AI Card Bao AICARDBAO.COM

Industry News

ChatGPT Search, Google AI Mode, and Perplexity: The difference isn’t those 20%; it’s that additional 1.3%.

“The statement ‘ChatGPT has taken away 20% of Google’s search traffic’ has been shared countless times in recent days. However, this is not the original meaning of those words; in fact, even if we interpret them literally, it doesn’t make any sense to claim that ChatGPT ‘has taken away’ a certain percentage of Google’s search volume.”

On November 2nd, Adam Fly of OpenAI stated that the number of queries on ChatGPT that trigger web searches accounted for approximately 20% of Google’s total query volume. This was merely a comparison between the two platforms; he did not provide the specific figures for Google. In February of this year, Ahrefs used publicly available data to estimate that ChatGPT generates 2.5 billion messages per day, while Google handles 13.7 billion searches daily. Based on these figures, the initial ratio between the two platforms was around 18%; however, when 65% of ChatGPT’s interactions were classified as search activities, this ratio dropped to 12%.

Therefore, this proportion roughly falls within the range of 12% to 20%. That’s certainly not a low figure; back in the day, it took Bing more than ten years to achieve a similar ratio.

But what exactly are the differences between these three options? The amount of searches conducted cannot provide an answer to this question. What I’m more concerned about are two other aspects: where the “answers” lead you, and whether the sources behind those answers are reliable or not.

For the same question, the number of people who clicked to view it differed by more than twenty times.

According to Ahrefs’ statistical sample, which included 76,000 websites, Google generated 39.98% of the total traffic, while ChatGPT accounted for only 0.21% – a difference of approximately 190 times. When converted based on the number of searches conducted, Google’s click-through rate is estimated to be around 29.2%, whereas ChatGPT’s rate is roughly 1.3%.

This click-through rate is an estimated value, and the statistical periods for the numerator and denominator also do not coincide exactly; however, judging by the order of magnitude is sufficient.

The scale of these numbers already speaks volumes. In the case of ChatGPT, the vast majority of people simply read the response and leave without even looking at the original text. This isn’t necessarily a bad thing – after all, many questions aren’t really worth spending extra time on. The problem is that if you don’t click to read the original text, you won’t notice that the references provided in the response are incorrect.

The accuracy rate has increased, but the source of this improvement has become even more uncertain.

Google has some rather awkward numbers to report. A company named Oumi was commissioned to conduct 4,326 searches using SimpleQA, and the accuracy rate of AI in providing relevant information increased from 85% during the Gemini 2 era to 91% in the Gemini 3 era. However, within the same set of tests, the proportion of cases where the answers were correct but the provided links did not fully support those answers rose from 37% to 56%. Oumi’s own summary is even more striking: only 39% of the responses were both correct and included complete sources.

A Google spokesperson responded that this study contained “serious flaws” and that SimpleQA did not represent the types of information people usually search for. This argument makes sense; indeed, those test questions covered rather obscure or less commonly known facts. What this test measured was merely an overview of AI capabilities, not the AI Mode feature within Google Search. The two cannot be equated directly.

The hard data available regarding ChatGPT and Perplexity are relatively outdated. In March 2025, the Tow Center at Columbia University tested eight different tools and asked them to identify the sources of news excerpts. The free version of Perplexity made 37% of mistakes, while the paid version achieved a accuracy rate of 45%. ChatGPT made 67% of errors in its searches, and Grok 3 scored 94% incorrectly. In the end, Perplexity came out on top, but it still made more than one-third of the mistakes – and its paid version performed even worse than the free one.

This set of data cannot be applied directly in today's context.

In the retest results recently released by Stanford this week, the median survival time for cases where the AI was identified as having deficiencies was only around 15 months. In the case of Tow, 19 months have passed since that retest, and several generations of models have been used since then. There are countless lists on the internet that claim things like “Perplexity: 92%, ChatGPT: 87%”, but when I tried to trace the sources behind these claims, I was unable to find any decent methodological explanations. So I’ll just skip mentioning them.

