Long Queries in AI Search: What Is Long Query Marketing?

"Long query marketing" is a term practitioners use. It is not an official standard, and it is not a standardised term. Please read this article on that basis. Related labels such as AIO, GEO and LLMO are not used consistently either, and none of them is the name of a formal public standard (§1-1).

This article uses the term in the following sense:

Marketing that prepares information able to answer queries carrying many conditions, and measures whether that information reaches AI answers.

If you take "long" to mean character count, you run into counter-evidence straight away. Google itself writes that in AI Mode it is seeing growth in both short and long queries. So this article reads the subject not in terms of length but in terms of the number of conditions a single query carries. That lens, however, is this article's own analysis. What Google has published is query length. It has not published any figure measuring the number of conditions.

What this article sets out is a number of things that were observed separately: what people type into AI Mode, searching by voice and image, searches that end without a click, the way Google Search's generative AI features may break a query apart to search for it, and the variety of queries that existed long before any of this. These are not measurements of one phenomenon on one metric. We interpret them as sharing a common direction, but that interpretation is set out under its own heading (§8-1).

How to take that is left to the reader. This article commits to two things only: setting the observations side by side, and showing a procedure for measuring where your own company stands today.

Here is the map first.

These are not measurements of the same feature or the same metric. Each was observed for a different product, a different user behaviour or a different search process. Much of the material comes from Google's own announcements; the rest comes from a research organisation and from vendors.

What links them through the lens of the number of conditions is this article's analysis. Each section keeps what has been published apart from how this article reads it.

This article writes only about what has happened. It does not say what comes next. Forecasts sit poorly with a page that is meant to stay up for a long time.

Issues around multi-condition queries that deserve a treatment of their own are left to separate articles. This one is the overall map.

1-1. On the name

The labels AIO, GEO and LLMO cover different ground depending on who is using them. None of them is the name of a formal public standard.

They are not all on the same footing, though. For the term GEO (Generative Engine Optimization), there is a peer-reviewed paper accepted at the international conference KDD in 2024. Google's guide on generative AI features also introduces AEO and GEO as terms in general use, and then explains that, from Google Search's point of view, optimising for generative AI search is still optimising for the search experience, and so it is still SEO.

"Long query marketing" has no paper or guide of that kind behind it. The sense in which this article uses it is limited to the one sentence set out at the top.

2-1. The figure Google has published

In a post published on 19 May 2026, Google wrote the following about the use of AI Mode in the United States:

the average AI Mode search is triple the length of a traditional Search query

On average, a search in AI Mode is three times as long as a traditional search.

Three limits need to be stated up front.

First, this is a US figure. The post covers the United States. Within the scope of this article's research, no primary data measuring query length in Japanese was found. Nor can a word count be applied to Japanese as it stands.

Second, the method of calculation has not been published. Because the post does not say what was counted as length, the figure cannot be set alongside other studies and compared.

Third, this is about what happens inside AI Mode. It is not primary evidence that search as a whole has become longer.

2-2. The same post says the opposite as well

This is the important part. The same post also contains this sentence:

we're seeing growth in both short and long queries in AI Mode

Short queries and long queries are both growing. Google says so itself.

In other words, short queries have not been replaced by long ones. What has grown is the range. Generalising to "search has got longer" collides with this sentence.

2-3. This article's analysis: read by number of conditions, not length

What follows is this article's analysis, not a published fact.

What Google measured is query length. It did not measure conditions. The unit of length has not been published either.

This article still adopts the lens of the number of conditions because a long query can contain several conditions, such as industry, region, budget or timing. Saying only "queries got longer" collides with the sentence about short queries also growing. Looked at by number of conditions, short queries with few conditions and long queries with many conditions growing side by side fit within a single view.

This view does not mean that Google's figure shows conditions increasing. A long query is not necessarily a query with many conditions. Some queries become long through preamble or rephrasing. This article has not found any primary data measuring the number of conditions a query carries.

The same post also covers the form of input, in its list headed "Here are some of the key insights about AI Mode in the U.S. one year since launch:":

More than one in six searches in the U.S. now use voice or images, with image searches growing over 40% month-over-month

Google lists, among its key insights about AI Mode in the U.S., that more than one in six searches in the U.S. use voice or images, with image searches growing by more than 40% month over month. That sentence alone does not establish whether these are shares of all Google searches or of AI Mode searches; this article does not treat them as shares of all searches.

This may bear on the number of conditions because when the form of input changes, the shape of the query can change too. Someone who keeps it short when typing may be more inclined to speak in full sentences. Attach an image, and conditions that the image conveyed can enter the query without any words at all.

