WebMemo

Team research at the University of Michigan School of Information

WebMemo is a Chrome extension for collecting details from several websites into tables. While a person reads, GPT-4o fills in the rows, returning JSON for the table’s columns. Clicking a value opens the page it came from, scrolls to the passage and highlights it.

I designed and built the extraction and provenance workflow, and revised the prompts, schemas and interface around the failures we observed.

Try it

Pick a case, then a field, to see the words that support it. The three examples are made up for this page and are not live model output.

Each field and its sentence

Illustrative source 1

An afternoon of reading

Hillside Library invites you to an open reading session.

Join us on Saturday, 17 October, starting at 14:00.

Admission is free. No reservation required.

Extracted fields

Date17 October Start time14:00 AdmissionFree

Each field links to the sentence it came from.

When the source says nothing about the price

Illustrative source 2

The evening session

A second reading session is planned at Riverbank Reading Room.

Doors open at 18:30.

Admission details will be announced separately.

Extracted fields

VenueRiverbank Reading Room Doors open18:30 AdmissionNot stated

The source does not state a price. The result stays “Not stated” until there is evidence.

An outdated time, and the notice that replaces it

Illustrative source 3

A revised opening time

Original notice: the Saturday reading begins at 10:00.

Update: the reading now begins at 11:00. This replaces the time in the original notice.

The venue and admission policy are unchanged.

Prepared result, needs review

Start time10:00

The corrected value is 11:00. The update explains why it takes precedence over the original notice.

Where the design came from

The design started from a study of 24 people and the web tasks they do often¹. Going back through the 150 tasks they had proposed themselves, we matched what people asked for to what the extension would do.

In the interviews	In WebMemo
87 of the 150 tasks needed information from more than one site.	Rows from different pages go into tables that can be made or changed at any time.
People pictured “an extension” or “a small window”, and worried that opening new sites would “increase mental load”.	It stays in the browser and adds rows one at a time while the person scrolls.
They wanted AI to gather the material but to confirm the final step themselves.	Each row links back to the passage it came from and can be checked as it arrives.

Compared with OttoGrid

OttoGrid, a table tool with AI assistance, was the comparison point in the study. The two tools make different choices in three places².

	OttoGrid	WebMemo
Tables	set up at the start	made or changed at any time
Rows	added in a group	added one at a time
Checking	at the end	as you go

What the study found

Twelve people used both tools, one tool per task, with the order counterbalanced³. All twelve finished their WebMemo task; with OttoGrid, one hit the 30-minute limit. They rated WebMemo’s results as more trustworthy: a mean trust rating of 4.42 against 3.33 (p = .027), on a scale the draft does not give.

Mean ratings on the draft’s 5-point scale, where higher is better.
Statement	WebMemo	OttoGrid	p
Collects data without interrupting browsing	4.92	3.33	.005
Organizes page content into a structure	4.67	4.42	.54
Less to remember across sources	4.75	4.25	.22
Saves time over collecting by hand	4.67	3.67	.039

Mean completion times were lower with WebMemo in both tasks: 223 against 315 seconds for the researcher task and 802 against 911 for the shopping task. The draft reports no test for the times. One participant put the trust result this way: “dynamically showing rows makes me understand that WebMemo is reading the website.”

The draft also says what WebMemo does not do yet. It does not flag entries that are probably wrong, and it does not learn from a person’s corrections.