Query fan out is Google's technique for splitting one question into many related searches, running them at once and building an AI answer from the results. Impressions from AI Overviews and AI Mode are counted in your Search Console totals (Google, 03/06/2026). My unconfirmed hypothesis is that fan-out adds low-click queries, so CTR falls while position holds.
What is query fan out and how does it work?
Query fan out is the technique Google's AI features use to answer one question with many searches. The model breaks the question into subtopics, writes related sub-queries, runs them at the same time against the Search index, retrieves pages and writes one answer. Each sub-query is a fresh chance for your page to be retrieved.
Google names the technique in its own documentation, which defines query fan-out as "A set of concurrent, related queries generated by the model to request more information and fetch additional relevant search results to address the user's query."
The Google blog post announcing AI Mode updates at I/O on 20/05/2025 says: "Under the hood, AI Mode uses our query fan-out technique, breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf." You will also see these called AI fan-out queries or fan-out queries.
How query fan out works, stage by stage
Each stage below is supported by Google's own published wording.
- Decompose the question. AI Mode works by "breaking down your question into subtopics".
- Generate related sub-queries. The model writes "concurrent, related queries" to fetch more results.
- Run them at the same time. Google describes "issuing a multitude of queries simultaneously".
- Retrieve pages from the index. Google's AI features rely on core ranking systems "to retrieve relevant, up-to-date web pages from our Search index".
- Synthesise the answer. The results are combined to "address the user's query".
What the sub-queries might look like
Google doesn't publish the sub-queries it writes. As an inferred example, a question such as "what do I get on Supabase's free plan?" could fan out into strings I see in my own Search Console data:
- supabase free tier limits 2026
- supabase pricing official 2026
- reddit supabase free tier limits projects 2026
Google hasn't confirmed that link; it is my reading of the strings' shape. The rest of this article shows what fan-out looks like in one site's real Search Console data, and how to check your own.
What does query fan out look like in real Search Console data?
In real data, query fan out looks like a page with a strong average position, a very large impression count and almost no clicks. Everything below is my own Search Console data for aiagencyplus.com. What this example adds is per-query data from a single page.
One page on this site, a tutorial on Supabase free tier limits, recorded the following over 90 days, 08/05/2026 to 05/08/2026.
| Metric | Value |
|---|---|
| Impressions | 493,975 |
| Clicks | 76 |
| Average position | 7.58 |
| CTR | 0.015% |
I expected far more clicks at that position. My first assumption was the obvious one: rewrite the title and meta description. About ten minutes in the query data pointed to a different cause.
This is a measurement problem inside a bigger shift, where search is increasingly answered rather than listed. For the wider picture, read our guide to generative engine optimisation. This piece stays narrow: what your own Search Console numbers are telling you.
What do AI fan-out queries look like in Search Console?
They look like long, compressed keyword strings that already contain the answer, or whole sentences copied from documentation. A person types a short question. A retrieval system types what it expects to find on the page. Sort a page's queries by impressions and the difference usually shows within the top twenty rows.
Here are that page's queries, ordered by impressions.
| Query | Impressions | Clicks | Avg position |
|---|---|---|---|
| supabase free tier limits 2026 | 13,884 | 0 | 5.01 |
| supabase pricing free plan 2026 | 10,717 | 0 | 8.16 |
| supabase free plan limits 2026 500mb two projects unlimited api calls | 8,037 | 0 | 6.60 |
| supabase pricing official 2026 | 6,758 | 0 | 8.75 |
| supabase free tier limits 2026 500 mb two projects unlimited api requests | 5,298 | 0 | 5.74 |
| reddit supabase free tier limits projects 2026 | 3,285 | 0 | 7.01 |
| supabase free tier limits | 2,389 | 3 | 6.74 |
| is supabase free | 1,188 | 2 | 10.73 |
Look at the third row. Nobody types "supabase free plan limits 2026 500mb two projects unlimited api calls" into a search box. That string already holds the answer: 500mb, two projects, unlimited API requests. Whatever generated it knew what it was looking for.
Sentences used as search queries
A second page on the site made the pattern unmistakable. Its top query, with 68 impressions, was this:
> free projects are paused after 1 week of inactivity. limit of 2 active projects.
