Deep web research that writes its report into Notion

Deep web research that writes its report into Notion

An n8n workflow that researches a question in rounds. A form takes the question, a model asks clarifying questions, and a loop
searches the web, reads the top pages and writes down what it learns. A reasoning model then writes a long report, and the
workflow saves it to a Notion database page.

What is inside

• workflow.json: the workflow as n8n exports it, 86 nodes, 20 of them sticky notes. It holds no credential. It starts its own
  sub-runs through three Execute Workflow nodes that call this same workflow, so save it before you run it.
• README.md: these notes.

What we changed from the original copy

• Every credential link (OpenAI, Google Gemini, Notion and Apify) belonged to the author's n8n. The links are gone, so select
  your own.
• Five OpenAI nodes used o3-mini, which OpenAI shut down on 23 October 2026. They now use gpt-5.6-sol, the replacement
  OpenAI names.
• The Gemini node used models/gemini-2.0-flash, whose shutdown Google set for 1 June 2026. It now uses
  models/gemini-3.6-flash, the replacement Google names.
• The search actor was serping~fast-google-search-results-scraper, which Apify now answers with "not found". Web Search
  now runs Apify's own apify~google-search-scraper. Its request body uses queries and maxPagesPerQuery where the old one
  used searchTerms and resultsPerPage, and Top 5 Organic Results reads organicResults with title and url where it
  read origin_search.results. Web Search used a header credential, and both Apify nodes now use the same query-token one.
• The two sliders in the question form are HTML fields without an element name. Current n8n leaves the value of such a field
  out of the form's output, so Set Variables always fell back to depth 2 and breadth 4. The fields now have the element names
  depth and breadth. Set Variables reads them, with the sliders' own starting values of 1 and 2 as the fallback.
• The Notion database was the author's own, with its id, name and link in Create Row, Get Existing Row and
  Get Existing Row1. They now hold <YOUR_NOTION_DATABASE_ID>.
• Six sticky notes were rewritten, because they held the author's Notion database link, a referral link for Apify, a chat
  invite or the retired model name. One note that held only the author's picture is gone, and so are the pictures the author
  hosted in the question form, the clarifying form and the confirmation form.
• The export's instance id and the webhook ids are gone.
• In the page function of Page Contents, the jQuery handle is renamed from $ to jq. It does the same job and keeps n8n
  from reading $('title') as a reference to a node.
• Nothing else.

Before you start

• An n8n instance, on n8n Cloud or your own server. The form must be reachable by whoever uses it, so publish the workflow.
• An OpenAI API key.
• A Google Gemini API key.
• An Apify account with an API token and some credit.
• A Notion account, a Notion integration, and a database with the five properties named in the steps.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. In Notion, create a database with these properties: Name (title), Description (text), Status (status, with the options
   Not started, In progress and Done), Request ID (text) and Last Updated (date). Share it with your Notion integration.
3. Paste the database's id into Create Row, Get Existing Row and Get Existing Row1, replacing <YOUR_NOTION_DATABASE_ID>.
   Add a Notion credential and select it in those three nodes, in Set In-Progress and Set Done, and in
   Upload to Notion Page.
4. Add an OpenAI credential and select it in the five OpenAI Chat Model nodes, OpenAI Chat Model to OpenAI Chat Model4.
5. Add a Google Gemini (PaLM) API credential and select it in Google Gemini Chat Model.
6. Add a Query Auth credential with the name token and your Apify API token as the value. Select it in Web Search and
   Page Contents.
7. Save the workflow and publish it. The form's address is the production address shown in On form submission.
8. Open the form. Ask a question, keep depth and breadth low for a first run, tick the box about time and cost and answer
   the clarifying questions. The report appears on a new page in your Notion database when the run is done.

