SEO Keyword Research & Search Demand

Keyword research is finding out what your audience is trying to accomplish, in the language they use to try it, and deciding which of those needs is worth doing something about. That is the whole definition, and notice what is missing: a tool, a volume threshold, and the word "rank".

The belief I want to kill first is that a keyword is a thing that exists. It is not. A query exists: one person typed something, once, probably badly, at eleven at night. A keyword is what the industry calls a slice of those queries once somebody has grouped them and attached an estimated number. It is a unit of measurement, and treating a unit of measurement as the unit of work is the root of almost every mistake below.

Five words get used interchangeably here and they are five different sizes of the same thing. Getting them apart is the single highest-value thing on this page, because once you can see the difference, "one keyword, one page" stops being an unfashionable opinion and becomes visibly a category error.

Take five real strings: how to learn seo, learn seo, seo learning roadmap, seo course for beginners, where to start seo. Those are five queries. They are one need. A person who typed any of them would be satisfied by the same page, and would be mildly annoyed to find five pages from the same site competing for them.

This matters more in 2026 than it did in 2016, and not because of anything an SEO said. Google's own guidance on generative features states directly that you do not need to worry that you have not captured every variation of how someone might seek content like yours. And Search Console Insights now ships a feature called Query groups, which groups similar queries together and names the group after its best-performing query, because there are so many different ways to write the same question that a raw query list stopped being readable.

Google is grouping queries for you, in its own product, and telling you the variations do not need chasing. Any advice that still treats a keyword as the unit of work is arguing with the search engine about how the search engine works.

Keyword research still matters because it is the only cheap way to find out what a market wants in its own words, and because it is where you decide what not to build. Both halves of that are worth more than the traffic estimate everybody focuses on.

The opponent here is the 2026 version of "SEO is dead": search is fragmenting, AI answers the question, so why bother sizing demand. It is a reasonable thing to think and it collapses the moment you look at the demand sitting underneath a complex prompt, which is chapter twenty-one. The short version: the needs did not go anywhere, and your database still prices them.

Three jobs, and only one of them is about traffic.

• Understand demand. What are people trying to do, and how many of them. This is the one everybody does.
• Understand language. How customers describe the problem, which is almost never how your team describes the product. This is free, it takes an afternoon, and it improves your homepage more than a year of blogging will.
• Make decisions. Which opportunities deserve a page, a feature, a tool, a video, a section, or nothing at all. This is the one that gets skipped, and it is the one the whole exercise exists for.

Search Engine Journal puts the connective sentence better than I can: keyword research is an extension of understanding your audience by first considering their needs. Which makes it market research, conducted through search behavior, and not tool usage.

The practical consequence is a test you can apply to your own last session. If it ended with a file, it did not finish. If it ended with somebody saying "so we are building these three things and deliberately not these four", it did.

Answer seven questions about the business before you open anything, because a keyword tool cannot tell you what you sell and it will happily let you find out what somebody else sells instead.

The opponent is the workflow that starts at a seed keyword. It feels efficient and it inverts the whole thing: the tool suggests, you accept, and three weeks later you have a content plan built around a category rather than around a customer. Ahrefs starts its own process by telling you to put yourself in the customer's shoes for exactly this reason.

Here is the rule I would put on a wall, and it is the most useful sentence on this page:

Do not let a keyword tool teach you what your business does. Understand the business first, then use the tool to validate and expand what you already know.

The mechanism is simple and it is about ordering. Whatever you put into a tool first anchors everything that comes out of it. Seed it with project management and you get four thousand rows about project management, of which perhaps forty are about the specific thing you built. Seed it with the sentence a real buyer would say, and the four thousand rows come back centred on a problem you can actually solve.

The seventh question is the one that does the work: write the sentence a real buyer would say out loud, before they searched anything. Not a keyword. A sentence, with a situation in it. Everything downstream is a filter on that sentence, so if it is wrong the rest of the process will faithfully find demand for the wrong product.