That reads as a line from Supabase's own published plan limits, full stop included, used as a search query. Others on the same page included "supabase quickstart wait a few minutes create project" (1,249 impressions, no clicks) and "supabase create project takes about 2 minutes" (461 impressions, no clicks).
These read as retrieval queries. My hypothesis is that they are the sub-queries a machine writes when it breaks a person's question into pieces and goes looking for sources. The CTR section below sets out what Google's own attribution rule says and the other explanation that fits.
Four types of fan-out query in my data
Grouping my own strings by what they seem to do gives four types. The labels are my reading of the data; Google publishes no such list.
| Type | Example from my data | Impressions | What it appears to do |
|---|---|---|---|
| Recency-tagged | supabase free tier limits 2026 | 13,884 | Asks for current information |
| Source-targeted | reddit supabase free tier limits projects 2026 | 3,285 | Looks for a named source, such as Reddit or the official page |
| Answer-confirming | supabase free plan limits 2026 500mb two projects unlimited api calls | 8,037 | Checks a specific answer |
| Verbatim documentation | free projects are paused after 1 week of inactivity. limit of 2 active projects. | 68 | Matches a sentence from documentation |
Six signs a query came from a machine
- Compound strings of eight or more words.
- Queries that contain the answer, such as a storage limit or a project count.
- Full sentences, sometimes with punctuation and full stops.
- Phrases that read like product documentation.
- A "reddit" prefix bolted onto a long tail.
- High impressions with zero clicks, row after row.
How are fan-out queries different from keywords?
Keywords are what people type, repeatedly, in volumes a tool can measure. Fan-out queries are written by a model for one question and may never recur. My own table shows the gap: answer-laden strings with thousands of impressions and no clicks, which no keyword tool would list as a target.
Ahrefs describes fan-out queries as zero-search volume queries, since "over 95% receive no recurring searches". That is why a string like "supabase free plan limits 2026 500mb two projects unlimited api calls" never appears in a keyword tool, even though it logs impressions in Search Console.
| Long-tail keyword | Fan-out query | |
|---|---|---|
| Who writes it | A person | A model |
| Measurable search volume | Usually | Over 95% receive no recurring searches (Ahrefs) |
| Contains the answer | Rarely | Often, in my data |
| Example from my data | is supabase free (1,188 impressions, 2 clicks) | supabase free plan limits 2026 500mb two projects unlimited api calls (8,037 impressions, 0 clicks) |
Keyword research still tells you what to write about. Fan-out may explain why the page then logs impressions you never targeted.
Does Google count AI Mode impressions in Search Console?
Yes. On 03/06/2026, Google announced Search Generative AI performance reports in Search Console, in a post by Hillel Maoz and Moshe Samet. On impressions from Google's generative AI features, the post says "this data is included in the overall performance report". They sit inside the same totals you have always read.
The June launch gave those impressions a separate view for the first time. The impressions themselves were already counted in your totals, mixed in with human searches.
Most sites are included by default
Inclusion in Search generative AI features is the default setting. Google's Search Console help puts it this way: "Your site can receive impressions and traffic from these features. This is the default control for all properties." Unless you have changed that control, your property is in.
Eligibility follows normal Search rules. Google's generative AI optimisation guide says that to appear in these features "a page must be indexed and eligible to be shown in Google Search with a snippet". The same guide adds that "a site must be included in Search generative AI features in Search Console to be eligible", which is the default control described above.
What the generative AI reports show
The reports first rolled out to a subset of websites, and they have been available to all sites since 31/08/2026. Google's Search Console help page for the Generative AI performance report now says: "As of August 31, 2026, we've rolled out these insights to all websites worldwide."
A property still sees nothing there if it has too few AI impressions, or none because its control is set to exclude.
The reports show impressions, pages, countries, devices and dates. They carry no clicks, CTR, average position or query breakdown, so the query check later in this article is still how you see fan-out.
Can query fan out lower your CTR?
CTR is clicks divided by impressions. When a large share of your impressions come from queries that earn almost no clicks, the denominator grows and CTR falls. Your real performance with people can stay exactly where it was. Google confirms AI impressions sit in your totals; my hypothesis is that fan-out sub-queries add many low-click rows on top.
