Crop, mark up, zoom and pan

The editor is laid out like a design tool's canvas. The capture fills the stage, and a few controls float over it: in the top-left corner, what you scanned (the site's icon and the page's title, cut short past 25 characters, with the whole title when you point at it); at the top center of the canvas, the toolbar; in the bottom-left corner, the zoom. The panels that decide what the file looks like open from a rail at the right edge. The panel starts folded, so the page gets the whole width; click an icon on the rail, or the expand button at its head, to open one, or press ⌘/Ctrl and the panel's number: 1 is Output Format, 7 is the Assistant. Hold ⌘/Ctrl for a moment to see each button's key badged on it. The toolbar stays centered over the canvas, so it moves over when the panel opens or folds.

Crop

The capture opens with the crop on the whole page and the crop tool active. Drag the frame's edges or corners to resize it, and drag inside it to move it. Crop to the main content on the toolbar puts it on the page's article or column of text, past the site's header, its sidebars and its footer, and the button is off while the crop is on it. Reset crop, beside it, or Esc, puts it back on the whole page.

To place something in the frame by eye, the Frame panel's last group, Guides, draws lines inside the crop: the rule of thirds, the golden ratio, center lines, quarters or the two diagonals. They are off until you switch them on, stay as you left them for the next capture, and are never in the file.

The main content is found while the page is scanned, from the page's own markup (its main and article) and, where the markup says nothing, from where the paragraphs of real text are. A page with no one column, such as a front page or a search result, has none, and the button is off. A page that sets its text beside a sidebar, as a GitHub issue or a forum topic does, opens on the column of posts and comments without the sidebar. On a news site the crop can run on past the story into the recommendations below it; drag the bottom edge up. Settings can make every capture open on the main content.

The crop is exact rather than approximate. The preview you drag over and the PDF are taken from the same layout of the page, so the frame lands on the file exactly where you drew it. On a Retina screen the bottom of a very long page can be a few pixels off; the FAQ says why.

Mark up the capture

The toolbar at the top of the editor holds the tools, in groups. Each group is one button showing the tool you last used from it, with a small arrow beside it that lists the rest of the group:

The tool in use is filled green. A tool's key picks it from any group. After the tools come undo and redo, then the two crop buttons, and Extract at the right end.

Remove element takes something out of the page itself: a banner, a sidebar, a card in the middle of an article. Point at it and the label beside the pointer names it; ↑ picks the box around it, ↓ goes back in, and ← and → step to the box before or after it. Click to mark it (as many as you like), then press Remove 1 element (or Remove 3 elements, by how many you marked) at the bottom of the stage. The page is scanned again without it, so what was below moves up to close the gap. The tab you scanned has to still be open on the same page, and the new capture starts without your crop and marks. Put back scans it again with everything in.

Pick a shape and drag on the capture to draw it. Hold Shift to keep a square, a circle, or a line on a multiple of 45°. As soon as you let go the new mark is selected and the tool goes back to Select, so you can adjust it at once: drag it to move it, drag a handle to resize a box or move the end of a line, press Delete to remove it. Click on nothing to deselect; drag on nothing to pan.

While a mark tool is in use or a mark is selected, a second bar appears under the toolbar: eight colors and a picker for any other, four stroke widths, Fill for a rectangle or an ellipse (a translucent tint of the stroke color), Rounded for a rectangle, Both ends for an arrow. With Remember the last choices on in Settings, the bar keeps what you last chose, so the next capture's first mark is drawn the same way; it is off at first. A change restyles the selected mark and becomes what the next mark is drawn with.

Text is written in place. Press T, click where the words go (or drag to set the width), and type. The words wrap inside the box, whose height follows them; click elsewhere or press Escape when you are done, and double-click the words to change them later. The bar offers the color, five sizes, three faces (Sans, Serif and Mono), bold, italic, and a box behind the words in a color of your choosing, for text over a busy page. In the PDF the words are text: you can select and search them. They are set in the PDF's own standard fonts, which cover Latin script and a few extras but not Korean, Japanese or Chinese. Words in those scripts go into the PDF as a picture, drawn in your computer's font the way you see them on the stage, with the words laid over the picture invisibly, so you can still select, search and copy them. The box says so as you type.

On a page in Korean, Japanese or Chinese there is a better way. The Page list in the bar holds the fonts the page itself is set in. Pick one and your words are drawn in the page's own font, and in the PDF they are real text in that font. The page only carries the characters it shows, so if you type one it never used, the box names it and the words go in as a picture again. Bold and italic come with the font, so those two buttons are off while one is chosen.

