Does AI change its tone when cited sources turn negative?

A re-ask is the next time the same engine answered the same prompt and described the same brand. We found 792 re-asks within seven days where every cited page about the brand read positive the first time and every cited page read negative the second time.

AI described the brand positively in 74.0% of the first answers and in 74.4% of the second answers. The cited pages switched tone. The AI tone did not.

AI tone changes between re-asks even when nothing else changes. When every cited page read positive both times, the AI tone changed in 25.0% of 78,095 re-asks. That is the normal rate of change.

When the cited pages switched from all positive to all negative, the AI tone changed in 28.4% of re-asks. That is 3.4 points above the normal rate, and the extra changes went in both directions.

In 781 re-asks, every cited page read negative the first time and positive the second time. AI described the brand positively in 74.1% of the first answers and 73.2% of the second.

Better sources did not produce better wording within the same week. Both switches point the same way: the tone of the cited pages did not steer the tone of the answer.

On Google engines, AI positive rate went from 79.8% to 80.4% across 613 positive-to-negative re-asks. On ChatGPT engines it went from 54.2% to 53.6% across 179.

With a 30-day limit between asks, the sample grows to 1,467 positive-to-negative re-asks. AI positive rate went from 74.1% to 72.1%, a 2.0-point drop.

A brand description is what one AI answer says about one brand. When every cited page about the brand read negative, AI described the brand positively in 65.0% of 19,273 brand descriptions and negatively in only 3.9%.

When every cited page read positive, the positive rate was 72.1%. All-neutral pages produced 65.3%, the same as all-negative pages. Seer and Trustpilot found that review content shapes how AI presents brands. This measurement shows how little of the page tone carries into the AI sentence.

The source gap is the AI positive rate when every cited page is positive minus the rate when every cited page is negative. It was 3.2 points in March, 5.9 in April, 7.1 in May, 5.0 in June, 6.0 in July, 8.5 in August, and 8.0 in September.

Over the same months, the overall AI positive rate moved between 61.6% and 76.5%. That 14.9-point swing from month to month was larger than the source gap in any month.

The prompts Parse tracks change over time, so we repeated the monthly cut on a fixed panel. The panel is the 7,006 prompts that Google engines answered in both March to April and July to August.

On that panel the source gap was 1.4 points in March and between 5.2 and 9.6 points in every later month. It did not grow steadily. The ChatGPT panel had too few all-negative brand descriptions per month to chart.

ChatGPT Search answers came from gpt-5-5 through August 7, 2026 and from gpt-5-6 from August 8. On the 2,087 prompts answered by both models, the AI positive rate rose from 52.8% to 69.1%.

The source gap also widened, from 6.9 points under gpt-5-5 to 16.0 points under gpt-5-6. Even so, gpt-5-6 still described 57.7% of brands positively when every cited page was negative.

ChatGPT Search described brands positively in 54.0% of brand descriptions, against 71.5% for Google AI Mode. Its source gap was 4.6 points. Google AI Mode had the largest gap at 9.0 points.

Each engine cited a different mix of pages and brands, so this split is descriptive. ChatGPT had only 103 all-negative brand descriptions, so its gap is the least certain.

Source type is the editorial category of a cited page. When a reference page was negative about a brand, AI still used positive wording in 51.2% of tone pairs. For e-commerce pages the figure was 54.5%.

Negative developer pages left AI positive in 70.8% of tone pairs, close to the 72.6% after positive developer pages. Social pages, mostly Reddit, carried the most negative tone pairs at 19,696 and left AI positive in 67.4%.

Among the 33 domains with at least 100 negative tone pairs, a negative Hacker News page led to a negative AI sentence 43.7% of the time. No other domain passed 20%. Parseable followed at 19.6% and ClickUp at 13.6%.

At the other end, AI used positive wording after more than 83% of negative pages from Proposify, ProofHub, FreeTaxUSA, and Orb. Reddit, with 19,405 negative tone pairs, led to a negative AI sentence 3.7% of the time.

Among 16 brands with at least 75 all-negative brand descriptions across at least 30 prompts, AI described Splunk negatively 42.9% of the time when every cited page was negative. Datadog followed at 21.6% and PagerDuty at 19.8%.

For monday.com, AI stayed positive in 83.0% of all-negative brand descriptions. LinkedIn was at 74.8% and New Relic at 74.4%. The table describes AI wording next to negative pages, not brand quality.

We kept 990,319 of 3,182,797 tone pairs with readings on both sides. We excluded 1,993,100 read by other tone readers, so each side uses one reader for the whole year. We also excluded 141,226 from prompts that were not organic, 42,974 with a mixed or hedged label, and 15,178 from October 2026, a partial month.

Tone labels are automated. Each cited page was read once, between May 25 and August 21, 2026, so for older answers the page may have changed since it was cited. Most ChatGPT tone pairs before May 2026 came from gpt-5-mini, and Google AI Mode cited fewer pages per answer in September than in May, so the September level is less comparable. Keeping only pages cited for that specific brand gave a 2.9-point source gap. Requiring confidence of at least 0.8 on both sides gave 6.6 points.