paper demo · ShareChat (arXiv:2512.17843)

Source links

ShareChat keeps each message's outbound links — the paper's SourceLink field — so citation behaviour is directly measurable. — 5 questions, shown as one conversation. · × exit demo

round 1 of 5

What share of conversations include links or citations, per platform?

% of conversations with links/citations by platform

grok 25.17

chatgpt 0

claude 0

gemini 0

perplexity 0

What this shows: Each bar is one platform; the value is % of conversations with links/citations. grok is highest at 25.17%; chatgpt is lowest at 0% across 5 groups. Search-centric platforms ground far more of their answers in sources than chat-first ones.

5 row(s) · 0.055s on the full corpus

compiled SQL (read-only, parameterized)

SELECT grp, AVG(t.has_links) * 100.0 AS value, COUNT(*) AS n
FROM (
    SELECT c.platform AS grp,
        c.id AS cid,
        c.n_messages AS n_messages, c.turns_count AS turns_count,
        cs.has_thinking AS has_thinking, cs.has_code AS has_code,
        cs.has_links AS has_links, cs.is_complete AS is_complete
    FROM conversations c
    LEFT JOIN conv_signals cs ON cs.conversation_id = c.id
    WHERE 1=1 AND c.platform IS NOT NULL
) t
GROUP BY grp
ORDER BY value DESC
LIMIT 20

-- params: []

open this round in the analyzer ⎘

round 2 of 5

Are conversations that cite sources longer than average?

avg messages per conversation

avg messages per conversation	n
9.69	3,309

What this shows: One overall figure: avg messages per conversation is 9.69, computed over 3,309 rows. Scope: has_links = true. Round 2 narrows to link-bearing conversations only — compare with the next round.

1 row(s) · 0.007s on the full corpus

compiled SQL (read-only, parameterized)

SELECT AVG(t.n_messages) AS value, COUNT(*) AS n
FROM (
    SELECT c.id AS cid,
        c.n_messages AS n_messages, c.turns_count AS turns_count,
        cs.has_thinking AS has_thinking, cs.has_code AS has_code,
        cs.has_links AS has_links, cs.is_complete AS is_complete
    FROM conversations c
    LEFT JOIN conv_signals cs ON cs.conversation_id = c.id
    WHERE cs.has_links = ?
) t

-- params: [1]

open this round in the analyzer ⎘

round 3 of 5

And how long are conversations without any links?

avg messages per conversation

avg messages per conversation	n
9.25	126,275

What this shows: One overall figure: avg messages per conversation is 9.25, computed over 126,275 rows. Scope: has_links = false. Same metric, inverted filter: a follow-up refines the previous plan instead of starting over.

1 row(s) · 0.02s on the full corpus

compiled SQL (read-only, parameterized)

SELECT AVG(t.n_messages) AS value, COUNT(*) AS n
FROM (
    SELECT c.id AS cid,
        c.n_messages AS n_messages, c.turns_count AS turns_count,
        cs.has_thinking AS has_thinking, cs.has_code AS has_code,
        cs.has_links AS has_links, cs.is_complete AS is_complete
    FROM conversations c
    LEFT JOIN conv_signals cs ON cs.conversation_id = c.id
    WHERE cs.has_links = ?
) t

-- params: [0]

open this round in the analyzer ⎘

round 4 of 5

How did citation usage change month by month?

% of conversations with links/citations by month

What this shows: % of conversations with links/citations over 20 months: it went from 0% in 2023-05 to 0% in 2024-12, peaking at 0% in 2023-05. SourceLink × Timestamp: the share of cited conversations becomes a trend line.

20 row(s) · 0.149s on the full corpus

compiled SQL (read-only, parameterized)

SELECT grp, AVG(t.has_links) * 100.0 AS value, COUNT(*) AS n
FROM (
    SELECT substr(cs.created_day, 1, 7) AS grp,
        c.id AS cid,
        c.n_messages AS n_messages, c.turns_count AS turns_count,
        cs.has_thinking AS has_thinking, cs.has_code AS has_code,
        cs.has_links AS has_links, cs.is_complete AS is_complete
    FROM conversations c
    LEFT JOIN conv_signals cs ON cs.conversation_id = c.id
    WHERE 1=1 AND substr(cs.created_day, 1, 7) IS NOT NULL
) t
GROUP BY grp
ORDER BY grp
LIMIT 20

-- params: []

open this round in the analyzer ⎘

round 5 of 5

Which topics cite sources most?

% of conversations with links/citations by topic

伏，李伟民的眼睛却是被那对胸脯吸引了，骆雪雁的脸色渐渐变 100

specific_info ソース: https://www.un.org/en/sc/itc/treaties/cybercrime/index.shtml 100

i'm ready to classify the conversation. however, i need the conversation transcript from the txt file to proceed. 100

html_editing 100

agriculture 100

# this is a code block with the conversation transcript and the classification of the user's capability ## conversation transcript 1. you are a conversation topic 100

sql 50

there is no previous conversation to classify. 33.33

explain_in_simple_terms 33.33

game_and_role_play 25

i'm ready to classify the conversation. 22.77

**data_analysis** 16.67

philosophical_discussion 13.33

**other** 9.52

purchasable_products 7.14

specific_info 3.34

creative_ideation 2.93

asking_about_the_model 2.90

health_fitness_beauty_or_self_care 2.85

data_analysis 2.68

What this shows: Each bar is one topic; the value is % of conversations with links/citations. 伏，李伟民的眼睛却是被那对胸脯吸引了，骆雪雁的脸色渐渐变 is highest at 100%; data_analysis is lowest at 2.68% (37.3× spread) across 20 groups. Final round: grounding behaviour differs by subject matter.

20 row(s) · 2.143s on the full corpus

compiled SQL (read-only, parameterized)

SELECT grp, AVG(t.has_links) * 100.0 AS value, COUNT(*) AS n
FROM (
    SELECT DISTINCT m.topic AS grp, c.id AS cid,
        c.n_messages AS n_messages, c.turns_count AS turns_count,
        cs.has_thinking AS has_thinking, cs.has_code AS has_code,
        cs.has_links AS has_links, cs.is_complete AS is_complete
    FROM messages m
    JOIN conversations c ON c.id = m.conversation_id
LEFT JOIN conv_signals cs ON cs.conversation_id = c.id
    WHERE m.topic IS NOT NULL
) t
GROUP BY grp
ORDER BY value DESC
LIMIT 20

-- params: []

open this round in the analyzer ⎘

end of the walkthrough That was “source links” in 5 rounds. try your own question → · timestamps demo · conversation completeness demo