How work moves across the web | Composite Research

AI assistants now take 14% of the active time in our panel, more than four times what goes to search engines (3.2%). Rather than acting as an end destination however, AI assistants serve as a reference during tasks. When someone leaves Coursera, 19% of the time they go to ChatGPT, and 84% of those trips come straight back. The same ping-pong effect shows up next to Canva, Upwork, Google Colab and university course portals.

Here we map that kind of movement across the sites where people do their work. Every data point comes from real browser sessions. The charts below let you look up your own site or role.

Where do people go next?

The map below puts 700 of the most-used sites on one page. The more people use a site, the closer it sits to the center. Each slice is one kind of site, from AI assistants to shopping, and within its slice a site leans toward the kind of site its visitors go to unusually often. 79% of moves between these sites cross into a different kind of site. Hover any dot for more information.

Tap any dot to see where people go next from that site.

Search & browsing
Learning & research
News & portals
Shopping, money & travel
Business & jobs
Docs & productivity
Email, chat & meetings
Building software & IT
AI assistants
Design & creative
Social, video & games

AI assistants: 82% of moves from these sites go to a different kind of site, most often Social, video & games or Email, chat & meetings.
Google Colab → ChatGPT: 30% of moves out of Google Colab go to ChatGPT, and 84% of those trips come straight back.
Meesho → WhatsApp: 21% of moves out of Meesho go to WhatsApp, and 70% of those trips come straight back.

Hover or tap any dot to see where people go next from that site; click to keep its card open and load it in the explorer below.

The explorer below zooms in on a single site. Search for a site to see where its visitors arrive from and where they go next, with each neighbour’s share of moves and how long people stayed before moving on. Google, YouTube and ChatGPT top almost every list, so each neighbour also carries a lift: how many times more likely the move is than the destination’s overall popularity predicts. A few patterns stand out:

Product pairs that behave like integrations. People leaving Linear go to GitHub 14% of the time, 26× more often than chance. SharePoint hands people to Excel at 60× the baseline rate.
AI as the side panel. For tools like Colab and Canva, most trips to ChatGPT come straight back: 84% of Colab-to-ChatGPT trips return straight to Colab, and 76% of Canva-to-ChatGPT trips return to Canva.
Short hops, not context switches. The median time on Coursera before jumping to ChatGPT is 24s; on Google Docs it is 26s.

Pick any site

Around ChatGPT

Shares are of moves into and out of this site. Lift compares with how often people move to that site from anywhere.

Came from

Google Search 7.4% 0.8× 23s there first
WhatsApp 6.2% 1.4× 40s there first · 60% come straight back
YouTube 5.7% 1.0× 55s there first
Gmail 4.8% 1.1× 37s there first
Google Docs 4.0% 2.3× 26s there first · 73% come straight back
Facebook 2.7% 1.1× 49s there first
Gemini 2.7% 1.4× 37s there first
Claude 2.7% 1.8× 45s there first
Canva 2.6% 3.2× 38s there first · 76% come straight back
Google Sheets 2.1% 0.7× 27s there first · 61% come straight back
LinkedIn 1.2% 1.4× 40s there first
Outlook 1.0% 1.3× 42s there first
Google Drive 1.0% 0.7× 38s there first
Google Translate 0.9% 2.7× 33s there first
Instagram 0.9% 0.9× 39s there first
GitHub 0.7% 1.4× 31s there first
Google Colab 0.6% 5.0× 24s there first · 84% come straight back
Pinterest 0.6% 1.9× 48s there first
NotebookLM 0.5% 1.5× 48s there first
Google Forms 0.5% 1.3× 23s there first · 75% come straight back

Went next

Google Search 8.6% 1.0× after 1m 40s
WhatsApp 6.3% 1.4× after 1m 10s · 60% come straight back
YouTube 5.7% 1.0× after 1m 7s
Gmail 4.9% 1.1× after 59s
Google Docs 4.0% 2.2× after 26s · 74% come straight back
Gemini 2.8% 1.4× after 41s
Canva 2.7% 3.3× after 24s · 73% come straight back
Claude 2.7% 1.9× after 51s
Facebook 2.7% 1.1× after 58s
Google Sheets 2.1% 0.7× after 33s · 62% come straight back
LinkedIn 1.1% 1.3× after 45s
Outlook 1.0% 1.3× after 57s
Google Drive 1.0% 0.7× after 1m 10s
Google Translate 0.9% 2.7× after 1m 2s · 67% come straight back
Instagram 0.8% 0.8× after 51s
GitHub 0.7% 1.4× after 56s · 64% come straight back
Google Colab 0.6% 5.0× after 25s · 84% come straight back
Pinterest 0.6% 1.8× after 1m · 60% come straight back
NotebookLM 0.5% 1.5× after 54s
Google Forms 0.5% 1.3× after 19s · 74% come straight back

Unexpected next stops

Pairs where people go from one site to the other far more often than the second site's overall popularity would predict.

