Do AI-cited sources agree about brands?

AI cited at least one positive and one negative page about the same ranked brand in 568 of 15,730 observed brand-answer evidence sets, or 3.6%. Each eligible set contained at least two distinct cited pages with validated, unambiguous source tone.

Most cited source sets did not contain this sharp form of disagreement. The 568 exceptions deserve separate review because a single answer can hide opposing evidence behind one recommendation.

Positive and neutral pages appeared together in 7,101 evidence sets, or 45.1%. Another 4,990 sets, or 31.7%, were positive only. The 568 positive-versus-negative conflicts consisted of 390 positive-plus-negative sets and 178 sets containing positive, neutral, and negative pages.

Not every difference in tone is a contradiction. Neutral context alongside praise is common, so the headline deliberately reserves conflict for evidence sets containing both positive and negative pages.

When the cited pages included both positive and negative source tone, the reviewed AI language was positive in 444 of 568 evidence sets, or 78.2%. It was neutral in 72, negative in five, mixed in one, and unavailable in 46.

Visible optimism does not mean the model resolved the disagreement correctly. It means teams should inspect whether the answer acknowledged the negative evidence instead of assuming the final tone summarizes every cited page.

Google AI Mode cited positive and negative pages in 496 of 12,391 eligible evidence sets, or 4.0%. ChatGPT Search did so in 72 of 3,339, or 2.2%, a difference of 1.8 percentage points.

The engines had different corpus sizes and source mixes, so this is an observed comparison rather than a causal ranking. Use an engine-specific baseline when reviewing a brand's citations.

The conflict rate was 3.8% for first-place brands, 3.2% for brands ranked second or third, and 3.8% for brands ranked fourth or lower. The widest gap among the three groups was 0.6 percentage points.

Source disagreement was not concentrated at the top or bottom of recommendation lists. Review source tone across ranks instead of limiting the audit to the winner.

Positive-versus-negative conflict appeared in 306 of 12,267 two-page sets, or 2.5%. The rate rose to 5.7% with three pages, 11.1% with four or five, and 27.2% with six or more pages.

This pattern is partly mechanical: more pages create more chances to observe both tones. Compare evidence sets with similar page counts before treating one brand or engine as unusually conflicted.

Among industries with at least 250 eligible evidence sets, Internet Services had 16 conflicts in 286 sets, or 5.6%, while Financial Services had nine in 573, or 1.6%. Software had the most observed conflicts: 49 in 958 sets, or 5.1%.

Industry differences identify where reviews may be more productive, but they do not establish why the sources differed. Prompt mix, brand mix, and cited-page count can all move the observed rate.

LinkedIn appeared in 50 conflicting evidence sets among 678 eligible sets. Atlassian had 23 among 197, Datadog had 20 among 185, Linear had 15 among 231, and Bright Data had 12 among 93.

This is a review-volume list, not a brand-quality ranking. Start with the named answers and pages, then compare prompts, engines, and page counts before drawing a reputation conclusion.

The main result was 568 of 15,730 evidence sets, or 3.6%. Starting June 1 returned 558 of 15,449, also 3.6%. Requiring every page in the set to carry sentiment confidence of at least 0.8 returned 502 of 14,519, or 3.5%.

The study measures opposing source tone, not factual contradiction, source quality, citation support, or what caused the answer. Prior citation research tests whether sources support claims; this cut asks whether the attached source set itself points in different brand-sentiment directions.

Positive-versus-negative source conflict appeared in 568 evidence sets, and 444 of those still carried positive AI language. The conflict rate increased with cited-page count and appeared at similar rates across recommendation positions.

Track source tone at the brand-answer level. Flag sets containing both positive and negative pages. Read the answer beside every cited page, verify which claims each page supports, and check whether important negative evidence was omitted or softened. Compare like-sized source sets, then repeat the fixed-window study next quarter.