A query such as "something like the thing in this photo, but cheaper, that I can pick up today" is quicker to say with an image and a voice than to build from text alone. One reading is that conditions increased because input became easier. That, too, is this article's reading. Google's figure concerns searches that use voice or images, not the number of conditions those queries carry.

This figure is also a US one. A breakdown for Japan could not be confirmed within the scope of this article.

3-1. With voice, what is longer is the "conversation"

There is one more published figure on voice. In a post on 20 May 2025, Google wrote the following about Gemini Live:

the conversations are five times longer than text-based conversations on average

Conversations are, on average, five times longer than text-based ones.

This needs careful reading. What is described as longer is the conversation, not each individual query. Neither the unit nor the method behind "length" has been published. It cannot be restated as "asking by voice makes a query five times longer."

From the standpoint of the number of conditions, this much can be said: if an exchange continues, conditions can be added along the way rather than stated all at once. Even when the first query is short, assumptions can be added later. Whether conversations are long because conditions are being added to the same topic, or because they move on to other topics, cannot be told from the published figure.

3-2. What can and cannot be said so far

Not writing the right-hand column is how this article is written. Where the article does use something from the right-hand column as a lens (the number of conditions), it is kept apart as analysis, as set out in §2-3.

Next come observations of how search results are used.

In an analysis published on 9 June 2026 by Rand Fishkin of SparkToro, 68.01% of Google searches in the United States from January to April 2026 ended without a click (Similarweb's clickstream panel). The same post gives 60.45% for 2024 (Datos's panel).

These two figures come from different panels. The author writes, "these aren't the same users or devices," noting that neither the users nor the devices are the same and that the demographics do not fully match either. The difference between 2024 and 2026 is not a change tracked within the same population.

A study published by Pew Research Center on 22 July 2025 analysed Google searches (68,879 of them) made in March 2025 by 900 US adults, using records of their actual browsing. When an AI summary appeared, a link in the traditional search results was clicked on 8% of visits; when there was no summary, the figure was 15%. A link inside the summary was clicked on 1% of visits. On pages where a summary appeared, 26% of visits ended browsing there, more than the 16% on pages without one.

The background to these figures is covered in the article on the overall picture of AIO.

4-1. This article's hypothesis: does narrowing move into the query?

What follows is a hypothesis.

If people do not narrow down by conditions on the destination site, one possible flow is that the conditions go into the first query instead. This article considers that the more the steps of looking at results, visiting a site and selecting filters are skipped, the more this may create an incentive to bring that narrowing forward into the wording of the query.

The data on searches ending without a click, however, does not measure this causal link directly. Searches that end without a click include cases where the answer on the page was enough, cases where the person searched again, and cases where they left unsatisfied. Pew's 26% likewise shows only that browsing ended there; it does not show whether queries became more detailed.

If the hypothesis is right, the work of narrowing moves into the query, and if the query cannot be answered once it gets there, the answer is produced with those conditions left unmet. Whether the hypothesis is right cannot be tested with the material within this article's scope.

What this section describes is Google's official explanation of the generative AI features in Google Search (AI Overviews and AI Mode). This article does not generalise about the internal workings of other companies' AI products.

5-1. The definition Google publishes

Google's guide to optimising for generative AI features (last updated 10 July 2026) defines query fan-out as follows:

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

In Google Search's generative AI features, the model generates several related queries at the same time and issues them, in order to answer the user's query.

A separate page (AI features and your website, last updated 10 December 2025) states that AI Overviews and AI Mode may use this technique. It does not say they always do.

So when Google Search's generative AI features receive a query with many conditions, it is not necessarily handled as one search string as it stands. It may be broken apart and searched for in separate pieces.

The mechanism is covered in more detail in the article on what AI cites when it describes your company. The wording in Google's official documentation is collected in the article on how long information stays in AI answers.

5-2. If queries are broken apart, should you add more pages?

If the pages are added in order to manipulate, the answer is no. Google says so itself. From the same guide:

While it might be tempting to create separate content for every possible variation of how people might search (for example, by focusing on other queries that people have asked, or fan-out queries), doing so primarily to manipulate rankings or generative AI responses in Google Search violates Google's scaled content abuse spam policy.

Creating separate content for every possible variation of how people might search, when done primarily to manipulate rankings or generative AI responses in Google Search, is stated to violate the scaled content abuse spam policy.

Note the condition. Creating a separate page that is useful to people is not prohibited in itself. What is prohibited is mass-producing a page per variation with manipulation as the main purpose.

