Do ChatGPT Queries Average 23 Words Compared With 4 for Google?

Where do the 23 and 4.2 figures come from?

Semrush analysed more than 80 million lines of clickstream data from July to December 2024. That period includes the launch of ChatGPT search on 31 October 2024. According to its summary and coverage of the study:

• Prompts without search averaged 23 words, with the longest at 2,712 words.
• Prompts with search averaged 4.2 words, with the longest at 301 words.
• About 54% of queries were answered without search, and about 46% used it.
• About 70% of prompts fell into an "unknown" intent category. They did not fit the navigational, informational, commercial and transactional labels used for search keywords, reflecting creative, problem-solving and brainstorming use.
• Unique domains receiving ChatGPT referrals rose about 300% from July to December 2024.

Keep the limits in mind. Clickstream data comes from a panel of users, not from OpenAI. The period was early in ChatGPT search's life, and the 2,712-word maximum shows that averages include pasted documents as well as questions.

Semrush has since updated the study with data to February 2026, and the gap has narrowed. Comparing early 2025 with early 2026, prompts without search fell from 24.9 to 13.5 words on average, and prompts with search rose from 4.7 to 8.7 words. Search ran on 34.5% of queries in February 2026, down from 46% in late 2024.

How did it become "ChatGPT versus Google"?

The study said the prompts with search "resembled traditional search engine behaviour" at 4.2 words. Summaries shortened this to a comparison between ChatGPT and Google, which was easier to repeat. The 4.2 figure is close to the three or four words commonly assumed for Google searches, which made the shorthand feel right. But Google has not published an average query length, and the study did not measure one. Quoted as "ChatGPT prompts average 23 words compared with 4 for Google", the statistic is wrong about its source. The direction of the difference is broadly right.

What does other data say about prompt length?

• Google's AI Mode. A year after launch, in May 2026, Google said the average AI Mode search in the US was triple the length of a traditional search query. Short and long queries were both growing. It also said searches starting "where to", "where should I" and "ideas for" were rising.
• What people use ChatGPT for. An NBER working paper on ChatGPT usage classified messages mainly as practical guidance, seeking information and writing. Each category tends to carry context: a situation, a draft, a constraint.
• Follow-ups. Conversations add context over several turns. A three-word question followed by "we're a 12-person team in Leeds on a tight budget" adds up to a long prompt, even if no single message is long.

Why are prompts longer than searches?

People write to an assistant the way they would brief a person. A search for "CRM small business" becomes a prompt such as: "What's a good CRM for a 12-person recruitment agency that already uses Outlook, needs GDPR compliance and can't spend more than £30 per user?" The extra words carry:

• The situation: industry, size, location, stage.
• Constraints: budget, tools already used, rules to follow.
• Preferences: what matters most, what to avoid.
• The task: compare, recommend, explain, draft.

Each of these changes the answer. Two people asking about "the best CRM" in these different ways can get entirely different recommendations.

Does the long prompt reach your page?

Not directly. When ChatGPT searches, it rewrites the prompt into shorter search queries, sometimes several of them. Browser developer tools can show the queries it runs, as covered in how to use browser tools to reveal ChatGPT's search queries and fan-outs. Those shorter queries decide which pages are retrieved. The full prompt then decides which of those pages are used and how the answer is framed.

So your pages have two jobs. They need to be findable for the short queries the engine runs. They also need to contain the specifics that match the long prompt: the constraint, the size band, the use case. A page that ranks for "CRM for recruitment agencies" but never mentions GDPR or Outlook may be retrieved and then passed over. How far ChatGPT's citations reach beyond the top search results is covered in whether 90% of ChatGPT citations come from pages ranked 21 or lower on Google.

What should content teams do with this?

• Write for situations, not only keywords. Add sections that answer for specific cases: "for teams under 20", "if you already use Outlook", "for regulated industries".
• State constraints and fit plainly: prices, limits, integrations, locations, compliance, who the product is not for.
• Keep keyword research, then extend it into full questions with the context buyers add. Sales calls and support tickets are the best source.
• Do not create a page for every variant. Google's guidance treats pages made at scale for every search variation, primarily to manipulate rankings or AI responses, as scaled content abuse. One strong page with clear sections covers many long prompts.

How to write in a style that suits conversational questions is covered in how to write conversational content for ChatGPT and AI search.

How should this change the way you test visibility?

Test with realistic prompts, not keywords. A three-word prompt such as "best CRM software" gives a generic answer dominated by the biggest brands. A 25-word prompt with a real buyer's situation often names different, more specialised products. If your visibility tracking uses only short phrases, it measures a kind of question few buyers ask. Build your question set from real buyer language, include the constraints, and keep a few short versions for comparison.

What might an illustrative case look like?

Consider an illustrative scheduling software company that tracked its AI visibility with 20 short prompts such as "best staff scheduling app". It appeared in almost none of the answers and concluded it was invisible. It then rewrote its test set from sales call notes, using prompts such as "What scheduling software works for a 40-person care home that needs to track staff training expiry dates?". It appeared in about a third of those answers, since its pages covered care-sector compliance in detail. The company kept both sets. It reported the long prompts as its main measure and the short ones as a view of general category visibility.

What do people ask about ChatGPT prompt length?

Is it fine to quote the 23-word figure?

Only as a late-2024 figure: ChatGPT prompts without search averaged 23 words then, against 4.2 words with search, according to Semrush. Its newer data puts prompts without search at about 13.5 words and with search at about 8.7 words in early 2026.

Are Google searches getting longer too?

Google says AI Mode searches average triple the length of traditional queries, and that long and short queries are both growing.

Should pages target long-tail phrases word for word?

No. Cover the situations and constraints clearly. Assistants match meaning, not exact long phrases.

Does prompt length affect which brands are recommended?

Often. Added constraints narrow the field, which can favour specialised products that state their fit clearly.