How to do keyword research when a third of the answers never get clicked

Keyword research has not changed. What changed is the answer to the question at the end of it, which is and would this be worth winning. On a growing share of queries the answer now appears above the results, the searcher reads it, and nobody visits anything. Research that ignores this produces a content plan full of pages that will rank and be ignored.

So this is the method, with the newer judgement built into it rather than bolted on.

Step 1: get the phrasings from something other than your own head

The single most common failure in keyword research is that every candidate was invented by the person doing the research. What you get is a list of terms your industry uses internally, which is reliably not what buyers type.

Three sources produce real strings rather than plausible ones:

• Autocomplete. Type your seed and go through the alphabet, then again with question words. These are ranked by prediction likelihood, which correlates with frequency. They are free and they are real. This is not a small source: the last time we ran it systematically, 142 seeds returned 16,332 distinct real queries at no cost.
• The results page itself. Related searches, People Also Ask, and the phrasing the ranking pages use in their own titles.
• Your own inbox and site search. The exact words customers used when they asked you a question. This is the highest quality source anybody has and almost nobody mines it.

Expand until you have far more candidates than you could ever build. Discovery should be cheap and wide, because the cut comes next and it should be brutal. If you would rather run the expansion as a prompt than by hand, the seed keyword expansion prompt and the long tail keyword mining prompt do the alphabet sweep and the question sweep respectively.

Step 2: attach real volume, and only then

Volume data comes out of advertising platforms, which is why free tools either estimate it or do not have it. Google’s Keyword Planner is the origin for most of the market, directly or through resellers, and its well known limitation is that without meaningful ad spend it reports wide buckets rather than numbers. If a figure matters to a decision, pay for it once and write the whole response to a file so you never buy it twice.

Two habits are worth building here. First, pull volume after discovery, not before, because pulling is billed per request and a single request holds hundreds of keywords. Second, look at the twelve month trend rather than the average. A term averaging a thousand a month that has been falling every month for a year is not a thousand a month, it is a decline you are about to build a page for.

Which tool you buy it from matters less than people think, and what actually differs between keyword tools goes through why.

Step 3: the cut, which is the actual work

Everything up to here is mechanical. This is the part that decides whether the plan is any good, and it is three questions in order.

Can we say something true here that is not already being said? If the honest answer is that we would be writing a slightly different version of what already ranks, the page will be a slightly different version of what already ranks, and it will place accordingly. This kills more candidates than anything else and it should.

Would the person typing this ever buy from us? Volume with no buyer is a traffic number that flatters a report and moves nothing. This is where most large content plans quietly go wrong: they optimise for the total at the bottom of the spreadsheet.

Could we win it from where we stand? Authority is the constraint that decides this, and it is not negotiable by writing better. A site with almost no referring domains competing against established publishers is not going to place on a head term regardless of how good the page is. The honest move is to build where the competition is beatable and to fix authority as a separate workstream rather than pretending content will do it.

We are describing our own position here rather than a hypothetical one. This domain has a small number of referring domains against rivals with hundreds, and the whole query set sits well down the first several pages. That is not a reason to publish nothing; it is a reason not to price a plan on head terms.

Step 4: the newer question, which is whether the click survives

This is the part that did not exist five years ago.

For each surviving candidate, ask what the results page looks like when the query runs. If a generated answer sits at the top and fully satisfies the question, ranking first underneath it is worth much less than the volume implies. Published measurements of how often that answer appears vary widely depending on who is counting and how, and the honest range is wide enough that anybody quoting a single precise figure is telling you more about their methodology than about the web. What is not in dispute is the direction, and that it lands hardest on short definitional queries.

Two things make this harder to reason about than it looks. Google documents that AI Overviews only trigger where its systems judge them additive, so trigger rate is a property of the query rather than of your page. And both AI Overviews and AI Mode use query fan-out, issuing several related searches behind one typed query, which means the thing being answered is often not the string you researched. What Google AI Overviews are and how they work covers the mechanism, and how to rank in them covers what is actually actionable about it.

The pattern that falls out of this is consistent:

• Definitional and short queries. High volume, increasingly answered in place. Still worth covering for the sake of being the source that gets cited, but forecast the clicks low and do not build a business case on them.
• Specific, comparative and situational queries. The click survives, because a generated summary is a poor substitute for a real comparison and searchers know it.
• Anything with a decision or a spend behind it. Untouched. Nobody buys from a summary.

That reordering is the practical output of this whole exercise. It moves effort down the tail, towards queries that are longer, more specific and less contested, and away from the head terms that look best in a plan.

The dimension nobody had to research before

There is now a second demand surface, and it does not have a keyword tool.

People ask assistants things they would never type into a search box, in full sentences, with context attached. There is no volume figure for these because no advertising platform sells against them. What there is instead is a different measurement: take the prompts your buyer would plausibly ask, run them against the assistants, and count how often you are named. That is what the AI visibility checker does, and the prompt library is the working set we test with.

That is a rate rather than a ranking, and it has to be re-run because the answers vary between runs on an unchanged site. It is a genuinely different instrument, and it answers a question keyword volume cannot: not how many people ask this but when they ask it, does anyone mention us.

Both matter. Keyword research tells you what to build. Prompt testing tells you whether what you built made you nameable. Doing only the first is how sites end up ranking respectably and never being recommended, and doing only the second is how people end up chasing tactics Google has published as ineffective while their canonical tags are broken.

For the classic half, position tracking still answers the ordinary question of where you actually sit, and what a keyword rank tracker does and does not tell you is the short version of how to read one.

What to keep from all this

Discovery should be wide, cheap and sourced from real strings. Volume should be bought once, read as a trend, and never treated as a forecast. The cut is the work, and it runs on three questions about differentiation, buyer fit and reachable authority. And the last filter is new: ask whether anybody will still click, because on a growing share of queries the answer is no and the plan should say so out loud rather than counting those impressions as wins.

The one measurement that is genuinely yours sits in Search Console. Queries where impressions held and clicks fell are where this is happening to you specifically, and no industry benchmark is needed to read it; the Search Console keyword analysis prompt does that split on an export.

Sources

Read on 15 August 2026.

• Google Ads, use Keyword Planner, for where volume data originates and how it is bucketed without spend.
• Google, AI features and your website, for AI Overviews only being shown where Google’s systems judge them additive, and for how they are counted in Search Console.
• Google, optimizing for generative AI search, for the query fan-out description and for the list of tactics Google says you can ignore.
• Google, creating helpful, reliable, people-first content, for the differentiation question in step three, put as a self assessment.