A seed is a phrase you feed a tool or a search box to generate ideas. It is a starting point, and it is almost never a target, which is the part that trips people up.

The opponent is the belief that your seeds are your keywords. They usually are not. For an AI meeting assistant, the honest seed list is something like meeting notes, meeting transcription, ai meeting assistant, meeting summaries, zoom transcription, meeting productivity. Six phrases, and in the worked example further down exactly one of them survives to the final map. Two of the highest-value clusters were not on the list at all, and they turned up in Google autocomplete inside four minutes.

Moz's framework builds seeds from three places, and it is still the cleanest way to generate them: what you want to rank for, what you already rank for, and what your competitors rank for. I could not fetch Moz's current guide on the day I wrote this, so treat that as the established structure of the framework rather than as a quote from its current wording. The three-way split is right either way, and the middle one is the one people forget they have.

One test for a seed, and it takes two seconds. Say it out loud inside a sentence a customer would say. "I need a meeting productivity" fails. "I need meeting notes without typing them" works. A seed that only survives as a category name will generate a category-shaped list, and category-shaped lists are where vanity traffic comes from.

There are eight places keyword ideas come from and the keyword tool is the fifth. That ordering is the most useful structural thing on this page and it is not a contrarian pose: the four above it produce the language, and the tool's job is to tell you whether the language is real and how much of it there is.

The opponent is a tutorial genre rather than a claim. Watch almost any keyword research video and it opens with a tool already on screen. That is not dishonest, it is just that the tool vendor made the video, and the four free sources above it do not need a subscription to demonstrate.

Backlinko's current workflow makes the same move, and it is worth noting because Backlinko is owned by a keyword tool company: it explicitly uses Google related searches, Reddit, niche forums and YouTube auto-suggest before validating anything with keyword data. When the people selling the database tell you to go somewhere else first, that is a strong signal about where the language lives.

My own rule is a count rather than a philosophy. Four sources minimum, and never four that are all databases. Two databases and two humans will beat four databases every time, because the databases mostly agree with each other and the humans do not.

If your site has any history at all, start in Search Console and not in a keyword tool. An impression means Google has already decided you are relevant enough to show for that query, which makes it the shortest distance between where you are and traffic you do not have.

The opponent is every guide that teaches one process to two completely different readers. Somebody launching a product and somebody sitting on four years of impressions are doing different jobs, and running the new-site process on an established site wastes a fortnight finding demand you already had. Semrush's current workflow is the honest exception: it starts established sites on their existing positions to find quicker opportunities before expanding into new topics.

The single most valuable filter is positions five to twenty, sorted by impressions. That is a list of things you are close to and not converting, and it costs nothing. Underneath it is a second list nobody looks at: queries where you get impressions and almost no clicks, which is Google offering you the visibility and readers declining it. Those are investigation candidates, not automatic title-tag fixes. Impressions without clicks has at least eight causes and the snippet is only one of them: you may be sitting at position 14, an AI Overview or a SERP feature may be absorbing the click, the query may not match the page, the query may be peripheral to what the page is about, the brand may be unknown, or the numbers may be an average across devices and countries that hides two different stories. Check position, intent, SERP features and query-to-page fit before you decide what to change. Often the answer is a better snippet. Often it is that you should never have been on that query.

One more thing worth switching on if your property is eligible. Search Console can now split branded from non-branded queries, and Google notes the filter is not available for sites with a low number of impressions. Run it once. The number of businesses whose "SEO traffic" turns out to be mostly people typing their own name is higher than anybody wants to say out loud, and you cannot plan discovery until you know which half you are looking at.

A competitor ranking for something is evidence that demand exists and that somebody thought it was worth funding. It is not evidence that it is worth funding for you, and the gap between those two sentences is where a lot of wasted content lives.

The opponent is the gap report used as a to-do list. Three competitors rank for it, we do not, therefore write it. That logic produces a site that is a weaker copy of the market leader, and it produces it efficiently.

