What do people actually type into AI search?

People type something closer to a sentence than to a search box entry, and the shift that matters for a page is not length on its own but what the extra words are doing.

Three changes show up in Dombrowski's matched-panel study. Questions become far more common: about 27% of AI prompts against 8% of Google searches, roughly 3.3 times the rate. Keywords and bare entities fall from 85% of Google searches to 46% of AI prompts. And context appears, with 8.0% of AI prompts naming the asker's role, situation or purpose against 1.2% on Google.

That third one is the most useful and the least quoted. A prompt that says "we are a 40-person manufacturer replacing a legacy system" is self-qualifying as well as longer. The engine is being handed the segmentation your marketing team guesses at, and the page that matches it is the one that mentions the situation rather than only the product.

What has not happened is a wholesale replacement of keywords by conversation. Nearly half of AI prompts in that panel were still keywords or entities, and the same study found that a user's Google habits explained only about 4% of their AI behaviour, which means the two are closer to separate behaviours by the same person than to one behaviour migrating.

For a B2B company the practical read is that your pages need to carry the question, the role and the situation, and that the vocabulary for all three has to come from somewhere other than your own product naming.

How much longer is an AI prompt than a Google search?

Longer by somewhere between two and seven times, depending entirely on whose measurement you accept, which is itself the most useful thing to know about this statistic.

The matched-panel figure is 8.4 words for AI prompts against 4.2 for Google searches, with a median of 9 against 5 on 267 matched pairs covering the same need, where the AI version was longer 60% of the time. Semrush, in an article dated 11 December 2025, puts the average ChatGPT prompt at 23 words against 3.4 words for a Google search. Sundar Pichai told Alphabet's third-quarter earnings call on 28 October 2026 that queries in AI Mode are three times longer on average than traditional searches, and added that one in five AI Mode searches globally involves voice, images or documents.

Source	Figure	What it sampled
iPullRank, 29 Oct 2026	8.4 words vs 4.2 on Google	16,310 matched users, Perplexity, Copilot and you.com, November 2024 data, US desktop
Semrush, 11 Dec 2025	23 words vs 3.4 on Google	Its own ChatGPT search study, with the Google figure taken from a third party
Google, 28 Oct 2026	Three times longer	AI Mode, internal data, no baseline definition published

Nothing reconciles those three. They sample different assistants, in different years, counting words in different ways, and only one of them publishes a method you can check. A page that quotes any single one of them as the average prompt length is reporting a measurement artefact as a fact about buyers.

The ratio survives better than the level. Every source agrees prompts are longer than searches and that the gap is large, and that is the part worth building a page strategy on.

Why do the published prompt length numbers disagree so much?

Published prompt length numbers disagree because each one measures a different population with a different instrument, and the limitations are usually disclosed and rarely carried forward when the figure gets quoted.

Take the most carefully documented of them. The matched-panel study is built on Datos clickstream panels filtered to November 2024, which is before Google AI Mode existed and days after ChatGPT search launched, and the author's own list of what the study cannot prove includes how users behave today. The AI panel is 99.9% US desktop, skews 65% male and 65% aged 35 to 44, and ChatGPT makes up only 0.1% of it. About 70% of the Perplexity rows were truncated at 30 characters, which is why the length analysis runs on a subset rather than the full panel.

None of that makes the study weak. Publishing the constraints is what makes it the most trustworthy figure in the set. What it does mean is that the number describes Perplexity, Copilot and you.com users on US desktops two years ago, and that quoting it as "the average AI prompt is 8.4 words" drops every one of those qualifiers.

The other two are harder to interrogate. The vendor figure of 23 words comes with a linked study but no visible sample description for the comparison, and the Google figure is a sentence on an earnings call with no baseline definition or measurement window attached. A three-times multiple is unfalsifiable without knowing what it is three times of.

The honest position for anyone selling work on this is to stop quoting a single average. Our own research shows why that matters: in the deep version of our pharmaceutical process skid study we ran 27 buying queries across four AI engines, and in 20 of the 27 no fabricator won all four. How the question was phrased changed which firms appeared, which means the phrasing is a variable in your results rather than a statistic to cite.

