Reading: how people use AI to consume docs

Data on how people use agents to read documentation on their behalf.

How people use agents to consume documentation.

Mintlify’s embedded AI assistant lets users ask questions and get answers directly from a site’s documentation. This is the layer between human users and documentation pages. Data from April 21, 2026.

Scale

Metric	Value
Total requests	95k
Unique sessions	29k
Unique documentation sites	1.2k

Failed sessions and abusive sessions are not included.

Message types

Most agent-mediated documentation interactions are not questions. Statements and commands dominate.

Message type	Share
Statement	62%
Question	30%
Troubleshoot	5%
Command	3%

First message intent

Intent	Share
How to	30%
Topic lookup	16%
Capability check	14%
Conceptual	12%
Troubleshoot	10%
Other	15%
Show example	1%
Why	1%

Message length

Most users are terse.

Metric	Value
First messages under 15 words	71%
First messages 1-5 words	32%
Median user message	9 words
Average user message	46.9 words
Median assistant response	10 words
Average assistant response	29.6 words

Verbosity by turn

Message length isn’t stable across a conversation.

Turn	Avg user message length
1	~50 words
2	~32 words
3	~43 words
6-7	70-150 words (frustration signal)

The spike at turns 6 and 7 is a key insight: users who persist through multiple failures eventually dump everything they know into one message. Intervening at turn 5 is a design opportunity.

Conversation length

Conversation length

Turns per session	Sessions
1	14,443
2	5,898
3	2,868
4	1,593
5	999
6	669
7	449
8	388
9-14	730
15+	391

Half of sessions end after a single message.

Frustration rate

The frustration rate is the share of sessions where the last user message was negatively-toned. It rises with conversation length.

Turn	Frustration rate
1	5%
2	6%
3	7%
4	8%
5-7	9%
8	19%
9-11	11%
12	15%
13	11%
14	10%
15+	14%

The spike at turn 8 (19%) is a key signal. Intervening at turn 5 or 6—before the spike—is a design opportunity.

Sentiment arc in multi-turn sessions

Among 3,712 sessions with 3 or more turns, four trajectory shapes emerge:

Shape	Sessions	What it means
Flat	1,524	Task-oriented, neutral throughout — probably successful
Rising	1,086	Starts frustrated, improves — recovery
Falling	1,023	Starts okay, ends frustrated — failure pattern
U-shape	79	Dips in the middle, recovers — successful after struggle

1,023 falling-arc sessions out of 3,712 multi-turn sessions (27.5%). More than one in four multi-turn conversations ends worse than it started.

Sentiment overall

From a sample of ~3,300 records (keyword-based, directional):

Sentiment	Share
Neutral	77.9%
Positive	15.4%
Negative	6.7%

Response time

Metric	Value
Average response time	3.7 seconds
Median response time	2.9 seconds
Average time to first token	1.8 seconds

6 errors out of 94,668 requests.

Benchmark

One-and-done: 55%

55% of sessions end after a single message.

How to measure

What you have access to depends on which agent platform you use. At minimum, look for:

• Session length distribution (how many one-and-dones?)
• Repeat usage (do users come back?)
• Explicit feedback if available (thumbs up/down)
• Turn-level sentiment if you can get it