Responding to multi-condition queries is therefore not a question of page count. It is a question of substance: whether information that answers the conditions a query carries actually exists.

5-3. No basis for assuming an entirely separate ranking system for AI

The same guide also says this about the generative AI features:

our generative AI features on Google Search are rooted in our core Search ranking and quality systems

Google Search's generative AI features are rooted in its core Search ranking and quality systems. The same guide also explains that, through an approach called retrieval-augmented generation (RAG), they use the existing ranking systems to retrieve relevant pages.

From this explanation, it can be said that for Google Search's generative AI features there is no basis for assuming a ranking system entirely separate from existing search. That does not mean no further processing is added while the answer is built, such as retrieving and selecting information or generating text. How far that goes cannot be told from this passage.

5-4. A setting available to site owners

There is one more official fact relevant to whether information gets through. On Google, site owners can use Search Console to choose whether their site's links and content are included in, or excluded from, generative AI features such as AI Overviews and AI Mode. Google states that it rolled this control out to all websites worldwide on 31 August 2026.

Google explains that this setting is not used as a ranking or inclusion signal affecting other parts of Search. If your information does not appear in the generative AI features, the first thing to check is this setting.

This section differs in kind from the ones before it. It deals not with something new but with something that was already there.

One caveat first. The figures in this section are figures about the query side. They do not show that information on the answering side has grown or become more detailed. This article has not been able to confirm any primary material showing whether answering information has become more or less detailed.

6-1. 15% of daily searches are queries not seen before

In a post on 25 October 2019, Google wrote:

15 percent of those queries are ones we haven't seen before

Of the searches Google sees every day, 15% are queries it has not seen before.

What this figure shows is that the variety in how queries are phrased was already large before generative AI features arrived in search.

It does not, however, support the claim that "there were already huge numbers of finely conditioned queries." A query can be new for many reasons besides a combination of conditions: new events, new words or proper nouns, spelling variation, rephrasing, and so on. And "not seen before" does not mean "searched only once."

6-2. The published figure was 15% in both 2019 and 2022

This belongs on the counter-evidence side too.

In a post on 3 February 2022, Google wrote:

In fact, 15% of searches we see every day are entirely new.

At least in the published figures, it was 15% in both 2019 and 2022. Nothing says the ratio went up. This is not evidence of an increase.

At the same time, two rounded figures at two points in time being equal does not mean the ratio stayed the same throughout. What can be said is that the published figures do not show an increase.

In short, it cannot be written that "new queries have increased, so a response is now needed."

6-3. What this section supports

What this section supports is that the variety in how queries are phrased was already large, and no more.

On top of that, as §5 describes, Google officially explains that Google Search's generative AI features may break a query apart to search for it. As §4 describes, a high share of searches is observed to end without a click. Connecting these to read "the way answers reach an already varied set of queries has changed" is this article's interpretation. It is not a fact confirmed on a common metric.

6-4. What this section does not cover

This article's register still holds several developments that this section could not cover: people using memory features to have conditions remembered, agent systems searching with specified conditions, changes to conversational input surfaces themselves, the incorporation of first-hand experiences, and the increasing granularity of product data.

Because the material on these has not been checked, this article does not deal with them. It does not fill the gaps by guesswork. Separate articles are planned for those that can be confirmed.

Please do not skip this section. There are several pieces of material that cut against this article's reading.

The "Source type" column shows what kind of source each item is. It is not a ranking of how strong the evidence is. Google's announcements are figures from the company that runs the product, and some do not state how they were calculated. Pew's study observed records of actual browsing. Being able to reach the original is a different matter from being strong evidence. What this article distinguishes is whether or not it could check the original directly.

The first point is that Google itself says both. The same Google post this article quoted in §2 says that short queries are growing as well. You cannot quote only the convenient sentence.

The second is that an increase in the share of new queries cannot be claimed. As §6-2 sets out, the published figure was 15% in both 2019 and 2022. Some of what this article sets out was already there before.

The third is that the share of searches reaching Google's AI Mode is still small. In the same SparkToro analysis, 0.34% of Google searches in the United States from January to April 2026 reached AI Mode. That is a figure for Google's AI Mode, not for AI search as a whole. It does not include use of AI products other than Google's.

The fourth is that the zero-click comparison rests on different panels. As set out in §4, 68.01% and 60.45% do not come from the same population. The size of the rise cannot be read directly as the size of a change.

