Key Takeaways
- Across 92 days, 178 real queries reached my site. 44 of them ran ten words or longer, so roughly a quarter of the query list, and 21.7% of the impressions
- The two groups barely overlap in length. Short queries have a median of 4 words, long ones a median of 31, and only 16 queries sit in the 7 to 9 word gap between them
- Thirty of the 44 long queries are written in the first person. They name a job, a climate, a body part, a product already tried and rejected, sometimes all at once
- Those 44 queries produced zero clicks at an average position of 38.9. You get seen on page four, which is a ranking problem sitting on top of a research opportunity
- Google's own generative AI report, global since August 31, 2026, gives impressions only. No queries, no clicks, no position, so a ten-word regex on the normal report is the only proxy anyone has
Edu here. I opened Search Console last Monday to see whether the ChatGPT ads post had picked anything up, which it mostly hadn't. Then I scrolled past the summary into the actual query list and found this sitting there:
i work outside in texas heat and need deodorant for pits groin butt and feet which mando product covers all areas
Twenty-one words. A named competitor. A climate, a job, and four body parts. Nobody typed that into a search box the way we've spent twenty years thinking about search boxes.
My first thought was that it was a fluke, some scraper or a bored person. So I exported the whole thing and counted, and I'm now fairly sure it isn't a fluke, it's just the part of the list I'd never bothered to scroll to.
Across 92 days of Google Search Console data for a small site, 44 of 178 genuine queries ran ten words or longer, accounting for 21.7% of impressions. The median long query was 31 words against 4 words for the rest, and 30 of the 44 were written in the first person, naming the searcher's own situation.
What's actually in the export?
Ninety-two days, June 7 to September 6, 2026, web search only: 7,753 impressions and 25 clicks, a 0.32% clickthrough rate. Small site, and I want to be upfront that everything below sits on a small sample. The query-level export carries 199 rows, 4,419 impressions and 3 clicks, because Google anonymises most queries and hands you a filtered slice.
The first number I wrote down was wrong, and I'll leak that here since it's the most useful thing in this section. One query, youtube87177, accounts for 3,508 of those 4,419 impressions. I have no idea what it is (it ranks at position 7, on my homepage, for a string that means nothing). Nearly 80% of my export is one piece of nonsense, and if I'd trusted the headline figure I'd have written a very confident post about traffic I don't have.
Another 18 queries are search operator strings, the kind ending in a wall of -site:reddit.com -site:twitter.com. Those are monitoring tools rather than people. They average position 5.7, so my best rankings by a distance belong to robots. Strip both groups and 178 queries are left, carrying 844 impressions and exactly one click.
Then I sorted those 178 by word count, and the shape surprised me.
A hundred and eighteen queries of six words or fewer. Sixteen in the seven to nine band. Then forty-four at ten words or more, with a median of 31 and a longest of 57. The middle is where the queries aren't.
The average barely moved, so everyone missed it
Published averages say almost nothing is happening. Similarweb data has US Google query length flat between 3.33 and 3.36 words from June 2024 through May 2025, then rising to a little over 3.51 words by May 2026 (OfficeChai, Google Search Queries Have Become Longer Since The Introduction Of AI Mode, 2026). Fifteen hundredths of a word in a year. You would never build a strategy on that.
Semrush found something bigger by splitting the surface instead of averaging it. Analysing clickstream data from roughly 69 million US desktop Google sessions between May 1 and July 5, 2025, they put the average AI Mode query at 7.22 words against 4.0 for a traditional search (Semrush, Google AI Mode's Early Adoption and SEO Impact, 2025). Nearly double, on the same search engine, in the same week.
My short bucket has a median of exactly 4 words, which lines up with Semrush's traditional figure so neatly that I double-checked it. My long bucket sits at 31. There is no version of an average that describes both of those honestly, and an average is what most people look at.
Average query length is a poor instrument for this shift. Similarweb recorded US Google queries moving from about 3.35 words to 3.51 between mid-2024 and May 2026, while Semrush measured AI Mode queries at 7.22 words against 4.0 for traditional search across 69 million sessions. The change lives in the tail rather than the mean.
