Product page data scraped from screenshots by a Gemini agent

Product page data scraped from screenshots by a Gemini agent

An n8n workflow that reads a list of page addresses from a Google Sheet and takes a full-page screenshot of each through ScrapingBee. An AI agent on Gemini reads the products, prices, brands and promotions off the picture. When the picture is not enough, the agent calls a second route that fetches the page HTML as Markdown. The result becomes rows in a Results tab.

What is inside

• workflow.json: the workflow as n8n exports it, twenty-nine nodes: fourteen that do the work and fifteen sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The credential links are gone. Select your own for the Google service account and Gemini.
• The model was models/gemini-1.5-pro-latest. Google's models and deprecations pages, read on 12 November 2026, list no Gemini 1.5 model at all. The node now names models/gemini-3.8-flash, which the models page lists as Stable with image input and structured outputs.
• The ScrapingBee key was already a placeholder, <your_scrapingbee_apikey>. It is <YOUR_SCRAPINGBEE_API_KEY> now in both requests.
• The screenshot request sent only screenshot_full_page. ScrapingBee's documentation lists screenshot and screenshot_full_page as separate options, and its own example sends both, so this copy sends both.
• HTML-based Scraping Tool held the id of the author's own workflow. It is blank now. After you import, select this same workflow in it.
• The sheet id is <YOUR_SHEET_ID>, with the tabs List of URLs and Results.
• The Create Rows node wrote the product title into a column named product_url. The column is product_title now.
• The agent's prompt ended with a stray Copy code line and an example that used a true or false value and a number under promo_percent. The parser's example uses text values and the key promo_percentage. The prompt's example now matches the parser.
• The notes are rewritten. One of them linked to the author's sample sheet, and that link is removed.
• Nothing else.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• A ScrapingBee account and API key. Its documentation says a request that renders JavaScript, which is the default, costs 5 credits.
• A Google service account with your sheet shared to its address, and a Gemini API key.
• A Google Sheet with two tabs. List of URLs has the header url in row 1 and one address per row. Results has the headers category, product_title, product_price, product_brand, promo and promo_percent.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Google service account credential and select it in Google Sheets - Get list of URLs and Google Sheets - Create Rows. Replace <YOUR_SHEET_ID> in both and check the tab names.
3. In ScrapingBee- Get page HTML and ScrapingBee - Get page screenshot, replace <YOUR_SCRAPINGBEE_API_KEY> with your key.
4. Add a Gemini credential and select it in Google Gemini Chat Model.
5. In HTML-based Scraping Tool, choose this same workflow in the Workflow field.
6. Add a few addresses to List of URLs and click Test workflow.

How a run goes

1. The manual trigger starts Google Sheets - Get list of URLs, which reads the rows.
2. Set fields keeps the url of each row.
3. ScrapingBee - Get page screenshot calls ScrapingBee with the address, screenshot and screenshot_full_page set to true and a browser-style User-Agent header. The picture comes back as a file, with the row's JSON beside it.
4. Vision-based Scraping Agent receives the picture and a message with the address. Its system message asks for the titles, prices, brands and promotions, and says to call the HTML tool only when something is missing.
5. The tool starts HTML-Scraping Tool in this same workflow with the address as query. Set fields - from AI agent query turns it into url, ScrapingBee- Get page HTML fetches the page, and HTML to Markdown converts it. The data field goes back to the agent.
6. Structured Output Parser shapes the answer as a list of products, and the agent returns it as output. Split Out makes one item for each product.
7. Google Sheets - Create Rows adds a row for each. category holds the page address, and promo_percent takes the value of promo_percentage.

Change the look

• What to extract: the system message of the agent and the example in Structured Output Parser.
• The columns: the mapping and schema in Google Sheets - Create Rows.
• The model: Google Gemini Chat Model.
• The trigger: replace the manual one with a schedule.

Good to know

• Scraping may be against a site's rules. Check the terms of every site you list and the law where you live. ScrapingBee's terms, read on 12 November 2026, also bar use that breaks a scraped site's own terms.
• Each page costs one ScrapingBee request for the screenshot and a second one when the agent needs the HTML, plus the Gemini calls.
• The agent can misread a price, so check the Results tab.
• This copy was not run against ScrapingBee, Gemini or Google Sheets.
• The page addresses and screenshots go to ScrapingBee and to Google's Gemini, and the rows go to your sheet.
• The JSON holds no credential.