AI Mode queries in Search Console: what 155 sites show
Google doesn’t label AI Mode queries in Search Console, but they leave a trace: people reply to an AI, and the reply gets logged as a search. I classified every query of 155 sites over 16 months to see how often that happens, and how fast it’s growing.
Published · Jan-Willem Bobbink · Data: June 2025 to September 2026
When someone asks a follow-up question in AI Mode, Search Console counts it as a new query. Google says so in its help pages, and it doesn’t mark which queries came from AI Mode. So if you want to know how much of your search traffic is a conversation with an AI, the only evidence is in the queries themselves.
Most of those queries look like any other search. Some don’t: a query that just says “yes please”, or “what about salary”, only makes sense as the next turn of a chat. I counted those across 155 sites, week by week, from June 2025 to September 2026.
Key findings
- Replies typed at an AI are now normal. Bare replies such as “yes”, “ja” or “yes please” showed up as searches on 94 of the 155 sites. Per million impressions they grew from 5.6 to 116: 21 times as many in 16 months.
- German speakers say please to an AI, Indonesian speakers don’t. Three in four German yes-replies come with “bitte” or “gerne” (“ja bitte” on 52 sites). Dutch speakers add “graag” two times in three, English speakers say “yes please” 40% of the time, and Indonesian speakers never do.
- Follow-up questions are rising from a small base. Queries like “what about salary” are still rare, but they grew 39x over the window, most of it after May 2026.
- The typical site sees seven times more AI-shaped queries. The median site’s AI-shaped share of impressions went from 0.032% to 0.22%. Across all impressions pooled the share fell, because a few very large sites carry most of the volume.
- AI-shaped queries rarely get clicked. Their click-through rate was 2.9%, against 35.6% for all queries pooled.
- It’s still small. AI-shaped queries were 0.13% of all impressions. The growth is the story, not the size.
Show the data
| Figure | Value | Detail |
|---|---|---|
| Reply artefacts | 21x | 5.57 → 116 per million impressions |
| Pivot follow-ups | 39x | 0.12 → 4.56 per million impressions |
| Median site’s AI-shaped share | 7.0x | 0.032% → 0.22% of impressions |
| Sites with AI-shaped queries | 88 of 118 | in the week of 21 Sep 2026 |
What counts as an AI-shaped query
Google does not label AI Mode traffic in Search Console. It does count it: a follow-up question someone asks inside AI Mode is reported as a new query (Google's Search Console help). So the only way to find those queries is by their shape. A search engine query reads like a search engine query. A message typed at an assistant reads like a message.
GSC Wizard's Query Analysis report sorts every query into one of these shapes. Three of them are what this study calls AI-shaped:
- Reply artefacts: bare replies in the middle of a conversation, such as “yes”, “sure”, “show me” or “try again”. Someone answered an AI and the reply was processed as a search.
- Pivot follow-ups: comparisons asked mid-conversation, such as “what about gemini” or “how about in chinese?”. They only make sense as the next turn of a chat.
- Conversational questions: questions phrased at a listener, such as “is it free” or “can you…”.
The other shapes are reported but not counted as AI-shaped. LLM tracker probes (synthetic prompts from AI visibility tools) and agent harnesses (a whole machine instruction logged as one query) are machine traffic. Search operators, pasted strings (error messages, copied text) and long uncategorised queries (10 words or more with no clearer marker) are unusual, but not evidence of a conversation.
These are the most common examples in each AI-shaped group, counted by how many of the 155 sites had them. I only quote queries that appear on at least three sites.
| Shape | Query | Sites |
|---|---|---|
| Reply artefact | yes | 77 |
| Reply artefact | ja | 73 |
| Reply artefact | yes please | 62 |
| Pivot follow-up | what about this | 20 |
| Pivot follow-up | what about salary | 13 |
| Pivot follow-up | how about this | 12 |
| Conversational question | in english please | 35 |
| Conversational question | give me more | 20 |
| Conversational question | is it good | 19 |
A bare word isn’t proof on its own. “Ja” or “si” can be part of a brand or an abbreviation, and some people do type short searches. That’s why the trend over time matters more than any single query.
Replies typed at an AI are showing up as searches
In June 2025, reply artefacts were 5.6 per million impressions. By September 2026 they were 116, and they appeared on 94 of the 155 sites. The rise started in late 2025 and sped up from January 2026. It peaked in April and early May 2026 at 130 to 138 per million, dipped to 68 in the week of 22 June, and climbed back to about 120 by September.
