Does AI make brands sound better than cited sources?

A brand-citation tone pair links the tone used for one named brand in an AI answer to the tone used for that same brand on a cited page attached to the answer. Both sides use the same positive, neutral, or negative scale.

The observed cut contains 207,858 more-positive shifts and 41,615 more-negative shifts. A positive AI answer can therefore be more favorable than the pages it cites.

Seer observed that AI narratives amplify sentiment found in Trustpilot reviews. LREC-COLING research found that opinion summarizers tend to suppress negative opinion when positive text dominates the input data.

Those studies establish the concern. This observed cut supplies the direction of each pair, the total it is measured against, the engine split, and named source and brand leaderboards across cited pages.

AI tone became more positive in 40.55% of pairs, stayed the same in 51.34%, and became more negative in 8.12%.

The mismatch is directional. A single mismatch rate would hide the strong tilt toward more-positive wording.

The source page used a neutral tone in 237,096 pairs. AI used a positive tone in 80.60% of them, kept the tone neutral in 18.44%, and moved negative in 0.96%.

Neutral evidence does not usually produce neutral brand language in the observed answers.

The source page used a negative tone in 17,295 pairs. AI used a positive tone in 13,631, a neutral tone in 3,136, and a negative tone in 528.

Only 3.05% of negative-source pairs stayed negative. This is a tone comparison, not a finding that the criticism was false.

Google AI Mode produced 8.34 more-positive shifts per more-negative shift. Google AI Overviews was at 4.57, ChatGPT at 2.80, and ChatGPT Search at 2.27.

This split is descriptive. Each engine cited a different mix of pages and brands.

We restricted the current engines to the same 5,043 organic prompts. Google AI Mode produced a 9.99-to-1 ratio across 134,484 tone pairs. ChatGPT Search produced a 2.28-to-1 ratio across 44,682 tone pairs.

These matched prompts, meaning the same prompts asked on both engines, remove prompt selection as the full explanation. The engines still cite different numbers and mixes of pages, so the comparison remains descriptive.

Source type is the editorial category assigned to a cited page. Among the eight largest source types, social sources had the widest ratio at 6.55 to 1. Finance followed at 6.26.

Editorial sources had the narrowest ratio in this group at 3.93 to 1. The direction is broad rather than driven by one source type.

Among the 20 domains with the most tone pairs, newrelic.com produced 19.48 more-positive shifts per more-negative shift. Dash0 ranked second at 16.95, followed by YouTube at 11.58.

Soar was lowest in this high-volume group at 1.71 to 1. The table describes observed brand-citation pairs, not domain quality.

LinkedIn had the widest ratio among the 15 brands with the most tone pairs. New Relic followed at 10.17. Datadog, ClickUp, and Dynatrace were all above 9 to 1.

Every brand in the high-volume table had more positive than negative shifts. The result describes the answer-and-cited-page pairs attached to each brand, not brand quality.

The headline counts one tone pair per answer, so the same page and brand can come up in many answers. As a sensitivity check we kept only the most recent tone pair for each page-and-brand combination.

Across 178,007 such pairs, 42.94% became more positive and 7.82% became more negative. Repeat appearances of the same page and brand do not explain the headline direction.

We excluded 14,086,667 brand-citation pairs without both tone labels, 16,890 pairs with a mixed or hedged label, and 22,047 fully labeled pairs from prompts that were not organic. The headline uses only positive, neutral, and negative labels on both sides.

Tone labels are assigned automatically, not reviewed by a person. On 492,889 pairs where both labels had confidence of at least 0.8, the ratio remained 4.81 to 1. Source-type and domain cuts cover 511,085 of the 512,650 headline pairs.

AI turned 13,631 negative-source pairs positive. A positive AI answer can therefore cite negative source material.

Track the answer sentence and cited page together. Find criticism that repeats across pages, correct factual errors, and address the underlying product issue when the criticism is accurate. Then rerun the same prompts and measure the same page-brand pairs.