The highlighter has four pens. Drag across a passage and it is covered in a band of color that darkens the page and leaves the text at full contrast, the way a real one does; hold Shift to keep the band straight and one line tall. Two passes of one color look like one. The text under a band is still text in the PDF.

Start the drag on a word and the selection follows the text, line by line, like selecting text in a browser: the highlighter, the blur, the redaction and the erase all do this, and the pointer shows the text cursor over the words. Start the drag beside the text, or hold Alt, for a free box instead. A capture that fell back to raster, or a picture you opened, has no text to follow until the Sensitive Text panel has read it; after that the drag follows the words it read.

Emoji are stamped. Press E and the bar becomes a picker: type a word (thumbs, fire, check) to narrow the grid, click the one you want, then click on the capture to place it. A click stamps it at a fixed size; a drag sets the size. Once placed, drag it to move it, drag a corner to resize it, and pick another emoji in the bar to swap it. The emoji is drawn by your computer's own emoji font, and it goes into the PDF as a picture rather than as text, so it looks the way your system draws it and cannot be searched for.

Step numbers put 1, 2, 3 on the page for a how-to or a bug report. Press N and click where each step is; the tool stays in hand, so each click places the next number. A drag sets a badge's size. The number is the badge's place in the order you placed them, so deleting one renumbers the rest. The bar offers the color, and the number is white or black, whichever reads on it.

The magnifier shows a small detail larger. Press M and drag round the detail: the region gets an outline, and a copy at twice its size appears beside it, joined by a line. Drag the copy to move it, drag its corners to resize it (it keeps the region's shape), and drag the region's outline to look at another spot. The copy includes the marks under the region, so a blurred, redacted or erased detail stays hidden in it, even if you hide it after making the copy. The copy is a picture of the region, in a PDF as in a PNG, so it is as sharp as the capture and its words cannot be selected; the words on the page beside it still can.

Cut a band out (K) takes a stretch of the page out of the file and closes the gap: the ad between two paragraphs, the newsletter box in the middle of an article. Drag down across it; the band always runs the full width of the page. On the stage it is hatched and the page stays where it is, but the Output Format panel counts the pages and pixels without it, and in the PDF and the image what is below the band moves up to meet what is above. Its words are not in the PDF at all. Drag its top or bottom edge to adjust it, or delete it to put the band back.

Sensitive Text, the shield below the other icons on the editor's right edge, opens a panel that looks through the crop's text for what should not be shared. It looks for nothing until you press one of its scans, each a magnifier on the right of its title, and each as often as you like:

• AI Scan looks for what no rule can: tick Names (people's names, and user names and handles), Postal addresses, ID numbers (national ID and resident registration numbers, social security numbers, passport and driver's licence numbers, tax and business registration numbers) or Payment details (bank account and routing numbers, card expiry dates and security codes, and card or account numbers the rules missed). It runs on Chrome's own built-in model, on your computer, and the text is not sent anywhere; or on a model you serve on this computer, or a small model Page Scanner downloads to run in the browser, chosen in Settings › AI Model. Where there is no model to ask, the panel says why and the scan button stays disabled, and offers a small model to download from Hugging Face and run in the browser, with its size on the button. Once it is downloaded, AI Scan and tidying use it. Where the browser has no WebGPU, it points at a model you serve on this computer instead.
• Pattern Scan looks for what has a known shape: tick Email addresses, Phone numbers, Card numbers, IBANs, ID numbers, IP addresses or Keys and tokens. The words and patterns you added in the settings are looked for with them. ID numbers checks the digits of Korea's resident and business registration numbers, Japan's My Number, China's resident ID, Italy's codice fiscale and partita IVA, and US Social Security numbers, on your computer. A number written as bare digits, and any My Number, is found only with its name just before it (주민등록번호, マイナンバー, SSN...), since order and tracking numbers look the same.
• Custom Pattern takes a word or a /pattern/ to find in this capture only; press its magnifier or Enter.

The ticks of both scans are folded away at first; the arrow left of a scan's magnifier opens them. They are remembered for your next capture. Each run of a scan replaces its last run's finds, so a kind you untick is gone after the next run.

Everything found is listed under Finds, masked, with the count in a badge beside the title, and outlined on the stage. The list shows about seven at a time and scrolls. Each find has an eye that brings it into view at 100 %, and a trash can that takes it off the list when it is not what it was taken for: it stays off for that capture, even if you scan again, and the Undo in the message brings it back. Untick what should stay. The checkbox on the right of Finds ticks or unticks them all: it shows a check when all are ticked and a dash when only some are. Under Hide Them By choose Erase, Blur or Redact, with the blur's strength or the redaction's color right there, and the button at the foot of the panel, Redact 9 for instance, hides the rest in one step. They become ordinary marks, which you can still move or delete, and none of them is left selected. A find that is already under a blur, a redaction or an erase drops out of the list, so once they are hidden the panel says there is nothing left to hide; undo brings them back. The question mark beside the panel's title says what each way does to the PDF.