The most common reply is the plainest one: “yes”, on 77 sites, followed by the same word in other languages (“ja”, “si”, “oui”, “да”). That fits an assistant that ends its answer with an offer (“Do you want me to find jobs near you?”) and a user who accepts. These replies are clicked at 4.7%, more than other AI-shaped queries: when a reply asks for more results, a list of links is what it gets.
Show the data
| Week starting | Reply artefacts per million impressions |
|---|---|
| 2025-06-02 | 4.79 |
| 2025-06-09 | 5.74 |
| 2025-06-16 | 5.03 |
| 2025-06-23 | 5.25 |
| 2025-06-30 | 6.07 |
| 2025-07-07 | 6.17 |
| 2025-07-14 | 6.34 |
| 2025-07-21 | 5.15 |
| 2025-07-28 | 4.29 |
| 2025-08-04 | 3.67 |
| 2025-08-11 | 3.56 |
| 2025-08-18 | 5.24 |
| 2025-08-25 | 7.02 |
| 2025-09-01 | 6.02 |
| 2025-09-08 | 6.8 |
| 2025-09-15 | 6.05 |
| 2025-09-22 | 7.41 |
| 2025-09-29 | 5.22 |
| 2025-10-06 | 7.13 |
| 2025-10-13 | 10.4 |
| 2025-10-20 | 9.39 |
| 2025-10-27 | 7.46 |
| 2025-11-03 | 9.31 |
| 2025-11-10 | 15.4 |
| 2025-11-17 | 9.62 |
| 2025-11-24 | 13.3 |
| 2025-12-01 | 9.46 |
| 2025-12-08 | 13.8 |
| 2025-12-15 | 13.3 |
| 2025-12-22 | 14.4 |
| 2025-12-29 | 20.9 |
| 2026-01-05 | 25.3 |
| 2026-01-12 | 29.3 |
| 2026-01-19 | 36 |
| 2026-01-26 | 51.9 |
| 2026-02-02 | 61.8 |
| 2026-02-09 | 78.1 |
| 2026-02-16 | 57.1 |
| 2026-02-23 | 63.2 |
| 2026-03-02 | 83.6 |
| 2026-03-09 | 84.3 |
| 2026-03-16 | 106 |
| 2026-03-23 | 112 |
| 2026-03-30 | 95.9 |
| 2026-04-06 | 110 |
| 2026-04-13 | 138 |
| 2026-04-20 | 131 |
| 2026-04-27 | 132 |
| 2026-05-04 | 134 |
| 2026-05-11 | 96.1 |
| 2026-05-18 | 86.1 |
| 2026-05-25 | 99.9 |
| 2026-06-01 | 89.1 |
| 2026-06-08 | 78.3 |
| 2026-06-15 | 74 |
| 2026-06-22 | 68.5 |
| 2026-06-29 | 90.3 |
| 2026-07-06 | 82.5 |
| 2026-07-13 | 95.8 |
| 2026-07-20 | 91.6 |
| 2026-07-27 | 127 |
| 2026-08-03 | 127 |
| 2026-08-10 | 108 |
| 2026-08-17 | 107 |
| 2026-08-24 | 104 |
| 2026-08-31 | 119 |
| 2026-09-07 | 122 |
| 2026-09-14 | 120 |
| 2026-09-21 | 119 |
Follow-up questions are rising from a small base
Pivot follow-ups are the clearest sign of a conversation: “what about accounting” means nothing unless something came before it. They are also the rarest shape, with 13,129 impressions over the whole window. Their growth is the steepest, though: from 0.12 to 4.6 per million impressions, 39 times as many, with most of the rise after May 2026.
Treat the multiple with care. A small base makes a big ratio easy, and one site with a few hundred extra impressions moves this line. The direction holds across sites; the exact multiple does not.
