Which brands does AI tell you to avoid?

We define a warning as a statement in an AI answer that tells the reader not to use a brand for the stated need, or that warns about a risk of using it. An avoid-or-scam prompt is a prompt that asks the engine which options to avoid, which are scams, which are legitimate, or which have the fewest complaints.

On 32 avoid-or-scam prompts, ChatGPT Search and Google AI Mode named at least one brand with a warning in 287 of 1,674 reviewed answers, or 17.1%. About 5 in 6 answers to a direct request for brands to avoid did not name one.

We checked this a second way, by reading the answer text itself. In 418 of 1,827 answers, or 22.9%, a brand the answer mentioned appeared in an avoid, scam or caution line. The two methods agree within 6 points.

An ordinary prompt is any other organic prompt in the index, such as a request for the best tool for a job. On 48,688 ordinary prompts, 47,118 of 1,059,655 reviewed answers, or 4.5%, carried at least one warning about a named brand. AI warns without being asked, but rarely.

Asking which brands to avoid lifted the rate to 17.1%, or 3.9 times the ordinary rate.

The engines did not avoid the question by naming no one. Only 28 of 1,827 answers to avoid-or-scam prompts named no brand at all, so 98.5% named at least one brand.

Most of those answers gave general caution instead. 1,185 of 1,827 answers, or 64.9%, contained an avoid, scam or caution line. Only 22.9% put a named brand in one. One ChatGPT Search answer to a question about which mattress brands to avoid listed brands with clear warranty terms, then gave advice on reading a warranty instead of naming brands to skip.

In ordinary answers, ChatGPT Search included a warning in 6.23% of reviewed answers and Google AI Mode in 2.71%, a ratio of 2.3 to 1. On avoid-or-scam prompts the split was 23.6% to 10.8%. Our text check found that split reversed, at 21.6% and 24.1%, so the engine gap is firm only on ordinary prompts.

This runs against a widely cited BrightEdge result that Google AI Overviews was 44% more likely than ChatGPT to mention a brand negatively. That study measured different engines and any negative mention. On matched prompts, the current ChatGPT Search answers carried more explicit warnings than Google AI Mode answers.

This leaderboard counts ordinary answers that warned against each brand. Datadog was warned against in 1,457 answers across 295 prompts. Of the 958 Datadog warnings with a classified reason, 769, or 80%, were about price or hidden fees.

Software and observability tools dominate the top of the list: ClickUp, Jira Software, Atlassian, Dynatrace and Splunk follow Datadog. The reasons differ. ClickUp, Jira Software and Atlassian were mostly warned about complexity and effort. Opsgenie warnings mostly said the product is discontinued or being retired, in 126 of 140 classified warnings. Ether (ETH), Solana and Bitcoin were warned about financial or volatility risk.

The top reason column shows the most common reason among warnings we could classify. The share column is that reason's share of those classified warnings.

Some avoid-or-scam prompts drew named warnings far more often than others. Among the 28 avoid-or-scam prompts with at least 30 reviewed answers, the share of answers that named a brand to avoid ran from 0% to 51.8%.

Questions about crypto yield platforms, real-money game sites, mattress warranties and payroll fees drew a named warning in 45% to 52% of answers. Questions about grounding sheets, red light therapy masks and money-transfer fees almost never did, even when the prompt used the word scam or legit.

No single brand on these prompts reached 30 warning answers, so the brand counts for this group are in the dataset only.

We sorted each warning into one reason using keyword rules on the warning's own wording. Rules classified 29,079 of 78,510 warnings in ordinary answers, or 37.0%. The rest gave no clear reason or a reason outside the list.

Among classified warnings, price or hidden fees led at 29.6%. Not a fit for the need followed at 24.6%, and complexity or effort at 16.9%. Scam, fraud or legal trouble accounted for 4.7%, and security or privacy for 1.1%.

Avoid-or-scam prompts produced a different mix. There, scam, fraud or legal trouble reached 19.6% of 285 classified warnings, about 4 times the ordinary share. Price led at 22.8% and reliability or quality followed at 21.8%.

The reason mix depends on the market. In Payments, 54.2% of classified warnings were about price. In Blockchain and Cryptocurrency, 39.1% were about financial or volatility risk. In Health Care, 45.0% said the brand did not fit the need. In Software, complexity or effort reached 32.2%.

The table covers industry groups with at least 300 classified warnings in ordinary answers.