Four things worth pulling, in this order:

• Their top pages by traffic. Not their keyword list. The pages tell you which demand they actually monetised, and how they chose to satisfy it.
• Their commercial pages. Pricing, comparisons, alternatives, use cases. This is where a competitor tells you what they think the buying question is.
• Queries where several competitors rank and nobody is good. A crowded page one of thin results is a better opportunity than an empty gap.
• Questions nobody in the market has answered. The best of these come out of support threads rather than out of a gap report.

And one question to ask of every row before it survives: does this make sense for our business. A gap can exist because the competitor sells something else, to somebody else, in a country you do not operate in, on a business model where that traffic pays and yours does not. Every one of those is a real reason a gap is not yours, and none of them is visible in the report.

Use a model to generate language, cluster a list, and label intent across four hundred rows. Do not use one as a data source. Ahrefs puts the whole warning in one sentence: chatbots typically do not have access to real SEO data, so they often make things up and present them as facts.

The opponent is not people using AI, it is people using it for the one thing it cannot do. Ask a model "give me 50 different ways a small SaaS founder might search for software to summarize Zoom meetings" and you will get a genuinely good list in four seconds, better than most humans produce in an hour. Ask the same model "which of these gets 2,000 searches a month" and it will answer with total confidence and no data behind it.

So the principle, and it is worth memorizing:

AI generates hypotheses. Data validates them.

Where it is genuinely excellent, which is more places than the sceptics allow:

• Language generation. Fifty phrasings of one problem, including the ones your team would never write because you are too close to the product.
• Bulk operations. Four hundred rows deduplicated, grouped and labelled for intent, consistently, in a minute. This is the biggest real time saving and nobody writes about it because it is boring.
• Gap spotting. Give it your clusters and a competitor's, and ask what is in theirs and not yours. It is good at set differences.
• Argument. Paste your five priorities and ask it to make the case against each one. It will find at least one you had not thought about.

And the boundary. Anything with a number in it goes back to the data. If a tool will not tell you where its figures come from, treat the figures the way you would treat a stranger's estimate, because that is what they are.

Ten labels, six of which change what you would do and four of which only change how you sort a spreadsheet. Learn all ten so the words mean something when somebody uses them at you, then forget four of them.

The opponent is taxonomy as a lesson. It is the easiest part of this subject to teach, which is why it takes up so much room in guides, and it is close to the least valuable. Nobody has ever failed at keyword research because they could not define "navigational".

The four in the quiet rows are descriptions of where a query sits, not decisions. "Long-tail" describes a position on a demand curve. "Question" describes a grammatical shape. Useful for sorting, and neither one tells you whether to write anything.

The six that do change something all change it in the same way: they tell you what the searcher is at the point they typed. Somebody searching ahrefs login wants one specific destination and no amount of quality gets you that click. Somebody searching hire seo consultant is ready, which is why the volume being small is completely beside the point.

One warning on the branded row, because it is the one that fools people. Branded volume looks like demand and it is mostly awareness you already paid for somewhere else. A site whose search traffic is 70% its own name does not have an SEO channel, it has a reporting artifact, and the Search Console branded filter from the previous chapter is how you find out which one you have.

Long-tail is a position on the search demand curve, not a word count. It is the very large number of individually low-volume queries that sit below the head, and a three-word query can be long-tail while a six-word one is not.

The opponent is the definition every article uses: four or more words. It is easy to remember, it is wrong, and it leads directly to people padding queries to make them "long-tail" as though length were the thing being rewarded.

Almost every guide on this subject draws the curve freehand with no numbers on it. Here is the counted version, out of all 84,581 rows of one of my own Ahrefs exports, and it produced a finding I did not expect.

Read the caveat under that chart twice, because it is the finding rather than a disclaimer. A keyword export undercounts exactly the part of the curve it is named after: the genuine tail is millions of queries phrased in ways nobody else phrased them, and by definition a database of stored strings sees the least of that. Which is also the cleanest argument for Search Console over a keyword tool on any site with traffic. Your own impressions data contains tail queries no database has.