Show the data
| Week starting | Pivot follow-ups per million impressions |
|---|---|
| 2025-06-02 | 0.17 |
| 2025-06-09 | 0.08 |
| 2025-06-16 | 0.05 |
| 2025-06-23 | 0.25 |
| 2025-06-30 | 0.14 |
| 2025-07-07 | 0.1 |
| 2025-07-14 | 0.08 |
| 2025-07-21 | 0.06 |
| 2025-07-28 | 0.16 |
| 2025-08-04 | 0.07 |
| 2025-08-11 | 0.07 |
| 2025-08-18 | 0.03 |
| 2025-08-25 | 0.11 |
| 2025-09-01 | 0.08 |
| 2025-09-08 | 0.12 |
| 2025-09-15 | 0.06 |
| 2025-09-22 | 0.12 |
| 2025-09-29 | 0.08 |
| 2025-10-06 | 0.07 |
| 2025-10-13 | 0.16 |
| 2025-10-20 | 0.18 |
| 2025-10-27 | 0.91 |
| 2025-11-03 | 0.62 |
| 2025-11-10 | 0.18 |
| 2025-11-17 | 0.29 |
| 2025-11-24 | 0.19 |
| 2025-12-01 | 0.22 |
| 2025-12-08 | 0.3 |
| 2025-12-15 | 0.31 |
| 2025-12-22 | 0.25 |
| 2025-12-29 | 0.28 |
| 2026-01-05 | 0.45 |
| 2026-01-12 | 0.42 |
| 2026-01-19 | 0.44 |
| 2026-01-26 | 0.61 |
| 2026-02-02 | 0.69 |
| 2026-02-09 | 0.7 |
| 2026-02-16 | 0.74 |
| 2026-02-23 | 0.99 |
| 2026-03-02 | 0.89 |
| 2026-03-09 | 0.81 |
| 2026-03-16 | 0.75 |
| 2026-03-23 | 0.93 |
| 2026-03-30 | 0.84 |
| 2026-04-06 | 1.06 |
| 2026-04-13 | 1.19 |
| 2026-04-20 | 1.09 |
| 2026-04-27 | 1.24 |
| 2026-05-04 | 1.43 |
| 2026-05-11 | 1.46 |
| 2026-05-18 | 1.28 |
| 2026-05-25 | 1.5 |
| 2026-06-01 | 1.84 |
| 2026-06-08 | 2.36 |
| 2026-06-15 | 2.61 |
| 2026-06-22 | 3.04 |
| 2026-06-29 | 3.29 |
| 2026-07-06 | 3.27 |
| 2026-07-13 | 3.42 |
| 2026-07-20 | 4.37 |
| 2026-07-27 | 4.29 |
| 2026-08-03 | 4.51 |
| 2026-08-10 | 6.28 |
| 2026-08-17 | 4.64 |
| 2026-08-24 | 3.83 |
| 2026-08-31 | 4.01 |
| 2026-09-07 | 4.85 |
| 2026-09-14 | 4.01 |
| 2026-09-21 | 4.32 |
The typical site versus the total
Two ways of adding up the numbers give opposite answers. Take each site’s own AI-shaped share and look at the median site, and the share went from 0.032% to 0.22% of impressions: seven times higher. Pool every impression from every site, and the share fell from 0.31% to 0.10%.
The pooled figure is dominated by a handful of very large sites whose conversational-question traffic fell in the second half of 2025. The median site is the better answer to “what will I see in my own Search Console”. For the median I only used sites with data in at least 90% of the weeks, so sites connected partway through don’t distort it.