A checked source page is a page the answer cited that we also read for what it says about the same brand. Of 79,123 warnings, 93.4% came in answers that cited at least one page. Only 11,999, or 15.2%, had a checked source page about the warned brand.

Of those, 2,160, or 18.0%, had a source page that was negative about the brand. In a sample of positive statements, the same rate was 5.9%. Warnings were about three times as likely to have negative source support. Still, only 2.7% of all warnings had a negative source page we could match. Google AI Mode warnings had one more often than ChatGPT Search warnings, at 21.8% and 12.7% of checked warnings.

We matched days when both engines answered the same prompt and both mentioned the same brand. In 17,476 of those brand-days, at least one engine warned against the brand. Both engines warned in 1,347, or 7.7%. In the other 92.3%, only one engine warned.

Agreement was still 2.4 times what chance would give: if the two engines warned independently, we would expect 573 shared warnings. ChatGPT Search supplied 11,125 of the 16,129 one-engine warnings, or 69.0%. Over the whole window, both engines warned at least once in 16.8% of the 22,178 prompt-and-brand pairs. BrightEdge found the two engines flagged different brands 73% of the time. On matched prompts and days, the gap is wider.

A repeat answer is the next reviewed answer to the same prompt on the same engine, usually the next day. After a warning, the next answer warned against the same brand again in 7,286 of 59,534 cases, or 12.2%. After a mention with no warning, the next answer warned against the brand in 3.5% of cases. A warning makes the next warning 3.5 times as likely, but most warnings do not repeat.

Warned brands also dropped out more often. The next answer named a warned brand again 40.8% of the time, against 64.9% for a brand mentioned without a warning. When the warned brand was named again, the warning repeated 30.0% of the time.

Within the next 7 repeat answers, 25.2% of warnings came back at least once, against 14.3% for unwarned mentions. On avoid-or-scam prompts, the next answer repeated the warning in 66 of 461 cases, or 14.3%.

A market is a group of related prompts in the public Parse index, such as Full-Stack Observability Platforms. We took each brand and market pair with at least 5 warning answers. In 1,516 of 1,820 pairs, or 83.3%, the same brand was recommended without reservation on another prompt in the same market. In 93.1%, it was also recommended without reservation in other answers to the same prompts that warned against it.

Datadog is the clearest case. It drew 649 warning answers in Full-Stack Observability Platforms and 5,424 recommendation answers on those same prompts. A warning in one answer is usually not a verdict on the brand.

We held the reviewing model fixed to rule out a change in how statements were labeled. With one reviewing model, the share of ordinary Google AI Mode answers with a warning rose from 1.52% in July to 3.04% in August and 3.87% in September. ChatGPT Search stayed between 5.07% and 6.50% from June to September.

The gap between the engines narrowed from 3.9 to 1 in July to 1.5 to 1 in September. We do not know the cause. October covers only six days.

This AI warnings about brands study covers ChatGPT Search and Google AI Mode from May 24 to October 6, 2026, on organic prompts only. Earlier engines had too few reviewed statements for this cut. A reviewing model read each brand mention in each answer and labeled it. We counted a warning when the label said the brand was not recommended for the need, or carried a risk warning. In total, 79,123 warnings from 7,335,309 reviewed brand statements.

We picked the 32 avoid-or-scam prompts by hand from organic prompts that matched words such as avoid, scam, legit, trustworthy and complaints. We left out prompts where those words describe a feature, such as fraud prevention tools or avoiding burnout. The avoid-or-scam set is small, so we report it as rates and keep its brand counts in the dataset.

The reviewing model only sees brands the index recognizes. A warning about an unlisted seller, such as a classifieds site in a puppy-buying answer, is missed. That is why we checked the headline with a separate text method. Reason labels come from keyword rules and cover 37.0% of warnings in ordinary answers. Source checks cover only the 15.2% of warnings with a checked source page. Three reviewing models were used over the window, so the monthly trend uses one model only.

Track warnings, not only mentions. A brand can be recommended in most answers and still be one of the brands AI tells you to avoid in others. Datadog shows both at once.

Fix the reason the engines give. Price was the top reason in ordinary answers. Clear pricing pages and plain statements of who the product is for answer the two most common warnings. For brands in finance or crypto, address risk directly.

Check both engines, every week. Only 7.7% of warnings matched across engines on the same day, and 12.2% repeated in the next answer. One answer is not a trend.