takeUforward → LeetCode 791× 35% of moves out of takeUforward
NeetCode → LeetCode 504× 25% of moves out of NeetCode
Dzen → Yandex 346× 23% of moves out of Dzen
Visual Studio → Azure 377× 20% of moves out of Visual Studio
Wildberries → Ozon 648× 15% of moves out of Wildberries
Tokopedia → Shopee 124× 23% of moves out of Tokopedia
Rutube → Yandex 221× 16% of moves out of Rutube
Dhan → TradingView 152× 18% of moves out of Dhan
Meesho → Flipkart 267× 14% of moves out of Meesho
AliExpress → eBay 157× 17% of moves out of AliExpress

Ping-pong pairs

Share of moves from A to B that come straight back to A: people working across two tabs at once.

Character.AI ⇄ Google Translate 93%return straight back
GitHub Codespaces ⇄ DeepSeek 92%return straight back
Webflow ⇄ Figma 91%return straight back
NPTEL ⇄ Gemini 87%return straight back
n8n ⇄ ChatGPT 86%return straight back
AppSheet ⇄ ChatGPT 85%return straight back

The shape of the web at work

A handful of sites take a large share of the time people spend online. The 10 most-used sites account for 41% of all active time and the top 100 for 66%, but 22% of time spent online still go to sites ranked below 1,000: industry portals, niche software and regional services that rarely appear in top-sites lists.

41%of active time goes to the 10 most-used sites

66%goes to the top 100

22%goes to sites ranked below 1,000

How much time goes to those smaller sites depends a lot on the category. Almost all time on communication, productivity and social sites goes to the top 1,000, while 38% of personal-finance time and 35% of shopping time is spent on sites outside it. The chart below ranks every category by that share.

How much of each category happens on smaller sites

Share of the category's time spent on sites ranked below 1,000.

Sports 71%
Other tech & IT 66%
News 66%
Education & reference 58%
Travel & maps 56%
Gaming & hobbies 49%
Government & civic 48%
Business & sales 41%
Personal finance 38%
Shopping 35%
Jobs & careers 32%
Health 31%
Science & research 7%
Video & entertainment 5%
Developer & cloud 5%
Design & creative tools 3%
AI assistants & tools 2%
Productivity & docs <1%
Communication <1%
Search <1%
Social networks <1%

Each category also has its own weekly rhythm. In each person’s local time, work categories stay busy from Monday to Friday and drop off sharply at the weekend, while video peaks on Sunday evening and sports on Saturday afternoon.

When each kind of site is busiest

Local time, Monday to Sunday. Darker = closer to that row's busiest hour.

To see what people actually do on a site, pick one below. Every session is summarized into workflows, and each workflow gets one plain-language type.

What people do on Google Sheets

Spreadsheets & data83%
Project & task management2%
Research & reading2%
Sales, marketing & support2%
Meetings & calendar2%
Finance & money1%

Share of the workflows that touched the site.

A breakdown of tools by role

We grouped people into ~50 role archetypes built on the US Standard Occupational Classification, from software developers to online sellers, and compared what each role opens with what everyone opens. Ranked by lift rather than raw share, the signature tools read like job descriptions: Google Ads and Search Console for marketers, Seller Central and Shopify for online sellers, Zendesk and Freshdesk for support agents, Databricks and Colab for data scientists, journal submission systems like Editorial Manager and ScholarOne for researchers.

AI use is uneven. 66% of UX and product designers and 64% of marketers used ChatGPT in the window, against 25% of healthcare workers. Official skill lists have not caught up: O*NET, the US Labor Department’s occupation database, lists ChatGPT for 0.7% of occupations.

Role Compare with

Software developer

SOC 15-1252 · Software Developers

8.1% of people with an inferred role · 55% used ChatGPT

Signature tools share of people · lift vs. everyone

Cursor5%8.4×
Stack Overflow4%8.4×
Azure4%8.2×
AWS6%8.1×
OpenRouter4%7.8×
Jira9%7.6×
Vercel6%7.5×
Supabase5%7.5×

Kinds of work share of workflows

Coding & IT15%3.6×
Research & reading10%0.8×
AI assistants7%1.2×
Video & streaming7%0.7×
Project & task management5%2.1×
Messaging & chat5%0.7×

Daily rhythm local hour, vs. everyone (grey)

Marketer

SOC 13-1161 · Market Research Analysts and Marketing Specialists

3.8% of people with an inferred role · 64% used ChatGPT

Signature tools share of people · lift vs. everyone

Kinds of work share of workflows

Research & reading10%0.7×
Sales, marketing & support9%5.1×
Messaging & chat8%1.2×
AI assistants8%1.4×
Social media8%1.5×
Video & streaming6%0.6×

Daily rhythm local hour, vs. everyone (grey)

Each role's two most distinctive tools, in role order, so each role's signature runs down the diagonal and shared tools show up off it. Darker cells mean the role is many times more likely than everyone to use the tool (lift, capped at 30×).