Show the data
| Week starting | Median property AI-shaped share | All impressions pooled |
|---|---|---|
| 2025-06-02 | 0.025% | 0.37% |
| 2025-06-09 | 0.023% | 0.33% |
| 2025-06-16 | 0.039% | 0.32% |
| 2025-06-23 | 0.024% | 0.35% |
| 2025-06-30 | 0.026% | 0.29% |
| 2025-07-07 | 0.039% | 0.29% |
| 2025-07-14 | 0.043% | 0.27% |
| 2025-07-21 | 0.039% | 0.28% |
| 2025-07-28 | 0.032% | 0.23% |
| 2025-08-04 | 0.040% | 0.18% |
| 2025-08-11 | 0.036% | 0.17% |
| 2025-08-18 | 0.038% | 0.19% |
| 2025-08-25 | 0.049% | 0.19% |
| 2025-09-01 | 0.047% | 0.19% |
| 2025-09-08 | 0.046% | 0.24% |
| 2025-09-15 | 0.031% | 0.26% |
| 2025-09-22 | 0.037% | 0.16% |
| 2025-09-29 | 0.037% | 0.15% |
| 2025-10-06 | 0.029% | 0.14% |
| 2025-10-13 | 0.031% | 0.13% |
| 2025-10-20 | 0.033% | 0.14% |
| 2025-10-27 | 0.040% | 0.14% |
| 2025-11-03 | 0.039% | 0.12% |
| 2025-11-10 | 0.044% | 0.098% |
| 2025-11-17 | 0.083% | 0.095% |
| 2025-11-24 | 0.044% | 0.067% |
| 2025-12-01 | 0.052% | 0.080% |
| 2025-12-08 | 0.043% | 0.073% |
| 2025-12-15 | 0.047% | 0.066% |
| 2025-12-22 | 0.048% | 0.044% |
| 2025-12-29 | 0.044% | 0.046% |
| 2026-01-05 | 0.056% | 0.062% |
| 2026-01-12 | 0.073% | 0.056% |
| 2026-01-19 | 0.089% | 0.055% |
| 2026-01-26 | 0.069% | 0.059% |
| 2026-02-02 | 0.074% | 0.059% |
| 2026-02-09 | 0.067% | 0.058% |
| 2026-02-16 | 0.076% | 0.069% |
| 2026-02-23 | 0.095% | 0.081% |
| 2026-03-02 | 0.12% | 0.095% |
| 2026-03-09 | 0.10% | 0.084% |
| 2026-03-16 | 0.094% | 0.081% |
| 2026-03-23 | 0.11% | 0.082% |
| 2026-03-30 | 0.11% | 0.085% |
| 2026-04-06 | 0.11% | 0.093% |
| 2026-04-13 | 0.12% | 0.095% |
| 2026-04-20 | 0.14% | 0.11% |
| 2026-04-27 | 0.16% | 0.11% |
| 2026-05-04 | 0.18% | 0.10% |
| 2026-05-11 | 0.12% | 0.092% |
| 2026-05-18 | 0.10% | 0.082% |
| 2026-05-25 | 0.11% | 0.083% |
| 2026-06-01 | 0.13% | 0.086% |
| 2026-06-08 | 0.16% | 0.084% |
| 2026-06-15 | 0.16% | 0.083% |
| 2026-06-22 | 0.16% | 0.077% |
| 2026-06-29 | 0.17% | 0.073% |
| 2026-07-06 | 0.16% | 0.13% |
| 2026-07-13 | 0.20% | 0.080% |
| 2026-07-20 | 0.24% | 0.087% |
| 2026-07-27 | 0.23% | 0.093% |
| 2026-08-03 | 0.24% | 0.083% |
| 2026-08-10 | 0.23% | 0.091% |
| 2026-08-17 | 0.23% | 0.11% |
| 2026-08-24 | 0.23% | 0.11% |
| 2026-08-31 | 0.19% | 0.10% |
| 2026-09-07 | 0.20% | 0.099% |
| 2026-09-14 | 0.21% | 0.11% |
| 2026-09-21 | 0.25% | 0.10% |
Splitting the pooled share by type shows where the fall comes from. Conversational questions went from 0.31% of all impressions to 0.089%, while reply artefacts rose from 0.00056% to 0.012% and pivot follow-ups from 0.000012% to 0.00046%. Even pooled, with the largest sites weighing in, the two clearest signs of a conversation grew 21 and 39 times.