Two primary studies have counted the searches behind a single prompt. Nectiv (November 2025, nearly 9,000 prompts run through the Gemini 3 API) found 9.06 fan-out queries per prompt on average, with 59% of prompts at 5-11, 24% at 12-19 and a maximum of 28. Seer Interactive (21/11/2025, 501 prompts through the Gemini 3 API with Google Search grounding forced) found an average of 10.7.
Those figures are about a year old, and counts vary by engine. Peec's April 2026 analysis found averages of 1.4 fan-outs per prompt for Perplexity, 2.1 for ChatGPT and 6.8 for Grok.
Google describes AI Mode as issuing "a multitude of queries simultaneously" for one question. In my data, "supabase free tier limits 2026", "supabase pricing official 2026" and "reddit supabase free tier limits projects 2026" each logged thousands of impressions and no clicks.
My hypothesis is that they are sub-queries of this kind, but Google hasn't confirmed that sub-query strings appear as queries in Search Console. Its Performance report help treats a follow-up question in AI Mode as a new query and applies standard impression and position rules: "All impression, position, and click data in the new response are counted as coming from this new user query." Google documents no logging of sub-queries as queries.
The other explanation is third-party assistants or scrapers running Google searches, which the "reddit … 2026" strings fit just as well. The pattern in my data is consistent with fan-out; it doesn't prove it.
In classic search, one view of a results page gave your page one impression. The CTR column can't tell you which kind you're looking at.
What it did to my numbers
| Impressions | Clicks | CTR | |
|---|---|---|---|
| Reported site total | 676,794 | 504 | 0.074% |
| Top 30 queries by clicks | ~26,016 | 155 | 0.60% |
| That one page, human queries only | 4,512 | 14 | 0.31% |
Around 96% of my reported impressions sit outside the top 30 clicked queries, and the visible part of that tail is almost entirely machine-shaped. The top 30 queries run at 0.60% CTR against a reported site total of 0.074%, about eight times higher.
Where the other clicks sit
Machine-shaped strings don't account for every click. The tail outside my top 30 still holds 349 of the site's 504 clicks, spread thinly across lower-volume queries and queries Search Console doesn't list individually.
On the Supabase page, 14 of the 76 clicks sit on the human queries I classified. The other 62 sit in rows I didn't classify as clearly human and in queries the Queries tab doesn't show, so they can't be sorted either way. That makes 0.31% a lower bound for the page's human CTR.
The query data reset my priorities. Rewriting meta descriptions moved from the top of the list to a sensible small job, sized against human queries rather than the headline figure.
How do I check my own site for query fan out?
Open the Performance report, filter to one page with high impressions and low CTR, and read its queries. Ask of each one whether a person would type it. The whole check takes about ten minutes and needs nothing beyond the Search Console access you already have.
- Open Search Console, go to Performance and pick a page with high impressions and a suspiciously low CTR.
- Filter to that page, then switch to the Queries tab.
- Sort by impressions and read the top twenty.
- Ask one question of each: would a person type this?
- Recalculate CTR on the human-looking queries only.
For step 4, use the six signs listed earlier in this article. If most of your top impressions match them, your CTR is likely inflated by machine-generated queries, whether from Google's fan-out or third-party tools.
Recalculate your real CTR
Take only the queries a person would plausibly type, sum their impressions and clicks, and divide. That is the figure worth setting a target against.
On my Supabase page, the human queries came to 4,512 impressions and 14 clicks. That gives 14 ÷ 4,512 = 0.31%, against the 0.015% that my own Search Console dashboard showed for the whole page (493,975 impressions, 76 clicks, 08/05/2026 to 05/08/2026). Treat your own result as a lower bound too, because the Queries tab doesn't show every query.
You can run the ten-minute check yourself, using the Search Console you already have. Nothing in this article needs a paid tool.
Where I can help
On a discovery call I'll go through your Search Console query data with you, separate human queries from machine ones and work out your real CTR. If the fix is an afternoon's work you can do yourself, I'll tell you that.
If your dashboard shows the same pattern and you want it read with you, Book a Discovery Call.
What should I do about it?