What O*NET says vs. what the browser shows

O*NET, the US Department of Labor's occupation database, lists the "hot technologies" for each occupation. It names ChatGPT for 0.7% of occupations. In our data:

Role	Role-specific tools O*NET lists	Most distinctive tool we see	Used ChatGPT
Academic researcher / scientist	ESRI ArcGIS software, SAS, IBM SPSS Statistics	Editorial Manager	49%
Software developer	Eclipse Jersey, GitLab, Atlassian Bitbucket	Cursor	55%
Business owner / manager	Kronos Workforce Timekeeper, HubSpot software, Mozilla Firefox	Shopify	44%
Teacher / educator	Microsoft SharePoint	Padlet	40%
Marketer	HubSpot software, WordPress, Marketo Marketing Automation	Google Ads	64%
Healthcare worker	eClinicalWorks EHR software, MEDITECH software	Linktree	25%
Administrative assistant	SAP Concur, Canva, Google Sheets	iLovePDF	32%
Project manager	Asana, Procore software, Google Workspace software	Google Chat	47%
Graphic designer	Figma, WordPress, AJAX	Behance	66%
University faculty	Google Docs	Turnitin	36%
Systems / DevOps admin	Border Gateway Protocol BGP, Red Hat OpenShift, Microsoft Active Directory	Google Admin	38%
Logistics / supply chain	Oracle Primavera Enterprise Project Portfolio Management, Microsoft Power BI, Microsoft SQL Server	SharePoint	33%

O*NET's column leaves out the office software it lists for most occupations (Excel, Word, Outlook, PowerPoint and Office). Desktop apps are invisible to a browser recording, so a desktop tool missing from our column doesn't mean nobody uses it.

How we did it

For this post we used a large sample of recent real browsing sessions and built four layers on top of them:

An attention timeline. For each session we rebuilt what the person was looking at, tab by tab, carrying the current page forward until they navigated or switched tabs. Sign-in screens, redirects, new-tab pages and asset URLs are not counted as visits, and visits shorter than three seconds are skipped.
Products, not hostnames. Every hostname is mapped to a product: customer-specific subdomains collapse to the product they run on (a company’s Jira site is just Jira), Google Docs is split into Docs, Sheets and Slides, and Microsoft 365’s many hosts become Outlook, Teams, Word, Excel and so on.
Workflows and types. Each session is summarized into workflows (“Triaged email inbox”, “Reviewed a pull request”), and each workflow gets one of 28 plain-language types.
Roles. Each session also carries a short inferred description of the person’s likely role. We map those descriptions to role archetypes anchored in the Standard Occupational Classification and give each person their most frequent archetype.

Sites are categorized with the IAB Tech Lab Content Taxonomy 3.1. Because version 3 no longer treats news as a topic, general news publishers use a clearly labelled extension category.

Methodology and limitations

This is our population, not the internet. The panel is not spread evenly across the world. Treat every share as describing this population, not web users at large.
Browser only. Desktop applications, phones and anything outside the browser are invisible here.
Roles are inferred from behaviour. Because roles are inferred partly from what people do, “developers use GitHub” is partly circular. That is why we rank signature tools by lift and show kinds of work and daily rhythm rather than the role-defining sites alone. About 76% of people received a role.
Definitions. Shares of moves count consecutive visits on the attention timeline, with self-transitions removed. Lift is smoothed as (observed + 5) / (expected + 5), so thinly supported pairs drift toward 1. Pairs with little support are marked “limited data”. Active time counts only time with user interaction. Daily reach is the share of person-days on which a site appears. Local time uses each person’s time zone.
What we do not show. We publish shares, ratios and ranks only. A site is named only when enough people across more than one part of the panel used it; smaller sites are counted in every total but not named. Each listed pair of sites must pass the same test, so for some sites the lists cover only a small share of their moves. Adult, gambling and piracy sites appear only as one aggregate category. Company-internal tools, employer portals and single-program systems are left out so the data doesn’t point to anyone’s employer.
Map slices. Each slice groups sites by kind: our grouping of the IAB content categories into about ten plain-English groups, with every site on the map checked by hand. Within a slice, a site’s angle leans toward the slice its visitors go to more often than that slice’s overall popularity predicts (the same smoothed lift as above; sites with too little data sit mid-slice), and its distance from the center is its popularity rank. Government and health sites, country-wide groupings (such as Indian bank portals), sites without a daily-reach figure (such as Google Maps) and sites we couldn’t categorize confidently are left off the map; the explorer still lists them.
Coverage. Flows and active time use the browser recordings of every session in the window. Daily reach, workflows and roles use the subset of those sessions that our intelligence layer had fully processed, so they can differ slightly from the flow figures.

Work with this data

Composite’s data partnerships are built on this kind of material: real browser work, summarized into workflows, categorized and attributed to roles. If you build agents, evaluate them, or want to know where your own users go next, see composite.com/data or write to research@composite.com.