How a run goes

1. The form opens from On form submission with a button. Then Research Request asks for the question, a depth slider from 0
   to 3, a breadth slider from 1 to 5 and a box that accepts the time and cost. After that, Set Variables keeps the answers and
   the id of this execution.
2. Clarifying Questions asks the model for up to three questions. Feedback to Items, For Each Question... and
   Ask Clarity Questions show them one at a time and Get Initial Query joins the answers to the question.
3. Report Page Generator writes a title and a short description, and Create Row makes the Notion page with the status
   Not started and the request id.
4. Initiate DeepResearch starts a separate run of this workflow and does not wait for it. Confirmation and End Form close
   the form with a link to the Notion page, so the user can leave.
5. The new run enters at DeepResearch Subworkflow. Execution Data saves the request id and the job type, and JobType Router
   sends it down one of three paths. The first sets the page to In progress and starts the research.
6. Each round calls this workflow again for the learnings path. Generate SERP Queries asks the model for up to as many search
   queries as the breadth, each with a research goal. For each query, Web Search runs a Google search on Apify and
   Top 5 Organic Results keeps the first five links.
7. Page Contents runs Apify's Web Scraper on each link with a 90 second limit. Convert to Markdown turns each page into
   text. DeepResearch Learnings reads up to 25,000 characters of each page. It returns up to three learnings and up to three
   follow-up questions.
8. Accumulate Results counts the rounds. While the round count is below the depth, Set Next Queries turns the follow-up
   questions into new queries and the loop runs again. A depth of 0 stops after the first round, a depth of 1 after two.
9. When the depth is reached, Generate Report starts the report path. Then DeepResearch Report writes the report in Markdown
   from all the learnings. Next, Convert to HTML and HTML to Array split it into pieces. After that, Notion Block Generator
   asks Gemini to turn each piece into Notion's block format. Last, URL Sources to Lists builds a Sources list by hand.
10. For Each Block... sends the blocks to the page through Upload to Notion Page, and Set Done sets the status to Done with
    the time.

Change the look

• How far it goes: the depth and breadth sliders in Research Request.
• What the model asks and how it writes: the prompts in Generate SERP Queries, DeepResearch Learnings and DeepResearch Report.
• The models: the OpenAI nodes write the questions, the queries, the learnings, the report and the title. Gemini formats the
  Notion blocks.
• The searches and scrapes: the Apify nodes, or swap them for another search or scrape service.
• Where the report goes: replace the Notion steps with the wiki tool you prefer.

Good to know

• Time and cost: the file's own notes give about 10 to 15 minutes for depth 1 and breadth 2. They give 30 to 40 minutes for depth
  1 and breadth 4, and 2 to 5 hours for depth 3 and breadth 5. They were written for the old model. Each round multiplies the number of
  searches, page scrapes and model calls, so raise the sliders slowly.
• Models: OpenAI's deprecations page, read on 12 November 2026, lists o3-mini for shutdown on 23 October 2026 with gpt-5.6-sol
  as the replacement. Google's page, read that day, lists gemini-2.0-flash for shutdown on 1 June 2026 with gemini-3.6-flash
  as the replacement. Both swaps are untested here, so run a shallow request first.
• Search actor: Apify's store listing for apify~google-search-scraper, read on 12 November 2026, prices it per page of results
  scraped, from $0.0045 on its free tier, and Web Search asks for one page for each query. The scraper's own input and output
  fields were read on Apify's site and not run from here.
• Sources list: URL Sources to Lists cuts the links into chunks of 50 but takes 49 from each chunk. It ignores any leftover
  past the last full chunk, so a run with 50 or more links leaves some out of the list.
• The form is public once the workflow is published: anyone who has its address can start a run and spend your OpenAI, Apify and
  Gemini credit. The consent tick box protects nothing.
• Your data: the question and the answers go to OpenAI, the search queries and the links go to Apify, the page text goes to OpenAI,
  the report goes to Gemini for formatting and to Notion.
• The workflow calls itself, so a copy saved under a new name works, but an unsaved one does not.