Some of these aren't searches. They're prompts with the packaging still on
One query in my export reads, in full: "context: location: united states (not for language). do not include location references in your response. question: do protein shake subscriptions usually offer a discount, and what's the catch with skipping or pausing deliveries?" Three impressions, average position 5.33.
That's not a person. "Do not include location references in your response" is instruction scaffolding, and it arrived in my Search Console with the scaffolding still attached. Somewhere between a user and Google, a machine wrote that string, and Google logged it as a search query like any other.
The method for finding these isn't mine. Chris Long published it on Search Engine Land on February 27, 2026: filter the Performance report by query using the regex ^(?:\S+\s+){9,}\S+$, which matches anything of ten words or more (Search Engine Land, How to see AI search prompts inside Google Search Console, February 2026). I used the same ten-word cut here, partly so the numbers are comparable to anyone else running it.
His caveat matters more than the trick, and I'd repeat it louder. In his words, "we don't have direct evidence that these types of queries are directly from ChatGPT, AI Mode or any other AI platform." It could be people using Google like a chatbot. Maybe it's assistants querying on someone's behalf. Probably it's both, in some mix nobody outside Google can see.
And Google isn't going to settle it. The Search generative AI performance report launched on June 3, 2026 and finished rolling out worldwide on August 31 (Search Engine Land, Google Search Console AI performance reports rolling out globally, August 2026). It covers AI Overviews and AI Mode, and it reports impressions only, broken out by page, country, device and date. No queries. No clicks, no CTR, no position (Google Search Console Help, Generative AI performance report, 2026). The one report built for this question is the one that won't tell you what was asked.
So the regex is a proxy for something Google measures and won't show. I'm fine with that, honestly. It was good enough to find this:
hvac worker with swamp butt during 12 hour summer shifts needs whole body deodorant for butt crack mando spray vs cream which applies better to hairy areas
What a 31-word query carries that a 4-word one doesn't
Thirty of my 44 long queries are written in the first person, and 31 open by stating a role or a situation before they get anywhere near a product. Honestly that's the whole finding. The length isn't decoration, it's the person explaining themselves before they ask.
Take the HVAC one apart and you get a research brief that no keyword tool on earth would have sold me.
| Fragment | What it tells you |
|---|---|
| "hvac worker" | Occupation, which is a real segment nobody targets |
| "12 hour summer shifts" | Duration of wear, the actual performance spec |
| "swamp butt", "butt crack" | The words they use, none of which appear on a product page |
| "mando" | The competitor already in consideration |
| "spray vs cream" | The unresolved decision, so the objection to answer |
| "hairy areas" | The constraint driving the whole choice |
A keyword tool gives you "whole body deodorant" with a volume number beside it. This gives you a person. And when I ran the numbers on the deodorant cluster specifically, 11 long queries carrying 77 impressions, every one of them named a use case rather than a product feature: outdoor work in heat, jiu-jitsu, hormonal sweat, scent that doesn't read as cologne.
None of this is new information to me, and that's the slightly deflating part. Same language the whole body deodorant study pulled out of Reddit months ago, with the same themes in roughly the same order. What's new is the delivery channel. That vocabulary is now arriving in an SEO report, filed under queries, next to "voc research" and "dtc pricing."
Which means the method I keep writing about, going and reading how people describe the problem to each other, now has a second entrance that costs nothing. The step-by-step version is in how to find the exact words your customers use, and your own console is a smaller, free, much less representative sample of the same thing.
i need whole body deodorant for my groin that doesn't smell like cologne down there truly unscented or very subtle
You show up, and then nothing happens
Here's the unglamorous half. Those 44 long queries generated 183 impressions and zero clicks, at an impression-weighted average position of 38.9. Page four. The short queries did better and still only managed one click across 601 impressions, so I'm not claiming the long ones are uniquely broken.
But the position gap inside the long bucket is strange, and I can't explain it properly. Long queries about pausing a subscription land at position 6.9. Long queries about deodorant land at 45.5, and the skincare ones at 51.0. Same site, same template, same person writing them.
My best guess is competition: pause-versus-cancel is a thin SERP where the subscription study is one of very few pages that answers the question directly, while beauty and personal care are trench warfare. Maybe it's also that the pause queries are shorter and more decision-shaped, closer to a normal question. I don't know, and with 26 impressions in that cluster I'd be inventing a rule from almost nothing.