A capture taken as an image, or a picture you opened, has no text of its own, so the panel reads it first, on your computer, with the same engine a searchable PDF uses, and shows how far along it is. Recognition can misread a word, and a misread word is not found, so look the crop over as well. Whichever way you hide a find there, it is painted over the picture in every export. On a scanned PDF the panel offers Redact only, and says why: a blur or an erase would leave the scan's pixels in the file, while a redaction turns that PDF into a picture with the box painted in.

While AI Scan runs, a green spinner turns beside its line, which says what is going on and the time so far ("Scanning the crop with Chrome's model, 12 s so far."). The first time, Chrome downloads the model; once Chrome gives a figure the line gives its percent, and until then it says "It has not said how far along it is; 2 min so far". Chrome often gives no figure until the download is over, so a spinner that turns for minutes is a download still going, not a stuck one.

Blur, Redact and Erase text hide what is under a box, and they are three tools because they promise three different things. Blur softens the pixels so a thing cannot be read; the bar offers five radii, and the stage shows the blur as it will come out. Redact paints the box solid, black unless you pick another color, and removes what is under it from the file. Erase text takes the words out as if you had deleted them and paints nothing: on the stage and in an image the box takes the color of the page around it, and in the PDF the page's own background shows where the words were. An erase always sits under your other marks, so an arrow or a highlight drawn over it stays visible.

In a PNG or a JPEG the three are what they look like. In a PDF, read this before you rely on any of them. The PDF normally keeps the page's text as text, and a picture laid over text hides it from nobody who copies or searches the file. So under a blur the PDF's text is rewritten: each letter under the box becomes another letter of the same font, spaces stay, and what comes out of the file where the blur is reads like xxxx xxxxx. The blurred picture is laid over those. A photo, a chart or a drawing under a blur is only covered. Under a redaction the PDF's text is removed outright, and the rest of the page stays text. A redaction promises that nothing under it survives, and a photo cannot be cut out of the page, nor a drawing that runs past the box's edge: so a redaction over one makes the whole PDF a picture, with the box painted into it, and the price is that the text elsewhere on the page is no longer selectable. Under an erase the PDF's text is removed too, and nothing is drawn in its place. An erase promises only that the text is gone: a photo, a logo or a drawing under it stays in the file, in plain sight, and the rest of the page stays text. The hint on the Output Format panel's PDF option says which you will get before you extract, with the reason. If you need the rest of the page's text to stay text, crop the picture away first, or redact in a PNG. The marks go into every export. In a PDF they are vectors drawn over the page, and the page's text stays selectable around them; a mark that crosses a page break appears on both sheets, cut at the break. In a PNG or a JPEG they are drawn over the pixels. A mark outside the crop is not in the file; the stage draws it faintly there so you can bring it back. Marks are kept with the capture, so a reloaded editor tab still has them, and like the capture they never leave the browser.

⌘/Ctrl + Z undoes and ⌘/Ctrl + Shift + Z redoes. Esc backs out one step at a time: the selection, then the tool back to crop, then the crop back to the whole page.

The stage's menu

A right-click on the stage opens a menu of what the editor can do from there, each with its key: Copy as PNG (⌘/Ctrl + C, which copies the crop to the clipboard as a picture with its marks, whatever format the Output Format panel has), Extract, which writes the file at once, the crop (to the main content or the picked element, to what is on screen, or back to the whole page), undo and redo, deleting the selected mark, the tools, the zoom and the settings.

Zoom and pan

Actual size also scrolls back to the top of the crop, just under the toolbar, so it still does something when you are already at 100 %.

Fit to window shows the whole capture, however tall, with room left above it for the toolbar and below it for the zoom, so the crop's top edge never opens under the toolbar. Zoom goes from 2 % to 1,600 %, but a page too tall to fit at 2 % is fitted below it, 60,000 px at about 1 % on a laptop's screen, and zooming out stops there. The percentage in the bottom-left corner opens a menu of Fit to window and a few fixed sizes.

Accessibility

With Check accessibility on in the settings, the editor's rail has an Accessibility panel for the capture. It says how many WCAG rules the page failed and by how many elements, then lists each rule with its elements, numbered. The eye beside an element takes the view to it, and while the panel is open every failing element is boxed and numbered on the page, red for critical, orange for serious, blue for moderate and gray for minor. The boxes are only on screen; an export does not carry them. Download marked PDF saves the capture with the boxes drawn on and the rules listed after the page, to pass on.