Two limits, both of which the data above supports and both of which get ignored. Low volume does not mean low competition. And a long query does not mean a valuable one. Mangools puts the general version well: do not become a slave of search volumes, take them as a clue. Here is the counted version of both exceptions.

Search volume is an estimated average of monthly searches, for one country, one match type and one twelve-month window. Google says so about its own numbers: Keyword Planner figures are rounded, and averaged over twelve months by default.

The opponent is not a belief anybody would defend out loud. It is a habit: pasting a volume figure into a report, a deck or an article with nothing next to it. Here is why that is not a fact, using one word and one afternoon.

None of those five figures is wrong. Each one probably answers correctly the question it was asked, and the questions were different: country against global, one export scope against another. The problem is that a cell labelled "Volume" does not record which question it answered, and by the time it reaches a slide nobody can reconstruct it.

So the working rule is short. A volume figure without its tool, its country and its date next to it is a number, not a fact. I hold my own writing to it: every figure on this page carries all three in the component that prints it, and the whole set gets re-pulled rather than left to quietly age.

And stop asking whether a volume is big enough. Backlinko is right that there is no magic number and that it depends on intent, competition, industry and seasonality. The better question is the one that has an answer: what would these searchers be worth if they were the right people. A hundred a month in enterprise procurement is a market. Twenty thousand a month of people who will never buy is a hosting bill.

Traffic potential is the total organic traffic the current number one page receives from every keyword it ranks for. It is a better input than volume because pages do not rank for one phrase, and it changes more decisions than any other metric on this page.

The opponent is comparing two keywords by their volumes. Here are two real rows from 21 August 2026. ai note taker: 20,000 US searches, difficulty 54. ai meeting notes: 1,500 searches, difficulty 55. Same difficulty, thirteen times the volume, so the big one obviously wins.

It does not. The page ranking first for the 20,000-search term collects 3,100 monthly visits. The page ranking first for the 1,500-search term collects 58,000, because it owns the whole cluster around it. Volume ranks those two one way and reality ranks them the other, and nothing except traffic potential would have told you.

Here is the whole data set the page is built on. Sort it by volume, then sort it by traffic potential, and watch the order change. That reordering is the lesson, and it is worth ten seconds of clicking more than it is worth another paragraph from me.

Now the counterexample, because a page where every rule holds is a page whose author stopped checking. Look at seo tools in the third tab: 442,000 US searches and a traffic potential of 97,000. Traffic potential is lower than volume, and it is supposed to be the bigger number.

The mechanism is not mysterious. Traffic potential measures what the page ranking first actually receives, and on a 442,000-search results page with an AI Overview on it and 0.93 clicks per search, most of that volume never becomes a click on anything at all. So traffic potential is not a bigger version of volume. It is a different measurement, of a page rather than of a query, and it inherits every property of the page it measured.

Which produces the second rule for reading it: look at what the number one page is. If it is a brand homepage or a product, a high traffic potential tells you nothing about an article you could write. In the pull above, ai meeting notes has a parent topic of "read", which is a company. The 58,000 belongs to a product page, and a blog post is not going to inherit it.

Ahrefs Keyword Difficulty counts how many unique websites link to the current top ten pages, and turns that into a number out of a hundred. That is what it is. It is not Google's opinion, it does not know your site, and Ahrefs says both of those things in its own documentation.

The opponent is the filter. KD under 30, sort, export, publish. It is the most common keyword workflow in the world and it is how you end up writing about a Korean actress.

I am not being rhetorical. Of the 372 rows above 10,000 searches in my own SEO keyword export, 108 score KD 30 or lower. Read five of them and the mechanism is instantly obvious: seo ye ji, chrome://settings/searchengines, search google or type a url. Every one of those passes a volume-above-10k, difficulty-below-30 filter, and every one is worthless to an SEO site. The counted version of both exceptions is in the panel back in chapter ten.