The fifth is that the measurement side was unstable. In Search Console, from 13 May 2025 to 27 April 2026, a logging issue meant impressions were not reported correctly. Click-through rate and average position, which depend on impressions, were affected as well. A comparison across this period using values saved at the time, or data that has not been corrected, can pick up the effect of the issue, so care is needed. This article does not state how far the historical values now shown in Search Console have been corrected.

The sixth is that the average traditional query is still short. In an analysis Semrush published on 30 July 2025, covering roughly 69 million Google Search sessions (clickstream) on desktop in the United States between 1 May and 5 July 2025, traditional search queries averaged 4.0 words and AI Mode queries averaged 7.22 words. Even with long queries growing in AI Mode, the average traditional query is a few words. Note that the ratio between AI Mode and traditional search in this analysis is roughly twice, which does not match the three times Google published. The periods and devices differ, and Google has not published how it calculates length, so the two figures cannot be compared as they stand (§2-1). This article does not judge which is correct.

In addition, there is a study reporting that AI search leans toward citing third-party authoritative sources rather than a brand's own properties. It is a non-peer-reviewed preprint, but it suggests that even if you prepare information that answers the conditions yourself, it is not necessarily the source that gets cited.

And this article could not find primary data measuring query length in Japanese. That is not a claim that none exists; it records that none was found. Unless otherwise stated, the figures this article cites are US figures.

7-1. What remains after the counter-evidence

What still stands after all this is the individual observations laid out in the §1 map, and our interpretation that they share a common direction. The interpretation is set out separately in §8-1. This article claims nothing beyond that.

If you judge the counter-evidence to be stronger, that is a legitimate reading too. This article does not hide the opposing material in order to steer toward a conclusion. It puts both sides in front of you, as needed for a judgement.

Whichever judgement you reach, measurement is not the only way to learn where your company stands. It is one of several (§8-4).

Here are the observations once more.

In AI Mode in the United States, the average query is three times the length of a traditional search, and short and long queries are both growing. More than one in six searches in the United States use voice or images (a figure Google listed among its AI Mode insights for the U.S.; not established as a share of all searches). A high share of searches end without a click, and link clicks fall when an AI summary appears. Google officially explains that Google Search's generative AI features may break a query apart to search for it. And queries not seen before have long made up 15% of daily searches.

8-1. Our interpretation: a common direction

This is not a summary of facts. It is our interpretation.

We interpret these observations as sharing a common direction: the situations in which a query carrying conditions arrives, as it is, at the entrance to an AI answer are widening.

They are not, however, measurements of the same phenomenon on the same metric. Query length, form of input, whether a click happens, how the search process is described, and the share of never-seen queries are each separate metrics. No single figure supports this interpretation.

8-2. What it means that these are separate observations

These are not measurements of the same feature or metric; each was observed for a different product, a different user behaviour or a different search process. Much of the material comes from Google's own announcements; Pew is a research organisation, and SparkToro's is a vendor observation.

They are not the same feature as one another. But since much of the material comes from the same company, they cannot be called unrelated either. Separate observations lining up can be material for an interpretation, but it does not prove the interpretation right.

8-3. Still, no firm conclusion

What to draw from this is left to the reader. This article does not write "so you should respond to multi-condition queries." It does not have the grounds to write that.

The counter-evidence in §7 has not gone away. The average traditional query is still a few words, the share of searches reaching Google's AI Mode is still small, and no measurement in Japanese has been found. Writing "you should respond" while these remain would be a sales pitch, not an argument.

What this article took on was to set separately observed things side by side, in the same table, and no more.

8-4. Measurement is one input among several

With that position of not drawing a firm conclusion in place, here is the practical side.

If you want to know directly how your company is treated right now, measurement is one input to your judgement. Which way things go is not known, but how your company is treated inside AI answers today can be recorded if you measure it.

Measurement is not the only input. Whether your target customers actually use AI, the figures in Search Console (§10-5), conversations with customers, and estimates of cost and benefit are inputs too. How to use measurement results is something to decide together with those.

From here on, this is our position. It sits under its own heading because it is not a summary of facts.

We, Vaigate Inc., operate Vaipm, which measures AI-space perception. Below are the principles we hold when designing measurement, set out along the lines of this article's argument.

First, we measure by varying how the question is asked. A query that asks only about a company by name cannot show how that company is treated when the question is asked without naming it. The way a question is asked changes the AI's answer.

Second, we measure by repeating the same question. Generative AI does not return the same answer to the same question every time. A single result does not establish a state.

Third, we fix the conditions and disclose them: the number of repetitions, not carrying over prior exchanges, spanning multiple AI engines, and the language. A figure published without these cannot be interpreted by the reader.

