Did the Recommendation Hold Attention or Satisfy the Reader?

Evaluate AI engagement signals alongside explicit feedback, access and the people a metric leaves out.

A museum worker compares a visitor card with a display of aggregate charts.
Understanding a visit takes more than counting the minutes a person spent there.

An audience AI model can rank articles or videos according to an expected outcome. A long visit may mean someone found a detailed answer, left a tab open or struggled to find a transcript. The choice of outcome shapes what the system learns to show next.

Specify what success means for this audience

A woman in a futuristic setting gazes at an orange and yellow digital face, reminiscent of a computer-generated avatar or hologram.
A quick exit after finding the correct opening time can be a successful visit.

Define the reader's task before choosing a target. YouTube describes using several signals in recommendations, including surveys to understand satisfaction alongside viewing behavior. A publisher can ask whether the reader found the answer or would recommend the piece, while tracking effort, reach and return visits separately. YouTube on its recommendation and satisfaction signals

Account for whose feedback is visible

A young woman is intently looking at a computer screen while sitting in front of two monitors and various objects on her desk.
People who fill out a survey may differ from people who move on.

Log when feedback was offered, who responded and which surfaces could collect it. Distinguish a missing response from a negative one. Segment descriptive results by device or content format where sample sizes permit, and compare recommended content with a simpler editorial list in a controlled trial before attributing differences to ranking.

Check whether the format creates artificial engagement

An elderly woman sits in a wheelchair, holding a tablet with a garden scene on the screen.
A difficult interface can increase time spent while reducing the chance of completing a task.

Captions, transcripts, keyboard navigation and descriptive links change what a reader can do. Test task completion with those options available. Evaluate whether the recommender returns diverse accessible formats, and collect an explicit preference when a person chooses text or audio. Use engagement data to ask about friction rather than guessing at an individual's motives.

Try it with the evidence

Read a misleading engagement headline

Source notes. Fictional page. Of 200 visitors, 50 answered “Did this answer your question?” and 40 said yes. Median time on page rose after a transcript link was moved. No other feedback or comparison group was collected.

Prompt. Report the satisfaction result with both denominators, explain the missing evidence and propose a test involving the transcript link.

Check your reading

Eighty percent of respondents said yes. Twenty percent of all visitors explicitly said yes, with 150 visitors' answers unknown. Longer visits could signal interest or difficulty. Compare link placements using task completion and a consistently offered satisfaction question.

Illustrated field notes

Recommendations and audience signals

Identify who had a chance to respond and what a click or a pause can actually show.

An illustration of a computer monitor surrounded by digital tools and symbols associated with sentiment analysis.
Inspect the target of the opinion before labeling the audience.
A person is intently working on a laptop in a dimly lit room, surrounded by graphs and charts on the screen.
Specify which future behavior the forecast is intended to predict.
A computer monitor displaying a dashboard for data analysis or stock market tracking in an office setting.
Separate audience reach from evidence of a causal advertising effect.
Access and participation

Check whether the person could find a usable format and complete the intended task.

A woman interacts with a large screen displaying various packages.
A support conversation can be long because a problem remains unresolved.
A woman is intently looking at a computer monitor displaying stock market data.
A heavily shared post can distort an average across channels.
Email Marketing Campaign Template.
Count unsubscribes and undelivered messages alongside clicks.
A person holding a smartphone with a white screen displaying trees and text in a foreign language.
Repeated actions in a game may reflect the design of the task.