Show the data
| Week starting | Conversational questions, share of impressions | Reply artefacts, share of impressions | Pivot follow-ups, share of impressions |
|---|---|---|---|
| 2025-06-02 | 0.37% | 0.00048% | 0.000017% |
| 2025-06-09 | 0.33% | 0.00057% | 0.0000077% |
| 2025-06-16 | 0.32% | 0.0005% | 0.0000053% |
| 2025-06-23 | 0.35% | 0.00053% | 0.000025% |
| 2025-06-30 | 0.29% | 0.00061% | 0.000014% |
| 2025-07-07 | 0.29% | 0.00062% | 0.0000095% |
| 2025-07-14 | 0.27% | 0.00063% | 0.0000076% |
| 2025-07-21 | 0.28% | 0.00052% | 0.0000057% |
| 2025-07-28 | 0.23% | 0.00043% | 0.000016% |
| 2025-08-04 | 0.18% | 0.00037% | 0.000007% |
| 2025-08-11 | 0.17% | 0.00036% | 0.000007% |
| 2025-08-18 | 0.19% | 0.00052% | 0.0000034% |
| 2025-08-25 | 0.19% | 0.0007% | 0.000011% |
| 2025-09-01 | 0.19% | 0.0006% | 0.0000079% |
| 2025-09-08 | 0.24% | 0.00068% | 0.000012% |
| 2025-09-15 | 0.26% | 0.00061% | 0.0000062% |
| 2025-09-22 | 0.16% | 0.00074% | 0.000012% |
| 2025-09-29 | 0.15% | 0.00052% | 0.000008% |
| 2025-10-06 | 0.14% | 0.00071% | 0.0000068% |
| 2025-10-13 | 0.13% | 0.001% | 0.000016% |
| 2025-10-20 | 0.14% | 0.00094% | 0.000018% |
| 2025-10-27 | 0.14% | 0.00075% | 0.000091% |
| 2025-11-03 | 0.12% | 0.00093% | 0.000062% |
| 2025-11-10 | 0.097% | 0.0015% | 0.000018% |
| 2025-11-17 | 0.094% | 0.00096% | 0.000029% |
| 2025-11-24 | 0.066% | 0.0013% | 0.000019% |
| 2025-12-01 | 0.079% | 0.00095% | 0.000022% |
| 2025-12-08 | 0.072% | 0.0014% | 0.00003% |
| 2025-12-15 | 0.064% | 0.0013% | 0.000031% |
| 2025-12-22 | 0.043% | 0.0014% | 0.000025% |
| 2025-12-29 | 0.044% | 0.0021% | 0.000028% |
| 2026-01-05 | 0.06% | 0.0025% | 0.000045% |
| 2026-01-12 | 0.053% | 0.0029% | 0.000042% |
| 2026-01-19 | 0.051% | 0.0036% | 0.000044% |
| 2026-01-26 | 0.054% | 0.0052% | 0.000061% |
| 2026-02-02 | 0.053% | 0.0062% | 0.000069% |
| 2026-02-09 | 0.051% | 0.0078% | 0.00007% |
| 2026-02-16 | 0.063% | 0.0057% | 0.000074% |
| 2026-02-23 | 0.074% | 0.0063% | 0.000099% |
| 2026-03-02 | 0.086% | 0.0084% | 0.000089% |
| 2026-03-09 | 0.076% | 0.0084% | 0.000081% |
| 2026-03-16 | 0.07% | 0.011% | 0.000075% |
| 2026-03-23 | 0.071% | 0.011% | 0.000093% |
| 2026-03-30 | 0.075% | 0.0096% | 0.000084% |
| 2026-04-06 | 0.082% | 0.011% | 0.00011% |
| 2026-04-13 | 0.082% | 0.014% | 0.00012% |
| 2026-04-20 | 0.092% | 0.013% | 0.00011% |
| 2026-04-27 | 0.092% | 0.013% | 0.00012% |
| 2026-05-04 | 0.086% | 0.013% | 0.00014% |
| 2026-05-11 | 0.083% | 0.0096% | 0.00015% |
| 2026-05-18 | 0.074% | 0.0086% | 0.00013% |
| 2026-05-25 | 0.073% | 0.01% | 0.00015% |
| 2026-06-01 | 0.077% | 0.0089% | 0.00018% |
| 2026-06-08 | 0.076% | 0.0078% | 0.00024% |
| 2026-06-15 | 0.076% | 0.0074% | 0.00026% |
| 2026-06-22 | 0.07% | 0.0068% | 0.0003% |
| 2026-06-29 | 0.063% | 0.009% | 0.00033% |
| 2026-07-06 | 0.12% | 0.0082% | 0.00033% |
| 2026-07-13 | 0.071% | 0.0096% | 0.00034% |
| 2026-07-20 | 0.077% | 0.0092% | 0.00044% |
| 2026-07-27 | 0.08% | 0.013% | 0.00043% |
| 2026-08-03 | 0.07% | 0.013% | 0.00045% |
| 2026-08-10 | 0.08% | 0.011% | 0.00063% |
| 2026-08-17 | 0.096% | 0.011% | 0.00046% |
| 2026-08-24 | 0.1% | 0.01% | 0.00038% |
| 2026-08-31 | 0.091% | 0.012% | 0.0004% |
| 2026-09-07 | 0.086% | 0.012% | 0.00049% |
| 2026-09-14 | 0.095% | 0.012% | 0.0004% |
| 2026-09-21 | 0.089% | 0.012% | 0.00043% |
AI-shaped queries rarely get clicked
Across the whole window, AI-shaped queries had a click-through rate of 2.9%. All queries pooled had 35.6%, a figure lifted by large brands that people search for by name, so the gap is smaller for a typical site. The order is still telling: follow-up questions (1.5%) and conversational questions (2.8%) are clicked least. Someone talking to an AI is mostly reading the answer, not choosing a link.