Fourth, we do not design toward increasing the number of pages. As set out in §5-2, Google states that creating content for each variation of a query, when done primarily to manipulate rankings or generative AI responses, violates its spam policy. What we measure is not page count but how a company is treated inside the answer.

9-1. Questions this design does not answer

To be candid, what this design shows is how your company is treated right now in response to the questions we designed, and no more.

There are three things it does not tell you.

First: if you make an improvement, how long it takes to be reflected in AI answers is not built into the design. Google explains that crawling can take from several days to several months. The intervals for fetching and re-fetching differ by provider, and what each discloses varies.

Second: how the way you are treated inside answers connects to business outcomes cannot be told from this measurement. What is measured is how you are treated inside AI answers. What lies beyond that belongs to other metrics.

Third: what queries real users type cannot be told from this measurement. What is measured is the answers to questions set in advance.

We consider not overstating what measurement can answer to be part of the design.

What follows is a general procedure readers can try for their own company. It does not describe the features of our product. Our own way of measuring is set out separately in §10-7.

10-1. Measure with four ways of asking

The fourth ties directly to this article's subject. A company that appears when named can fail to appear in a query that stacks three conditions. That gap is a starting point for looking for the cause. There is more than one candidate cause, however (§10-3). The existence of a gap alone does not tell you whether you have responded adequately.

10-2. Add conditions one at a time

The fourth way of asking is easier to read if you add conditions one at a time rather than putting them all in at once.

Seeing at which added condition you stopped appearing gives you a lead for forming a hypothesis about the cause. A result that says only "we did not appear in the query with every condition in it" gives you no lead at all.

This procedure does not identify the cause, though. Generative AI answers also change with word order, phrasing, how conditions combine, and the search results at that moment. Adding conditions one at a time is only an aid to narrowing down candidate causes.

10-3. What to record

Whether you appeared or not is not enough. Record what was cited as the source, too. Was it your own page, a third-party roundup, or a competitor's comparison article?

If the cited sources change as conditions are stacked, there are several possible causes:

• Information that answers that condition is missing or thin on your side
• The information exists but does not rank high in search results
• A different source is chosen at the stage of retrieving information or selecting sources
• Your information is being treated as outdated
• Outside authoritative sources are being preferred (the lean reported by the study cited in §7)

Conversely, if your own page keeps being cited as conditions are added, then for those conditions your information is getting through, at least at that point in time.

There are three things to record: whether you appeared, roughly what position you were treated in, and what was cited as the source. Drop the third, and you lose the lead for working out which of the candidates above applies.

10-4. Why once is not enough

Generative AI answers vary. Ask the same question on the same day and the text that comes back is not the same.

Not appearing in one measurement does not prove that you do not appear. Equally, appearing once does not prove that you keep appearing. A single result is not a state; it is one output at one moment.

When you measure, therefore, the following conditions need to be fixed and recorded:

• Number of repetitions — how many times you asked. Once is a trial, not an observation. The fewer the repetitions, the larger the uncertainty in any share you read from the results
• Not carrying over prior exchanges — if login state or history remains, you cannot tell whether you are measuring your company or the history of the person measuring. Not carrying over history, though, does not remove conditions such as region, the search index, or changes on the provider's side
• Which AI you measure with — a study reports that AI search services differ markedly from one another in the range of sources they cite, the freshness of information, stability across languages, and sensitivity to phrasing (the non-peer-reviewed preprint cited in §7). Whether to measure across several AI services depends on what you want to know
• The names of the product and model used, the measurement date, and the date and time each answer was retrieved — the same product can change over time
• Whether web search is enabled, and the region — whether search is involved can change how sources are cited
• Language — as §7 notes, no measurement in Japanese was found within this article's scope. If you measure for your own company, record the language you measured in
• Question design — the four ways of asking in §10-1, and the way conditions are added in §10-2

Then disclose them. A figure published without its conditions cannot be read back even inside your own company a year later. It cannot be used to compare with competitors, or to compare before and after.

10-5. What Search Console can and cannot measure

For Google, there are official figures as well. Search Console's "Generative AI performance report" shows how many times links to your site were displayed in AI Overviews and AI Mode, broken down by page, country, date and device. Google states that it rolled this out to all websites worldwide on 31 August 2026.

Google also explains that Search Labs experiments are not included, that when results from the same site appear twice in one generative AI feature they count as one, and that a certain number of impressions is needed before data is shown.

10-6. The two methods look at different things

Search Console's figures are impressions to real users that Google recorded for your site. The procedure in this article, and our own measurement, record AI answers to questions set in advance; they tell you neither what real users typed nor how many times anything was actually shown.

