is worth a look — the free tier covers every capability with no card required.