Ahrefs is unusually straight about the limits of its own metric, and it is worth quoting because it is a vendor arguing against its own number. Its guide says the score does not account for your specific website authority, does not tell you whether the ranking pages actually satisfy the user, and that the only way to make the right bets is by thoroughly studying the search results.

So the practical rule is a reframe rather than a threshold. Difficulty is a cost, and a cost is only a reason to say no once you know the price of the thing. A KD 90 term that describes exactly what you sell can be worth three years of work. A KD 0 term that describes something you do not do is worth nothing this afternoon.

I am doing that live on this page. seo keyword research scored KD 90 on the day I wrote this, with a $4.00 CPC and a traffic potential of 280,000. I am writing it anyway, because it is precisely what I do, because I already have two ranking pages in the topic, and because the alternative is writing about SEO jobs.

Real difficulty is a list rather than a number: what actually ranks, what format those pages are, how good they are, how old, how many links, how strong the brands, how much topical ground you already hold, and what you would have to make. The score is a first filter over that list. Reading page one takes four minutes and beats it every time, which is the subject of the next lesson.

Cost per click is a vote, in money, from people who have already measured what that visitor converts at. It is the best free signal of commercial value there is, and it is a signal about their business rather than about yours.

The opponent is treating CPC as a verdict in either direction. High CPC means valuable, low CPC means skip. Both readings are wrong and the second one is more expensive, because a low CPC often means nobody has worked out how to monetise a query yet, which is an opportunity rather than a warning.

Here is the version worth remembering, from my own market and my own pull. ai tools: 56,000 searches at $2.00 a click. ai seo consultant: 400 searches at $13.00 a click. Advertisers, spending their own money, have priced one of those visitors at six and a half times the other. Then remember that the expensive one is 140 times smaller, and notice that the market has already told you which one an AI SEO consultant should care about.

Two more clues in the same family, both of them in the live table above and both of them routinely ignored.

• Clicks per search. what is seo has 25,000 searches and 0.54 clicks per search. Barely half of those searches produce a click on anything. Volume is not traffic and this is the column that says so out loud.
• What is on the results page. Of the 22 keywords I pulled, 20 carried an AI Overview and 21 carried People Also Ask. A results page that answers the question above your listing changes what the volume is worth without changing the volume at all.

A twelve-month average is precisely the statistic that hides a shape. Stable, seasonal, growing, declining and brand new are five different situations, and all five arrive wearing the same single number.

The opponent is the volume column read on its own. Two keywords at 5,000 searches, one of which does 4,000 of them in November and one of which has been falling for three years, look identical in a spreadsheet and are completely different investments.

Google Trends is the free fix, and there is one thing you have to know before quoting it. Google says each data point is divided by the total searches of the geography and time range it represents, then scaled 0 to 100 against the topic's proportion of all searches. It is a relative index. A value of 100 is the peak inside the range you chose, not a quantity, which means two screenshots taken over different windows cannot be compared and a Trends number can never be added to anything.

One habit worth adopting: set the range to five years, not twelve months. Five years is where a slow decline becomes visible, and a slow decline is the thing a keyword tool is worst at showing you. A term losing fifteen percent a year still looks perfectly healthy in a twelve-month average, right up until the page you built for it stops earning.

You need exactly enough search intent here to reject an obviously wrong keyword. Everything past that belongs in the next lesson with a results page open, and cramming it in here would wreck both.

The opponent is intent taught as a taxonomy to memorise. Four labels, definitions, examples, done. That version is genuinely useless, because the labels are easy and applying them to a real ambiguous query is not.

The four you need are the standard four, and one sentence each is the right amount:

• Informational. They want to understand something.
• Commercial. They are choosing between named options.
• Transactional. They are ready to do the thing.
• Navigational. They want one specific destination.

