How often does AI recommend against a brand?

A reviewed statement is one brand described in one sentence of an AI answer, using cleaned brand names so that aliases count as the same brand. An explicit rejection is a statement that says the brand is not recommended for the stated need.

The observed cut contains 5,403 explicit rejections across 1,290,741 reviewed statements. Explicit rejection is too rare to use as a general measure of negative brand language.

A negative claim is a criticism or limitation recorded in the same reviewed statement. The observed cut contains 84,531 statements with a negative claim, or 6.55%. It contains 5,403 explicit rejections, or 0.42%.

BrightEdge reports negative sentiment by engine. This study measures the narrower decision that a brand is not recommended for a stated need. The two measures should not be combined.

The explicit rejections appear in 3,816 of 314,909 answers. They appear somewhere in 2,308 of 16,769 prompts, or 13.763%, because prompts were run more than once.

A prompt can produce an explicit rejection on one run and omit it on another. The answer rate measures what a user saw in one response.

ChatGPT Search has a 0.710% rejection rate. ChatGPT is at 0.429%, Google AI Mode at 0.264%, and Google AI Overviews at 0.122%.

These rates count individual statements, and each engine produced a different mix of statements about brands. One average across all engines would hide the engine difference.

We restricted ChatGPT Search and Google AI Mode to the same 15,996 organic prompts. ChatGPT Search has a 0.710% rejection rate across 495,970 reviewed statements. Google AI Mode has a 0.265% rate across 487,876 statements.

These matched prompts, meaning the same prompts asked on both engines, remove different questions as the full explanation. The engines can still name different brands and produce different numbers of statements.

No reason was recorded for 44.12% of explicit rejections. Risk warnings account for 17.90%, inferior comparisons 15.60%, conditional decisions 15.12%, fallbacks 6.83%, and budget-only decisions 0.31%.

A reason can tell a marketer whether the problem is safety, fit, comparison, or price. The group with no recorded reason needs a look at the actual sentences.

General needs have a 0.470% rejection rate. Pricing and contract needs are at 0.279%, feature requirements at 0.255%, and workflow needs at 0.089%.

A buyer-need type is the repeated purpose behind a prompt. The result shows where an explicit rejection appears, not which need type has more demand.

Blockchain and Cryptocurrency has a 1.050% rejection rate. Food and Beverage is at 1.007%, Privacy and Security at 0.921%, and Financial Services at 0.801%.

Sales and Marketing is at 0.186%. Compare industry groups only after checking the sample threshold and the brands included in each group.

Datadog has 74 explicit rejections across 4,272 reviewed statements. Jira Software has 69 across 2,800. ClickUp has 42. HelloFresh and PayPal each have 38.

This table ranks volume, not risk. Brands with more statements about them have more chances to appear in an explicit rejection.

Blur has an 8.82% rejection rate across 204 reviewed statements. OpenSea is at 7.23% across 249. Tether is at 6.65%, PayPal at 6.10%, and Healthy Paws Pet Insurance at 5.69%.

The rate table requires at least 200 reviewed statements and 10 explicit rejections. The threshold removes small samples from the leaderboard.

The source cut contained 1,864,552 candidate statements and five duplicate records. We counted each record once.

We excluded 162,188 statements from prompts outside the organic prompt set, 272,043 statements we could not match to a known brand, and 155,946 statements that had not passed source review. These groups can overlap. That leaves 1,290,741 reviewed statements as the base for every rate here.

The observed cut contains explicit rejections for 2,557 distinct brands across 2,308 prompts. A general sentiment score cannot show which buyer need produced the decision.

Audit the answer sentence, buyer need, and reason when a brand is explicitly rejected. Correct a factual error when the answer is wrong. Address the product limitation when the answer is accurate. Then rerun the same prompt across the same engines.