Because they look at different things, neither replaces the other. If you want to know how your links are displayed in Google Search's generative AI features, use Search Console; if you want to fix the shape of the question and see how you are treated inside AI answers, use the procedure in §10-1 to §10-4. You can also look at both side by side.

10-7. How we measure

This is our position.

Vaipm measures AI-space perception through a total of 25 stateless queries across multiple AI engines.

Stateless here means measuring without carrying over login state or prior exchanges. It does not mean removing every condition that could affect an answer, such as region or the search index.

The count of 25 is the observation design we have adopted. It is not an optimal value derived from research. Any share read from 25 results carries uncertainty that depends on the number of trials. It makes a tendency easier to read than a single output does, but it does not show a settled state.

Vaipm asks about a subject in four ways: by name; by category, without naming it; alongside competitors; and by the name of a specific initiative. The way a question is asked changes the AI's answer.

The "ask with stacked conditions" procedure in §10-1 and §10-2 is a general method set out in this article, and it is not the same as the ways Vaipm asks. Recording how your company is treated inside AI answers, repeatedly and with varied ways of asking — that is the extent of what we take on through measurement.

"Long query marketing" is a practitioner's term, not an official standard. This article used it to mean marketing that prepares information able to answer queries carrying many conditions, and measures whether that information reaches AI answers.

What Google has published is that queries are getting longer in AI Mode in the United States, and that short queries are growing too. The lens of the number of conditions is this article's analysis. Form of input, searches ending without a click, Google's explanation that its generative AI features may break a query apart to search for it, and the variety of queries that was already there: these are separate observations, and reading a common direction in them is our interpretation.

The counter-evidence remains. The average traditional query is still a few words, the share of searches reaching Google's AI Mode is still small, the zero-click comparison rests on different panels, and no measurement in Japanese has been found. There is not yet enough material to act on one side of the picture alone.

The response is not a matter of page count. Google states that creating content for each variation of a query, when done primarily to manipulate rankings or generative AI responses, violates its spam policy.

Some of what this article set out was already there before. Never-seen queries have long made up 15% of daily searches, and the published figures do not show an increase.

If you want to know how your company is treated now, measuring under fixed conditions is one input to your judgement. Search Console shows how your links are displayed in Google Search's generative AI features. To see how you are treated inside AI answers to questions you set, one approach is to measure repeatedly with varied ways of asking.

This article did not say which way things are heading. It does not have the grounds to. How your company is treated today can be recorded if you measure it. How to use that is something to decide together with your other inputs.

Q1. Is "long query marketing" an official term?

No. It is a term practitioners use. It is neither an official standard nor a standardised term. AIO, GEO and LLMO are not used consistently either, and none of them is the name of a formal public standard. For the term GEO, however, there is a peer-reviewed paper accepted at the international conference KDD in 2024, and Google's guide on generative AI features introduces AEO and GEO as terms in general use while explaining that, from Google Search's point of view, this is still SEO. "Long query marketing" has no paper or guide of that kind behind it. This article uses it only in the sense defined at the top: marketing that prepares information able to answer queries carrying many conditions, and measures whether that information reaches AI answers (§1-1).

Q2. Have searches become longer?

What primary sources support is that, in the United States, AI Mode searches are on average three times the length of traditional searches. That figure has limits: it covers the United States, the method of calculation has not been published, and it describes what happens inside AI Mode. The same post says that short and long queries are both growing, so short queries have not been replaced by long ones; what has grown is the range. Within the scope of this article's research, no primary evidence was found that search as a whole has become longer. And what Google measured is length, not the number of conditions (§2-1, §2-2, §2-3).

Q3. Should we create a page for every query we expect people to ask?

Google states that creating separate content for every variation of how people might search violates its spam policy when done primarily to manipulate rankings or generative AI responses (§5-2). Creating a separate page that is useful to people is not prohibited in itself. What is prohibited is mass-producing a page per variation with manipulation as the main purpose. Responding to multi-condition queries is therefore not a question of page count but of substance: whether information that answers the conditions a query carries actually exists. This article recommends looking not at page count but at whether information that answers the conditions exists.

Q4. Do we need to optimise for a separate AI ranking?

Google explains that the generative AI features in Google Search are rooted in its core Search ranking and quality systems, and that through retrieval-augmented generation (RAG) they use the existing ranking systems to retrieve relevant pages (§5-3). There is no basis for assuming a ranking system entirely separate from existing search. That does not mean no further processing is added while the answer is built, such as retrieving and selecting information or generating text; how far that goes cannot be told from Google's description. This article does not generalise about the internal workings of AI products other than Google's.