Split your reporting into human queries and AI-driven impressions, treat the AI share as a retrieval signal, and keep improving titles on the queries people actually type. That way you report a real number, see where AI systems use your pages and size the meta work properly.
Re-baseline your reporting
If you report CTR to a client or a board, split human queries from AI-driven impressions first. Otherwise next quarter goes on explaining a decline in a number whose meaning changed underneath it. Measure with more care rather than dropping the numbers and going on feel.
Count AI impressions as retrieval
An impression inside an AI answer means a machine weighed your page as a source. That has value, and it measures something other than a person seeing your brand.
If those impressions are AI-driven, hundreds of thousands of them at positions 5 to 9 mean AI systems pull the page repeatedly when building answers. As a retrieval signal, that is strong. On my site the most-retrieved page by a wide margin is a developer tutorial, which a CTR column would never have shown me.
Being retrieved and being named in an answer are separate measures. Our guide to AI visibility tracking covers how to measure whether assistants mention you.
Keep improving titles
On human queries, that page runs at 0.31% CTR, so there is room to lift it. On my numbers the realistic gain is tens of clicks a quarter, which is still worth an afternoon.
Can you optimise for query fan out in AI Mode?
Partly. You can make each section of a page answer one question plainly, so it still makes sense when lifted out of context, and you can keep pages indexed and snippet-eligible. You can't see or target the fan-out queries themselves, and Google warns against building pages for every variation.
Write sections that stand alone
Write each section so it answers one question plainly, near the top. That helps a reader who lands mid-page, and it makes a section easy to quote. A reader asking about the free plan's storage limit wants one specific sentence that answers it. Google is clear this isn't a technical requirement: "There's no requirement to break your content into tiny pieces for AI to better understand it."
That is the same discipline as writing content AI search engines actually cite. Google's guide points the same way: "Creating content that people find unique, compelling, and useful will likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide."
Don't build a page for every variation
Google's guidance warns that making pages for every fan-out variation to manipulate results breaches its scaled content abuse policy. The same guide says "many suggested "hacks" aren't effective or supported by how Google Search actually works."
You don't see fan-out queries as a list, they change with each question and Google's AI reports carry no query data. Any fan-out keyword list is an inference from patterns, the same as my own reading here.
AI answers reach beyond page one
AI answers cite pages from outside the visible results. In its update published 02/03/2026, Ahrefs found, across 863K keyword SERPs and 4M AI Overview URLs, that 37.9% of URLs cited in AI Overviews also appeared within the first 10 blocks of results for the same query. Those blocks include ads and SERP features.
Ahrefs says this "indicates that Google is selecting far fewer pages straight from the original SERP (~76% in July 2025 vs. ~38% today)." Put simply, about 76% of cited pages came straight from the original results in July 2025, against about 38% in March 2026.
The two Ahrefs studies used different samples, so treat the comparison as indicative. Ahrefs reports the shift without proving its cause, though it fits a system that runs many searches per question. The practical version: answer each question plainly, in its own section, near the top of that section. There is more on the mechanics in our GEO optimisation strategy guide.
Which AI search systems use query fan out?
Google and Microsoft are the providers whose own documentation, cited here, describes generating queries to retrieve content: Google calls it query fan-out, and Microsoft reports grounding queries for Copilot. Third-party pages often name other assistants too, including ChatGPT, Perplexity, Grok and the Gemini app. I haven't cited a statement from those makers in this article, so I don't assert it for them.
Google's AI features page also says both AI Mode and AI Overviews "may use a "query fan-out" technique" to build a response, "issuing multiple related searches across subtopics and data sources".
| System | Official statement cited here |
|---|---|
| Google AI Mode | Yes: Google blog, 20/05/2025 |
| Google AI Overviews | Yes: Google AI features page ("may use") |
| Copilot | Yes: Bing Webmaster Tools AI Performance report (grounding queries), public preview 10/02/2026 |
| ChatGPT, Perplexity, Grok, Gemini app | None cited in this article |
Only searches run on Google Search can appear in Search Console. If another assistant retrieves from a different index, its sub-queries never reach your report, so Search Console can't measure fan-out outside Google.