Now the use. Take seo jobs: 5,300 searches, difficulty 0, the easiest thing in my whole export. If you sell SEO software or SEO consulting, the intent is somebody looking for employment, and no version of your page can serve that and also serve you. That is the rejection intent buys you at this stage, and it is worth the four sentences.

What you cannot do yet is decide the format. Whether a cluster wants a comparison, a listicle, a tool, a documentation page or a product page is a question you answer by reading what currently ranks, and doing that properly is the whole of the next step. Here, intent is a filter. There, it becomes a specification.

A hundred searches from people with your exact problem are worth more than ten thousand from people who will never buy. Business value is the axis a keyword tool has no column for, and it is the one that overrules the columns it does have.

The opponent is volume-first prioritisation, and I want to be fair about why it persists: volume is the only number in the room that everybody agrees on and that goes up. Business value is a judgement, it cannot be sorted, and somebody senior will argue with it. So it gets skipped, and the plan comes out looking impressive and earning nothing.

Ahrefs calls this Business Potential and scores it 0 to 3. That model is fine and I use four words instead, for one reason: a 0 to 3 score gets averaged into a total by the second week, and the whole value of this judgement is that it cannot be averaged.

Search Engine Journal puts the general principle plainly, and it is the sentence to carry out of this chapter: not all traffic is equal. Its version adds the part people resist, which is that low-volume keywords can be much more valuable because they can deliver users who are ready to buy.

Reading about a judgement is not the same as making one, so here are five. Every pair uses real rows from the pull, you have to pick a side before the reasoning appears, and in two of the five I take the option the metrics argue against.

Clustering is grouping queries that represent one underlying need closely enough that a single page can satisfy all of them. It is the step that turns a keyword list into a content plan, and it is the step that stops you publishing five pages that compete with each other.

The opponent is one keyword, one page. It was reasonable advice in 2012 and it is now actively harmful, and I do not need an opinion to show you why. Here are two rows from one keyword database on one day. free ai note taker: 2,100 searches, difficulty 57, parent topic "ai note taker". ai note taker free: 1,600 searches, difficulty 83, parent topic "ai note taking".

Same three words. Two volumes, two difficulty scores twenty-six points apart, and two different parent topics, because the word order changed. Nobody typing either of those wants a different page. The tool cannot see the need either, which is exactly why the grouping is your job.

So do it by hand once before you automate it. Twenty-six real queries, sorted by volume the way an export arrives, and two answer keys at the end: mine, and the parent topic the tool assigned.

Two things worth saying about how you group. Group by the sentence the searcher would say, not by the words they typed, and name each group as a need rather than as a keyword stem. "They want to know if there is a free version" is a group name you can act on. "Free keywords" is a folder.

Semrush defines clustering as grouping search terms that share the same search intent and targeting them together on a single page, which is the same idea from the vendor side. Its manual method leans on SERP similarity: if the same pages rank for two queries, cluster them. That is the right method and it needs a results page open, so it belongs in the next lesson rather than this one. What you can do here, with nothing but judgement, gets you about eighty percent of the way.

And notice again what Google itself shipped. Query groups in Search Console Insights groups similar queries, names the group after its best performer, and reports the clicks for the whole group rather than for the top query. That is clustering, built into the reporting, by the company whose results you are trying to appear in.

Mapping is deciding what should satisfy each cluster. A page, a section of a page you already have, a product page, a free tool, a video, an FAQ answer, or nothing at all. This is where research stops being research.

The opponent is the default. Every row becomes an article, because an article is the thing content teams know how to make, and because a keyword tool's output looks like a list of article titles. That default is how sites end up with two hundred blog posts and four commercial pages, wondering why the traffic does not convert.