That matters when you read your own reports. An impression on “what about logistics” means your page was shown inside a conversation, not that someone was ready to visit.
Show the data
| Query shape | CTR | Impressions |
|---|---|---|
| Long uncategorised | 36.35% | 23,962,103 |
| All queries | 35.55% | 10,389,726,965 |
| Search operators | 8.23% | 396,675 |
| Pasted strings | 6.26% | 1,385,520 |
| Reply artefacts | 4.67% | 581,473 |
| Conversational questions | 2.84% | 12,952,687 |
| Pivot follow-ups | 1.45% | 13,129 |
| LLM tracker probes | 0.00% | 54,290 |
Which languages
Half of all AI-shaped impressions (49.8%) were in English, followed by Dutch (6.0%), Portuguese, Danish, Spanish, French and Swedish. That partly reflects which sites are in the study: many are Dutch or international. More than a third couldn’t be given a language at all, because a one-word reply like “ok” is too short to detect one. In total, AI-shaped queries were found in 27 languages.
Show the data
| Language | Share of AI-shaped impressions | Impressions |
|---|---|---|
| English | 49.83% | 6,751,272 |
| Not detected | 32.97% | 4,466,777 |
| Dutch | 5.98% | 809,512 |
| Portuguese | 2.45% | 331,304 |
| Danish | 1.75% | 236,976 |
| Spanish | 1.59% | 214,776 |
| French | 1.35% | 183,070 |
| Swedish | 1.15% | 155,629 |
| Italian | 0.92% | 124,463 |
| German | 0.76% | 102,474 |
| Polish | 0.50% | 67,091 |
| Indonesian | 0.29% | 38,991 |
| Arabic | 0.12% | 16,722 |
| Turkish | 0.12% | 16,468 |
| Japanese | 0.10% | 13,144 |
| Russian | 0.03% | 3,653 |
| Czech | 0.02% | 2,984 |
| Croatian | 0.02% | 2,617 |
| Persian | 0.02% | 2,600 |
| Bulgarian | 0.02% | 2,429 |
| Ukrainian | 0.01% | 1,981 |
| Hungarian | 0.01% | 836 |
| Romanian | 0.00% | 661 |
| Finnish | 0.00% | 315 |
| Chinese | 0.00% | 245 |
| Hebrew | 0.00% | 149 |
| Slovak | 0.00% | 101 |
| Norwegian | 0.00% | 49 |
Who says please to an AI
When an assistant offers to do something, people answer it, and Search Console logs the answer. That shows how politely people talk to an AI. Of all yes-replies, I counted how many carry a please or thanks word: “ja bitte” rather than “ja”, “yes please” rather than “yes”.
German speakers stand out. Three in four of their yes-replies include “bitte” or “gerne”, and “ja bitte” appeared on 52 sites. Swedish, Danish and Norwegian speakers score higher still (88%, with “ja tak”, “ja tack” and “ja takk”), though on only 22 sites. Dutch speakers follow at 67%: “graag” and “ja graag” each appeared on 46 sites, more than a bare “ja”. English sits in the middle at 40%, with “yes please” on 62 sites and a plain “yes” on 77.
At the other end, Hungarian, Hebrew, Persian, Finnish, Croatian, Serbian and Ukrainian speakers add a please or thanks one time in ten or less, and the 40 sites with Indonesian replies never saw one: only “ya” and “iya”. That is a habit, not rudeness. A bare “kyllä” is a normal answer in Finnish, and in Japanese はい is already polite on its own; お願いします (“please”) still appeared on 17 sites.