What I'd stop short of concluding is that long queries are a traffic opportunity. On this evidence they're a visibility signal with no revenue attached yet. Read them for what they say rather than for what they send.
Your console shows you forty of these. Your category has thousands.
Insightios goes to the threads where your market explains the problem to itself, codes what repeats, and hands back the phrasings, use cases and objections ranked by how common they are, with the real quotes attached. Flat fee, fixed turnaround.
Frequently asked questions
How do you find AI-style prompts in Google Search Console?
Filter the Performance report by query with a custom regex matching ten words or more. Chris Long published the pattern on Search Engine Land in February 2026 as ^(?:\S+\s+){9,}\S+$. It finds unusually wordy queries, which correlate with prompt behaviour, though Google never labels a query as coming from an assistant.
Does Search Console show which queries came from AI Mode?
No. Google launched the Search generative AI performance report in June 2026 and finished rolling it out globally on August 31, 2026. It covers AI Overviews and AI Mode but reports impressions only, broken out by page, country, device and date. There are no queries, no clicks, no CTR and no position.
Are search queries actually getting longer?
The average moved very little. Similarweb data puts US Google query length flat between 3.33 and 3.36 words from June 2024 to May 2025, then climbing to just over 3.51 words by May 2026. The change is in the tail rather than the mean, since a small share of queries is now dramatically longer.
How long is a query in Google AI Mode?
About 7.22 words, against 4.0 words for a traditional Google search. Semrush reached those figures from clickstream data covering roughly 69 million US desktop Google sessions between May 1 and July 5, 2025. Close to double the length of a standard query.
Why do long search queries get so few clicks?
Mostly because they rank badly. Across 92 days, the 44 queries of ten words or more that reached my site produced 183 impressions, zero clicks and an impression-weighted average position of 38.9, so page four. Pages written for a category rarely answer a question phrased around one person's situation.
What can a DTC brand do with prompt-shaped queries?
Read them as customer language rather than as keywords. Thirty of my 44 long queries were first person and named a job, a climate, a body part or a product already tried and rejected. That detail is the raw material for ad copy, landing pages and the context hints used by ChatGPT ad targeting.
What to do next
Open your own Performance report, set the date range to three months, and filter queries by that regex. It takes about two minutes. Then read what comes back as language rather than as keywords, because as keywords most of it is worthless (single-digit impressions, page four, no clicks).
What you're looking for is the stuff a volume tool structurally cannot give you: the occupation, the season, the duration, the competitor they already tried. Write those down. They belong on your product page and in your ad copy, and if you end up buying inside an assistant they belong in the targeting field too.
Then be honest about the sample size. Forty-four queries over three months is a hint rather than a study. Mine skew to whatever I happen to have published, and a brand with real traffic would see a different and much bigger version of this. Call it a free look at one corner of the room.
All in all I think the interesting bit isn't that queries got longer. The extra words are almost entirely context about the person, arriving in a channel built to strip context out. Search Console was designed to tell you what people typed. For a growing slice of it, it's now telling you who they are, and I'm not sure the tooling has caught up with that at all.
Sources
- Insightios Google Search Console export, web search, June 7 to September 6, 2026. 199 query rows, 178 after removing junk and search-operator strings. Analysis by the author.
- Search Engine Land. (2026). How to see AI search prompts inside Google Search Console, Chris Long, February 27, 2026. Link Retrieved September 8, 2026.
- Google Search Console Help. (2026). Generative AI performance report (Search). Link Retrieved September 8, 2026.
- Search Engine Land. (2026). Google Search Console AI performance reports and Search generative AI control rolling out globally, Barry Schwartz, August 31, 2026. Link Retrieved September 8, 2026.
- Semrush. (2025). Google AI Mode's Early Adoption and SEO Impact (clickstream data, ~69M US desktop sessions, May 1 to July 5, 2025). Link Retrieved September 8, 2026.
- OfficeChai. (2026). Google Search Queries Have Become Longer Since The Introduction Of AI Mode, Data Shows (Similarweb data, June 2024 to May 2026). Link Retrieved September 8, 2026.