Run every cluster past five options before you reach for the sixth:

• A page you already have. Check first, always. Improving a page that already ranks in position eleven is faster and cheaper than anything else on this list, and it is the option a fresh keyword list never suggests.
• A commercial page. Pricing, a feature page, a comparison, an alternatives page, a use case. If the cluster is people choosing, the answer is usually not a blog post about choosing.
• A section rather than a page. Many clusters are a subsection of something you should be writing anyway, and splitting them out creates two pages that are each worse than one would have been.
• A different format entirely. A free tool, a template, a calculator, a video, a documentation page. Some needs are better served by something you use than by something you read.
• Nothing. The best row in a finished map is often the one that says ignore, with a reason next to it, so nobody has to rediscover it in six months.

Semrush treats mapping terms to relevant pages as a distinct part of the strategy rather than as a byproduct of the keyword list, and that separation is the useful bit. The list is not the plan. The mapping is.

Where those pages then sit, what links to what, and how a hub relates to its spokes is a site architecture question and it comes later. What you need here is one decision per cluster and a reason under it.

There is no honest formula. Priority is contextual, it changes with the business rather than with the keyword, and any tool that hands you a single opportunity score has multiplied together things that do not share a unit.

The opponent is exactly that score, and it is seductive because it sorts. Ahrefs makes the point from the inside: prioritisation changes based on the website, the team, the need for conversions, the time horizon and the resources. If you are chasing traffic fast, high volume and low difficulty is the right filter. If you are responsible for leads, business potential is the metric that matters. Same keyword, two correct answers, depending on whose quarter it is.

So use nine questions instead, and let two of them be vetoes. Relevance and business value can overrule every other answer, because a total of sixteen out of eighteen should never outvote "these are not our people".

The genuinely useful output of that tool is not the verdict. It is the weak answers it names back to you, because those are the questions you were quietly hoping nobody would ask. A plan that never writes down what is weak about its own choices is a plan that gets surprised in four months.

One prompt is not one keyword and is not one page. A generative system takes a long, specific question and fans it out into several related searches, and the searches it lands on are ones that were already in your keyword database.

This chapter has two opponents and they are the two loudest positions in the industry. "Prompts have replaced keywords" and "nothing has changed". Both are wrong, and the cheapest way to show it is to run a real prompt against a real keyword pull.

So what changed is most likely not the demand. It is the assembly. A person researching a CRM used to run four searches over two days, and a model now runs an unknown number of them in four seconds and hands back a synthesis. I say unknown deliberately: the count and the exact phrasing are Google's business and not observable from here. What is observable is the consequence, and it is enough to plan against. The thing being rewarded shifts from ranking for one query to being retrievable across a cluster, and the unit that gets retrieved is a passage rather than a page.

That has one concrete consequence for research, and it is not "collect prompts instead". It is: research the problem, the sub-problems, the comparison criteria, the constraints and the objections, because those are the needs a complex prompt is made of. For the CRM example that means pricing, integrations, team size, the founder-led sales constraint, and the Slack requirement. Every one of those is a query, most of them are in the database, and almost nobody targets them because they look too small on their own.

Prompt research is a real and separate input, and Semrush, who sell it, describe the boundary more honestly than most people quoting them. They define it as identifying and tracking the questions that cause AI systems to compare options and recommend brands, note that there is no history of search volume, CPC or trend data for prompts, and say directly that keyword research is still relevant and plays an important supporting role because it reveals how people describe problems and what intent sits behind them.

Two inputs, different maturity, both real. Keyword research measures conventional demand well and has fifteen years of history behind it. Prompt research shows you phrasing, context and the recommendation moment, and has almost no measurement behind it yet. Use both, and be suspicious of anybody selling the second as a replacement for the first while quietly using the first to build it.

If you want the deeper mechanism, I wrote up why fan-out is a passage problem rather than a page-length problem in query fan-out in SEO.

Eleven claims I have been told about this subject, all of them wrong, several of them by somebody selling the tool that creates the problem. Four are answered by data I pulled the day I wrote this, and two are answered by the vendors themselves.

Fourteen questions. Every answer is sourced or measured, and every one targets a distinction that changes a decision rather than a definition you could look up.