Q5. Is the same thing happening in Japan?

This is not known. Unless otherwise stated, the figures this article cites are US figures. The AI Mode figure (three times the length of a traditional search) covers the United States, and so does the voice-or-image figure of more than one in six searches (a figure Google listed among its AI Mode insights for the U.S.; not established as a share of all searches); a breakdown for Japan could not be confirmed. Within the scope of this article's research, no primary data measuring query length in Japanese was found, and a word count cannot be applied to Japanese as it stands. That is not a claim that no such data exists; it records that none was found (§2-1, §3, §7).

Q6. Where should we start?

If you want to know how your company is treated now, one approach is to measure, repeatedly and under fixed conditions, whether your company appears in queries that stack conditions (§10-1 to §10-4). Adding conditions one at a time, rather than all at once, shows at which condition you stop appearing, which gives a lead for forming a hypothesis about the cause (§10-2). Record what was cited as the source as well as whether you appeared. If you do not appear, there is more than one candidate cause (§10-3). Measurement is one input to your judgement, to be used together with other inputs (§8-4).

Q7. Is Search Console's generative AI report enough?

That depends on what you want to know. In Search Console you can see how many times links to your site were displayed in AI Overviews and AI Mode, by page, country, date and device. On the other hand, the queries users typed are not shown, and AI products other than Google's, and mentions of your name without a link, are out of scope. Search Console records impressions to real users; the procedure in this article records AI answers to questions set in advance. Because the two look at different things, neither replaces the other; use each according to your purpose (§10-5, §10-6).

Q8. Does searching by voice or image increase the number of conditions a query carries?

What primary sources support is that more than one in six searches in the United States use voice or images (a figure Google listed among its AI Mode insights for the U.S.; not established as a share of all searches), and that in Gemini Live conversations are on average five times longer than text-based ones. That people may be more inclined to speak in full sentences, and that conditions can be added along the way, is this article's reading. Google's figures do not show the number of conditions a query carries. And what is described as longer is the conversation, not each individual query (§3, §3-1).

Q9. Can it be said that searches ending without a click have increased?

In SparkToro's analysis, 68.01% of Google searches in the United States from January to April 2026 ended without a click. The 2024 figure is 60.45%, but the two come from different panels with different users and devices, not from tracking change in the same population. In Pew's study, when an AI summary appeared, a link in the traditional search results was clicked on 8% of visits, against 15% when there was no summary. The view that narrowing moves into the query remains this article's hypothesis (§4, §4-1).

Q10. Does "15% of daily searches are queries not seen before" mean there are many finely conditioned queries?

No, it does not go that far. What the figure shows is that the variety in how queries are phrased was already large. A query can be new for many reasons besides a combination of conditions, such as new events, new words, or spelling variation. "Not seen before" also does not mean "searched only once." And the published figure was 15% in both 2019 and 2022, which is not evidence of an increase. Two equal rounded figures do not mean the ratio stayed the same in between either; what can be said is that the published figures do not show an increase (§6-1, §6-2).

Q11. Can we keep our site out of Google's generative AI features?

On Google, site owners can use Search Console to choose whether their site's links and content are included in, or excluded from, generative AI features such as AI Overviews and AI Mode. Google states that it rolled this control out to all websites worldwide on 31 August 2026. Google explains that this setting is not used as a ranking or inclusion signal affecting other parts of Search. If your information does not appear in the generative AI features, the first thing to check is this setting (§5-4).

Q12. If we measure once and our company does not appear, does that mean AI is not treating us at all?

Not necessarily. Generative AI answers vary, and asking the same question on the same day does not return the same text. Not appearing once does not prove that you do not appear, and appearing once does not prove that you keep appearing. When you measure, the number of repetitions, not carrying over prior exchanges, the product used and the measurement date, the language, the question design and so on need to be fixed and recorded. If you do not appear, there is more than one candidate cause (§10-3, §10-4).