Show the data
| Language | Sites | Yes-replies | With please or thanks | Two or more words | Most common reply | Most common polite reply |
|---|---|---|---|---|---|---|
| Swedish, Danish, Norwegian | 22 | 51 | 88% | 69% | ja tak | ja tak |
| German | 55 | 209 | 75% | 48% | ja bitte | ja bitte |
| Bulgarian | 12 | 19 | 68% | 58% | да, моля | да, моля |
| Dutch | 50 | 210 | 67% | 37% | graag | graag |
| Romanian | 26 | 55 | 66% | 69% | da te rog | da te rog |
| Italian | 29 | 50 | 64% | 56% | si grazie | si grazie |
| Spanish | 43 | 129 | 54% | 49% | si por favor | si por favor |
| Japanese | 25 | 43 | 54% | n/a | はい | お願いします |
| French | 65 | 133 | 50% | 23% | oui | svp |
| Polish | 53 | 99 | 48% | 23% | tak | tak poproszę |
| English | 78 | 337 | 40% | 27% | yes | yes please |
| Czech, Slovak | 19 | 30 | 37% | 40% | áno | ano prosim |
| Portuguese | 59 | 102 | 32% | 34% | sim | sim por favor |
| Russian | 27 | 76 | 32% | 49% | да | да пожалуйста |
| Arabic | 62 | 107 | 22% | 22% | نعم | نعم من فضلك |
| Turkish | 46 | 88 | 18% | 10% | evet | evet lütfen |
| Hungarian | 45 | 50 | 10% | 8% | igen | igen kérem |
| Hebrew | 29 | 31 | 10% | 10% | כן | כן תודה |
| Persian | 29 | 46 | 9% | 7% | بله | بله لطفا |
| Finnish | 27 | 36 | 8% | 8% | kyllä | kyllä kiitos |
| Croatian, Serbian, Slovene | 34 | 69 | 7% | 16% | moze | da hvala |
| Ukrainian | 48 | 76 | 7% | 5% | так | так, будь ласка |
| Indonesian | 40 | 58 | 0% | 0% | ya |
How I counted: a yes-reply is a reply artefact made only of a language’s words for yes and for please or thanks. “Ok” is left out, because every language uses it. Each phrasing counts once per site, so one large site with thousands of bare “ja” impressions can’t decide the Dutch figure on its own. The language comes from the words, so these are languages, not countries: Spanish covers Spain and Latin America alike. A bare “ja”, “si” or “da” fits several languages; it goes to the language of that site’s other replies, and the 100 replies where that wasn’t possible are left out. The chart shows the languages seen on at least 20 sites; the table under it has all of them.
Run your own analysis for free
Every chart on this page comes from the Query Analysis report in GSC Wizard. Run the same analysis on your own sites and see which of your queries were typed at an AI. GSC Wizard is free for 7 days, and you don't need a credit card to start.
-
1
Log in
Sign in with Google or your email address. No credit card needed.
-
2
Connect Search Console
Pick the properties you want to analyse. Read-only access is enough.
-
3
Run Query Analysis
Open a property, go to Queries › Query Analysis and click Run Analysis.
How the study was done
The study covers 155 Search Console properties connected to one GSC Wizard account, a mix of client sites and my own. Only top-level properties are included, meaning domain properties and URL-prefix properties at the root of a host. Folder properties (such as example.com/blog/) are left out, and where a domain property and a URL-prefix property cover the same site, only one is counted, so no traffic is counted twice.
For every property, every query of the web search type was read week by week over 16 months and classified with the same classifier the Query Analysis report uses. Properties whose history is stored in GSC Wizard's warehouse were read from there; the others were read from the Search Console API.
- This is inference, not measurement. Google does not mark AI Mode queries. A reply artefact or a follow-up can come from another assistant whose messages end up in Google, or from a person who searches the way they chat.
- Trend figures use a balanced panel: the properties with data in at least 90% of weeks. Sites connected partway through would otherwise add impressions without history and read as growth.
- Search Console hides rare queries. Anonymised queries never appear with their text, so every figure here is a share of the queries Google does report.
- Recent weeks are slightly understated. Search Console keeps adding query rows for a few days after a date first appears. For most sites, the weeks after their history was first loaded were read early, and a fresh read of 36 sites showed about 16% more impressions for those weeks. AI-shaped queries were short by a similar amount or a little more, so recent per-million figures are if anything too low. The median share moved by about 0.01 percentage points.
- Partial weeks are left out, so the window runs from the first full week of June 2025 to the last full week of September 2026.
- Quoted examples only include queries seen on at least three sites, so none of them points to one client.
Reuse the charts and data
Use any chart on this page in your own article, deck or post, with a link back to this page. Each figure has buttons to copy the image, download the PNG or copy embed code that includes the link, and a table with the numbers behind it. The source line is part of every image.