What does a real AI prompt look like when it reaches a B2B site?

A real prompt reaching a B2B site looks nothing like a keyword, and our own logs are the clearest example we can show you because they are first-hand rather than averaged.

Bing Webmaster Tools is the useful view here, because Bing is the index behind ChatGPT's search. Our own Bing query log for this site is dominated by long conversational prompts rather than keywords. The single highest-position query on it, at position 1, is 38 words long and reads like somebody thinking out loud: a request to confirm whether a system is a pre-built retrieval setup, whether it resembles the one Perplexity or Google AI Mode uses, and then a request to name other systems that resemble those. Alongside it sit entries like "checklist for websites cited in ai answers" and "how does rag affect response differences between gemini chatgpt and claude".

The Google side of the same site shows the pattern in aggregate rather than in anecdote. Over the last 90 days, according to our own Google Search Console property, 132 distinct queries produced 1,148 impressions for this site, and 52% of those queries were six words or longer. For a site in a technical category, more than half of the demand arrives in a form no keyword tool would suggest writing for.

Two things are worth drawing from that. The first is that the specific long prompt almost never repeats, so there is no point targeting the string. The second is that the family it belongs to repeats constantly, and the family is identifiable by its vocabulary: the nouns, the engine names, the units. That is what a page can be built to match.

Any site with Search Console and Bing Webmaster access can produce the same view in an afternoon, which makes this the cheapest research in the discipline and the least often done.

How do you find the prompts that reach your own site?

Finding the prompts that reach your own site takes four sources, and none of them is a keyword tool. Run all four before you write anything, because they disagree with each other in informative ways.

Start with Bing Webmaster Tools, filtered to queries of six words or more. Bing sits behind ChatGPT's search, so its log is the closest thing to a direct read on conversational demand. Then Google Search Console, where long queries hide among short ones and where AI-feature impressions are counted inside the ordinary Search totals without being separated out, so you can see the phrasing but not which surface it came from.

Third, your own sales calls and support tickets, which is where the role and situation language lives. The study's finding that 8.0% of AI prompts name the asker's circumstances is a strong argument for mining the place your buyers already describe their circumstances at length.

Fourth, the sub-questions an engine generates on its own. Nick Haigler at Seer Interactive forced grounding on 501 prompts through the Gemini 3 API and recorded a mean of 10.7 fan-out queries per prompt, ranging from 3 to 28, with about 95% of the generated queries carrying zero global search volume (published 21 November 2025). Those sub-queries are what your page is actually being matched against, and no volume-based tool can show them to you because they have no volume.

Assemble the four into one list of questions in buyers' words, deduplicated by meaning rather than by string. That list is the input to everything else, and it is worth more than any published average.

What should you change on a page once you know the prompts?

Knowing the prompts should change three things on a page, and the first is worth more than the other two together.

Change the headings to the questions as they were asked. Not a compressed keyword version, the actual phrasing, including the words you would have edited out. A heading that reads like something a person would say aloud matches the way the query arrives, and an answer engine does not truncate headings the way a search results page truncates titles.

Change the first sentence under each heading so it answers the question outright. A reader arriving from a citation has not read the section above and will not read three sentences of preamble, and neither will the engine quoting you. Answer, then support.

Change what the page says about the reader. If buyers are naming their role and situation in the prompt, a page that names the same situation matches more closely than one that only describes the product. One sentence stating who the page is for does more than a paragraph of positioning.

What not to do is build a page per prompt. The long specific prompt does not repeat, the family does, and a thin page per phrasing produces a site of near-duplicates competing with each other. Our guide on two pages competing for the same AI answer covers that failure in detail.

If you want to see which phrasings the engines currently answer with somebody else's page, a free visibility assessment runs your buying questions through ChatGPT, Perplexity, Google AI Overviews and Claude and shows you the gap in their words rather than yours.

What else do people ask about AI prompts versus Google searches?

People ask these five questions about AI prompts versus Google searches while trying to work out whether their keyword research still describes anything real.