1. Google, "How AI Mode is changing and expanding the way people search," 19 May 2026 (key insights about AI Mode in the United States one year after launch). That the average AI Mode search is triple the length of a traditional Search query; that short and long queries are both growing in AI Mode; and, in the same list of AI Mode insights, that more than one in six searches in the U.S. use voice or images, with image searches growing over 40% month-over-month (the sentence does not establish whether these are shares of all searches or of AI Mode searches). The method for calculating length is not published. https://blog.google/products-and-platforms/products/search/ai-mode-us-insights/
2. Google Search Central, "Optimizing your website for generative AI features on Google Search," last updated 10 July 2026. The definition of query fan-out; the relationship between creating content for every possible search variation and the scaled content abuse spam policy (where this is done primarily to manipulate rankings or generative AI responses); that the generative AI features are rooted in core Search ranking and quality systems; that retrieval-augmented generation (RAG) uses the existing ranking systems; and the explanation of the terms AEO and GEO. https://developers.google.com/search/docs/fundamentals/ai-optimization-guide
3. Google Search Central, "AI features and your website," last updated 10 December 2025. That AI Overviews and AI Mode may use query fan-out, and that crawling can take from several days to several months. https://developers.google.com/search/docs/appearance/ai-features
4. Rand Fishkin, "In 2026, Less than One Third of Google Searches Still Send a Click," SparkToro, 9 June 2026. That 68.01% of Google searches in the United States from January to April 2026 ended without a click (Similarweb clickstream panel); the 2024 figure of 60.45% (Datos panel); the author's note that the two are not the same users or devices; and that 0.34% of searches reached AI Mode in the same period. A vendor analysis. https://sparktoro.com/blog/in-2026-less-than-one-third-of-google-searches-still-send-a-click/
5. Pew Research Center, "Google users are less likely to click on links when an AI summary appears in the results," 22 July 2025. An analysis of the March 2025 browsing records of 900 US adults (68,879 Google searches). When an AI summary appeared, a traditional search-result link was clicked on 8% of visits (15% without a summary); a link inside the summary was clicked on 1%; browsing ended on 26% of pages with a summary (16% on pages without one). https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/
6. Google, "Understanding searches better than ever before," 25 October 2019. That 15% of daily searches are queries not seen before. https://blog.google/products/search/search-language-understanding-bert/
7. Google, "Gemini App: 7 updates from Google I/O 2025," 20 May 2025. That Gemini Live conversations are on average five times longer than text-based conversations. The unit and method of calculation are not published. https://blog.google/products-and-platforms/products/gemini/gemini-app-updates-io-2025/
8. Google, "How AI powers great search results," 3 February 2022. That 15% of daily searches are entirely new. https://blog.google/products-and-platforms/products/search/how-ai-powers-great-search-results/
9. Google Search Console Help, "Data anomalies in Search Console." That from 13 May 2025 to 27 April 2026 a logging issue meant impressions were not reported correctly, and click-through rate and average position were also affected; that clicks were not affected; and that the issue has been resolved. https://support.google.com/webmasters/answer/6211453?hl=en
10. Mahe Chen, Xiaoxuan Wang, Kaiwen Chen, Nick Koudas, "Generative Engine Optimization: How to Dominate AI Search," arXiv 2509.08919, 10 September 2025 (non-peer-reviewed preprint). That AI search leans toward third-party authoritative sources over a brand's own properties, and that AI search services differ markedly from one another in the range of sources cited, freshness, cross-language stability and sensitivity to phrasing. https://arxiv.org/abs/2509.08919
11. Google Search Console Help, "Generative AI performance report (Search)." That impressions of links to your site in AI Overviews and AI Mode can be viewed by page, country, date and device; that it was rolled out to all websites worldwide on 31 August 2026; that the queries users typed are not shown; that Search Labs experiments are not included; that multiple results from the same site in one feature count as one; and that a certain number of impressions is needed before data is shown. https://support.google.com/webmasters/answer/16984139?hl=en
12. Google Search Console Help, "Search generative AI control." That site owners can use Search Console to include or exclude their site's links and content from generative AI features such as AI Overviews and AI Mode; that it was rolled out to all websites worldwide on 31 August 2026; and that it is not used as a ranking or inclusion signal affecting other parts of Search. https://support.google.com/webmasters/answer/16908024?hl=en
13. Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan, Ameet Deshpande, "GEO: Generative Engine Optimization," arXiv 2311.09735 (accepted to KDD 2024). A peer-reviewed paper on the term GEO. https://arxiv.org/abs/2311.09735
14. Luke Harsel, "Google AI Mode's Early Adoption and SEO Impact," Semrush, 30 July 2025. An analysis of about 69 million Google Search sessions (clickstream) on desktop in the United States between 1 May and 5 July 2025. That traditional search queries averaged 4.0 words and AI Mode queries averaged 7.22 words. A vendor analysis. https://www.semrush.com/blog/google-ai-mode-seo-impact/

Sources 1 to 13 were checked directly against the primary source on 23 September 2026, and source 14 on 24 September 2026. No primary data on query length in Japanese could be found within the scope of this article's research.