At present, I have not come across any serious test that compares these three companies and conducts it again this year. If someone claims that a particular company provides the most accurate data, ask them first what year those data were collected.

Regarding the issue of charging fees, these three companies took three different approaches.

The AI Mode is free for regular users and can be accessed through Google Search. The paid versions are designed specifically for the Gemini platform: AI Plus costs $7.99 per month, while AI Pro costs $19.99 per month. AI Ultra is available in two tiers, priced at $100 and $200 respectively. Launched in January this year, the Personal Intelligence feature can analyze your Gmail messages and photos to answer questions; it was initially made available to English-speaking Pro and Ultra users in the United States.

ChatGPT has started displaying ads in addition to its subscription-based services. Since last month, sponsored product images have begun appearing within conversations. If you ask “Which vacuum cleaner is good?” today, in the future, you’ll need to pay attention to see which brand’s products are displayed next to the answer.

The pricing structure of Perplexity is indeed the most complicated. The Pro version costs $20 per month or $17 when purchased annually, while the Max version costs $200. A third-party analysis at the end of September stated that the 4,000 Computer credits included with the Pro version are one-time benefits and are not renewed monthly; in contrast, the Max version comes with 35,000 one-time credits, plus an additional 10,000 credits awarded each month. Some users on Reddit have also complained that the most powerful OpenAI and Anthropic models are now available in the Max version. However, I have not found any official confirmation of this claim, so it remains merely speculation at this time.

Those working in development should pay close attention to the costs associated with APIs. Sonar charges $1 per million tokens for both input and output operations; separate fees apply for search functions. Sonar Pro offers pricing options of $3 or $15, in addition to a Pro Search fee of between $14 and $22 per thousand searches. If one attempts to estimate costs based solely on the price per token, the final figure will be significantly underestimated.

Every single aspect that makes one hesitate…

ChatGPT: It tends to be the least proactive when it comes to providing references. Its answers are often too comprehensive, which makes people too lazy to verify the information themselves. This is inherent in its design as a product; adding some annotations or labels won’t change this fact. Moreover, when questions related to buying products are asked, I certainly wouldn’t limit myself to just asking ChatGPT.

Google: It’s practically impossible not to use its AI. This week, its new policies for primary and secondary schools in the United States clearly reflect this stance. If schools want to disable the use of AI, they would have to turn off search functionality as well. The same approach is applied on the general search page – the AI overview appears at the very top, while 56% of those “answers that are correct but lack a reliable source” are difficult to spot.

Perplexity: Being caught in the middle is truly uncomfortable. The advantage lies in the fact that references are clearly available; however, since the model itself does not belong to one, it is necessary to compete with the free services offered by those two giants. What seems to be a viable strategy today may turn out to be ineffective next month.

For the same question, I will determine which category it belongs to.

For things that just happened, local businesses, or information that requires accessing the official website’s original text, simply use Google to search for the link and read it directly. Don’t limit yourself to just reading the first paragraph at the top of the page.

For situations that require multiple follow-up questions—such as comparing three different options or clarifying a complex concept—ChatGPT is the most convenient tool. However, when dealing with numerical information, it’s still advisable to verify the results again.

When writing a report, it is necessary to cross-check the sources or use search APIs in batches. In such cases, Perplexity remains a suitable option, provided that each citation is checked individually. In scenarios where there are many Chinese-language resources, it may also be worthwhile to compare the results obtained using Metta AI’s search function. None of these three overseas-based products excel particularly when it comes to providing reliable Chinese-language sources.

How much money is spent is actually secondary. Twenty dollars a month is not enough to buy something that “doesn’t require any verification”.

Relevant comparisons:For Claude Ultra and ChatGPT Pro, the 200-hour time frame is no longer available; only one provider provided information regarding how many times their weekly quota amount increased. As for Gemini 4 Carbon, Claude Opus 5.5, and GPT-6 Astra, it’s Carbon that leads by one point – this month, most people can only refer to these rankings.

Related