The nearest first-party view of fan-out queries anywhere comes from Microsoft. Bing Webmaster Tools' AI Performance report, in public preview since 10/02/2026, shows grounding queries, which Microsoft defines as "The key phrases the AI used when retrieving content that was referenced in AI-generated answers." Microsoft added Intents and Citation Share on 16/06/2026. The report covers citations in Copilot and Bing's AI summaries, and Microsoft describes the grounding query data as a sample of overall citation activity.
What can't be proven from Search Console alone?
Search Console shows these strings aren't typed by people, and Google confirms AI impressions sit in the totals. It can't show which system generated a given query. Google doesn't attribute them at query level and its AI reports carry no query data, so exact attribution stays an open question.
Google's AI features are one possible source, given the timing and the company's own statements, but its help treats a follow-up question in AI Mode as a new query under standard impression and position rules, and documents no logging of sub-queries as queries. The other explanation is third-party assistants or scrapers running Google searches, which the "reddit … 2026" strings fit. Search Console has no way to separate the two.
Bing's grounding queries give you a way to compare query shapes. If the grounding queries for a page look like the compressed, answer-laden strings in your Search Console report, that supports the fan-out hypothesis. If they look different, give the third-party explanation more weight.
How do you find fan-out queries for a page?
Read the page's queries in Search Console, using the check above: it is the one method here that shows real fan-out strings rather than predicted ones. Query-expansion tools predict sub-queries instead.
I ran an AI query-expansion tool to see whether it would produce strings of the same shape. It produced tidy natural-language questions such as "What are the limitations of Supabase's free plan?"
The keyword-compressed, answer-laden strings in my data look different. The exact mechanism that writes them is still unknown, and I'd rather say so than guess.
When is a low CTR not caused by query fan out?
When your top queries read like normal human searches, the low CTR is real and usually fixable with better titles and descriptions. Fan-out at my scale needs a page ranking well for a high-volume informational question. Check the query shapes before you put the blame on AI Mode.
Smaller sites
Most sites won't see this at anything like my scale. My affected page ranks well for a high-volume technical question. A local services site with a few thousand impressions a month most likely has a standard CTR problem, and the answer there is better titles.
Low CTR near the top of commercial results
On my site the fan-out pattern sits on informational queries. If you rank near the top for commercial terms and still get few clicks, look at your title, your snippet and what else sits on that results page.
What should I remember about query fan out and CTR?
Query fan out can inflate impressions with machine sub-queries that earn almost no clicks, which drags CTR down while position holds. Google confirms AI impressions are in your totals; that sub-query strings show up as queries is my hypothesis, consistent with my data. Read your query list, recalculate CTR on human queries, report the two separately and treat AI impressions as a retrieval signal.
Frequently asked questions
Why are my Search Console impressions high with no clicks?
Often because a share of those impressions come from Google's AI features, which Google includes in your overall performance totals. My hypothesis is that many come from machine-generated sub-queries that record your page being retrieved as a source. Google hasn't confirmed that sub-query strings appear as queries in Search Console, and third-party assistants or scrapers running Google searches could produce the same pattern. Those queries get almost no clicks, so CTR drops while position holds.
Does Google count AI Overview impressions in Search Console?
Yes. Google confirmed on 03/06/2026 that generative AI impression data is included in the overall performance report, and launched a separate view for it at the same time. Inclusion in these features is the default control for every property.
How do I separate AI impressions from real search impressions?
The Generative AI performance report, available to all sites since 31/08/2026, shows AI impressions by page. It shows nothing for sites with too few AI impressions or with the control set to exclude, and it carries no query data. To split by query, filter the Performance report to one page, read its queries and keep only those that read like human searches. Calculate CTR on that set.
What is query fan out in Google AI Mode?
Query fan out is Google's technique for breaking one question into subtopics, running many related searches at once and writing an answer from what comes back. Google says AI Mode and AI Overviews may both use it. Google's help treats a follow-up question in AI Mode as a new query under standard impression and position rules; it hasn't confirmed that sub-query strings appear as queries in Search Console, though the pattern in my data is consistent with it.
Should I still optimise meta titles and descriptions?
Yes. Judge the result against human queries rather than the headline CTR, and expect gains in proportion to that smaller set. On my own page the human-query CTR was 0.31%, which is the figure a rewrite should move.