Find in the capture

⌘/Ctrl + F opens the editor's own find bar at the top right of the window, where it stays when a panel opens. Chrome's find bar cannot read the capture, because the page is drawn as a picture, so the editor searches the capture's text itself. Type a word or a phrase: every place it appears is painted amber and the current one green, and the count says which of how many it is on. Enter goes to the next and Shift + Enter to the one before, and the stage moves to it, zooming to 100 % when the text is too small to read where you are. Esc closes the bar.

The whole page is searched, whatever its case, and a phrase is found where it wraps from one line to the next. A picture, or a PDF opened from disk that holds only scanned pages, has its text recognized on your computer first; the bar shows how far along that is. Words inside an image on a scanned page are not searched.

Tell the editor what to do

The last icon on the rail, a robot in a speech bubble, opens the AI Assistant, and ⌘/Ctrl + K opens it with the cursor in its box. Type what the capture should become, in any language, and press Enter (Shift + Enter starts a new line). Until you send one, the panel offers two to start from, and a click sends it:

• "A4 PDF with the stamp below"
• "Blur the email addresses"

It is a conversation, so a follow-up builds on what came before: "and put the stamp below" after "A4 PDF" changes only the stamp. A question about the file ("what format will it be?") gets an answer and no change. When a request is unclear, it asks what you meant, and when it cannot do what you asked, it says so and where in the editor it is done by hand.

The Assistant is a preview, badged Preview beside its title. Page Scanner can close it, or only its server, while it is one: closed, its icon and ⌘/Ctrl + K are gone, or a red note at the end of its conversation says the server is closed for now and the Assistant reads nothing.

A message can take the Assistant a few steps: it looks, acts, reads what came of it and goes on until it answers in a sentence, for at most 8 steps. Looking runs by itself, each a line with a check or a cross: reading the editor's choices, finding sensitive text, finding words, zooming and picking a tool. A change to the file waits. The Assistant lists every change it wants as a numbered plan and runs nothing yet, and above the box it asks Run these steps? on a card of its own, the same way the side panel asks before a scan. Press ⌘/Ctrl + Enter, or Run, to carry it out; the shortcut works from anywhere in the editor while the Assistant is open, unless you are typing in another field. Cancel, Esc in the empty box, or sending a new message drops the plan, which is marked "Not run.", and the Assistant goes no further with it. Each line turns into a check or a cross as it runs, with a sentence saying what happened; a line that cannot run says why, and the rest still run. The Assistant then says what changed.

Undo, under a plan that ran, puts back the choices it changed, the crop and what it hid, but only what is still as the plan left it: a padding you changed after a plan that set the window stays yours, and a hide is not taken back once the marks have changed since. When nothing is left to put back, it says so. A zoom, a tool, a find, a copy and an export are not taken back. "Undo that", said to the Assistant, sets the choices back too.

It can set the format (PDF, PNG, JPEG, Markdown or Word), the page size and the page breaks, text in pictures, PDF/A, tidying, the window and the padding, the stamp, the integrity record and the reference file; crop to the main content or the whole page; zoom; pick a tool; find sensitive text of a kind and hide it by redaction, blur or erase; find words; undo and redo; copy the crop as a picture; and export. It does it with the panels' own controls, so every change shows in its panel and on the stage. It does not draw marks, move the crop by a number, or set colors, the stamp's time, a brand, a watermark or a preset. It cannot see the page, so it answers nothing about what the page says.

Every message is read by a model on Page Scanner's server: 10 messages a month for each install, free, each counted once however many steps it takes. With each step the server is sent the conversation so far, the editor's choices, and counts such as "2 email addresses listed"; never the page's text, pictures, address or title. A message that fails before the model reads it is not counted. Nothing is read on your computer, so the Assistant needs the network, and Page Scanner's server switched on in Settings › AI Model. The first line of each answer says how many of this month's messages are left: "Read by Page Scanner's server · 7 of 10 left this month".

The conversation lasts as long as the editor tab, and a new capture starts a new one. A capture scanned from the side panel opens with the Assistant up and what you said there in it.

When a message would not be read, a red note at the end of the conversation says so before you send it: offline, this month's messages used, the server unable to take requests, or the server switched off or closed for now.

While the server reads a step, the seconds are counted and Stop (or Esc) stops it. When it cannot read one, the message stops where it was, and a note says why:

When the capture was cut short

A page taller or wider than 60,000 CSS pixels is captured up to that point and no further, unless Maximum page height in the settings lets it go taller. The message that confirms an export then says so first: what the page measured and how much of it is missing, so a partial page never passes for a whole one.