CitationLab
Analysis

How commercial intent develops in AI conversations

An analysis of 150,000 real AI conversations

Most assumptions about how users make purchase decisions in AI search are wrong. We analysed 150,000 real conversations to find out what actually happens when commercial intent enters an AI conversation.

Published March 2026 by CitationLab AS

The dataset at a glance

150,000
conversations analysed
6.8%
with commercial intent
2.92
avg turns (transactional)
48%
single-turn conversations

What do people actually use AI for?

Before examining commercial conversations, we mapped the full distribution of conversation types. Transactional conversations are a tiny minority of AI usage.

Other
66.8%
Writing
12.6%
Coding
12.3%
Information
3.4%
Creative
1.9%
Transactional
1.8%
Analysis
0.8%
Math
0.4%

Seven key findings

6.8%

Commercial conversations are a minority

The vast majority of AI conversations are about coding, writing, and information retrieval. Under 7% have commercial intent.

31%

Price is the strongest buying signal

When a user asks about price, it triggers a cooling event 31% of the time. Price questions are buying signals that AI consistently fails to convert.

-69 points

Price triggers sharp cooling

Average intensity drop after a price-related cooling event is -69 points. The user goes from ready-to-buy to back at square one.

91%

Users fall back to TOFU after AI hedging

Over 90% of post-price cooling events result in regression to Problem Focus. The AI answer fails to close the decision gap.

41%

Many buyers start at BOFU

Users don't follow a discovery arc. Over 4 in 10 already know what they want when they open the conversation.

4.4%

The progressive arc is rare

Only 4.4% of transactional conversations follow the traditional TOFU-to-BOFU progression. The linear funnel is a myth in AI search.

2.92 turns

Transactional conversations are short

Commercial conversations are brief and action-oriented. Nearly half are single-turn. Users want answers fast.

The intensity arc

Average commercial intensity across conversation turns. Note the sharp drop from Turn 1 to Turn 2, and the flat TOFU plateau from Turn 2 onwards.

BOFU
MOFU
TOFU
T1T2T3T4T5T6T7T8T9T10

The post-price collapse

Price questions signal buying intent. But the AI's answer typically fails to close the deal, sending the user back to square one.

31%
Cooling rate after price signal
-69 pts
Mean intensity drop

What happens after a price-triggered cooling event?

Regression (silent)
89.5%
Frustration
7%
Uncertainty
3.5%

Why does this happen?

The failure is not the price itself. It is the precision of the AI's answer. Three response patterns consistently trigger user regression.

Low closure

Vague estimate

AI gives a price range without helping the user navigate it.

'Prices range from X to Y depending on location and condition'

Medium closure

Hedged answer

AI answers correctly but adds a caveat that opens a new concern.

'It costs X, but you should check availability first'

No closure

Inspirational divergence

AI gives an answer that triggers the user to pivot to a larger, unrelated goal.

User asks about hiring cleaners → AI gives cost → user pivots to 'how do I start a cleaning business'

Real conversation examples

Three documented conversations from our dataset showing the price-to-cooling pattern in action.

Silent regression

User researching used cars asks about costs. AI gives vague estimates. User abandons the topic entirely.

TOFUTOFUTOFUTOFUBOFUBOFUBOFUBOFUBOFUBOFUBOFUTOFUTOFU
Price signal Cooling turn Normal turn

Uncertainty cascade

User with a booked Genius Bar appointment asks about diagnostics. AI hedges. User loses focus and drifts to unrelated topics.

TOFUTOFUMOFUTOFUTOFUTOFUBOFUTOFUBOFUTOFUTOFUTOFUTOFUTOFUTOFUTOFU
Price signal Cooling turn Normal turn

Scope expansion

User asks about day rates for window cleaners. AI gives a price range. User pivots from hiring cleaners to starting a cleaning company.

TOFUTOFUBOFUTOFUTOFUTOFUTOFUTOFU
Price signal Cooling turn Normal turn

What you can do about it

The root cause of low decision closure is website content that forces AI to hedge. Four content patterns on source websites reliably cause poor AI answers.

Vague price ranges

Replace 'prices from X' with scenario-based pricing. E.g. 'For a standard 150m2 house: X. For 200m2: Y.'

Conditional language in main text

Move all disclaimers to a separate FAQ section. Lead with the concrete answer.

Missing structured data

Implement JSON-LD schema markup (Product, Offer, PriceSpecification) for all transactional pages.

No 'next step' content

Every service page needs a clear, machine-readable conversion path. Tell the AI what happens after purchase.

Frequently asked questions

Only 6.8% of the 150,000 AI conversations analysed had commercial intent. The vast majority cover coding, writing, and general information retrieval.

In this analysis, 66.8% of conversations fell into an 'other' category, while 12.6% were writing-related, 12.3% were coding-related, and 3.4% were information-retrieval. Transactional conversations represented only 1.8% of all conversations.

When a user asks about price, a cooling event follows 31% of the time — an average intensity drop of -69 points. Over 90% of these cooling events result in the user regressing back to Problem Focus (TOFU), meaning the AI's answer fails to close the decision gap.

No. Only 4.4% of transactional conversations follow the classic TOFU-to-BOFU progression. In fact, 41% of users open their conversation already at BOFU — they know what they want before the first message.

Transactional conversations average just 2.92 turns. Nearly half (48%) consist of a single turn. Users in commercial mode want direct answers, not lengthy back-and-forth exchanges.

Measure your own decision closure

This analysis is based on aggregated data. With CitationLab, you get brand-specific conversation analysis across ChatGPT, Gemini and Google AI Overview.

Read the full article on LinkedIn