The Keyword Research Checklist That Tells You What Not to Write

Most keyword research checklists only know how to add

Open any of the checklists ranking for this term and you will find the same six steps: brainstorm seeds, run them through a tool, pull volume and difficulty, sort by intent, cluster into topics, map to a calendar. Every step adds keywords. Not one of them removes any.

That is why the output of most keyword research is a spreadsheet of 400 rows that nobody ever publishes against. The list is accurate. It is also longer than the company can write, and it contains no instruction about which rows to abandon. So the team writes from the top down, runs out of budget somewhere around row 30, and the other 370 rows sit in a tab that gets opened once a quarter.

I ran marketing for a 4x Inc. 5000 firm from startup to exit, and the keyword lists were never the constraint. Publishing capacity was. A B2B company with one marketer and a freelance writer publishes maybe four pieces a month. That is 48 a year. If your research produced 400 candidates, the research did not tell you what to write. It told you what exists, and then handed you a 12 percent selection problem it made no attempt to solve.

The checklist below solves for that. Each gate is a reason to reject a keyword. A term has to survive all four to earn one of your 48 slots.

Gate one: check whether the click still exists

Volume is a measure of how many people search. It has stopped being a measure of how many people arrive. In the first four months of 2026, 68.01% of Google searches ended without a click, according to Rand Fishkin's analysis of Similarweb clickstream data. Slightly under a quarter of searches send a click to the open web at all.

The damage is not spread evenly, which is what makes this a usable filter rather than a general worry. Ahrefs' Ryan Law found that when an AI Overview is present, clicks to the top organic result fall by roughly 58%, up sharply from the 34.5% drop Ahrefs measured in its earlier study. And AI Overviews are concentrated: they appear on 57.9% of question queries. Where no AI Overview appears, Ahrefs found organic CTR has held up or risen.

So the gate is mechanical. Search the keyword. If an AI Overview renders and it answers the question completely, a first-place ranking on that term is worth roughly 40% of what your volume-based forecast assumed. Discount it accordingly, or reject it.

This kills a specific category, and it is usually the category that looks most attractive on a keyword report: the high-volume "what is" and "how does" terms. A 12,000-search definitional keyword with a complete AI Overview sitting above the fold is a worse asset than a 200-search term that Google has no confident answer for. Run this gate before the volume column has a chance to make the decision for you.

Gate two: run the pipeline math before you read the volume number

Every guide on this topic says business value matters more than volume. None of them do the arithmetic, which is unfortunate, because the arithmetic is where the counterintuitive answers live.

Take a B2B company with a $4,000 per month engagement. Run two candidate keywords through the same chain of realistic numbers: searches, click-through at your position, the share of visitors who convert to an enquiry, and the share of enquiries that close.

A 15,000-search informational term, ranked third, with an AI Overview present: call it 4% CTR after the AIO discount, 600 visits, a 0.4% conversion rate because the reader is a student or a competitor or someone three years from buying, and a 10% close rate. That is 0.24 clients a year.

A 150-search term with obvious commercial intent, ranked second, no AI Overview: 18% CTR, 27 visits, a 6% conversion rate because the searcher has a live problem and a budget, and a 25% close rate. That is 0.4 clients a year.

The 150-search keyword outperforms the one with 100 times the volume. At a $48,000 annual contract value the gap is worth real money, and the smaller term is also cheaper to rank for and faster to write. This is the calculation that should govern the list, and it is the reason a Demand Engine gets built around commercial-intent terms rather than traffic. We wrote more about the selection logic in our piece on competitor keyword research, where the same problem shows up as a gap report that reads like a plan but is not one.

Two practical notes. Use your own close rate rather than a published benchmark. If you do not know your conversion rate from organic traffic, that is the thing to go fix before you buy another keyword tool. Then be honest about intent. "Best cold email software" has a buyer behind it. "What is cold email" has a reader. Both can be worth writing. Only one of them should be at the top of a list of 48.

Gate three: find the sentence the top five results will not write

If you cannot name, before drafting, the specific thing your page will say that the current top five do not, you are about to publish a summary of the SERP. Summaries do not outrank the pages they summarise, and they give a language model nothing new to cite.

Google's own guidance on creating helpful content asks whether a page provides original information or original analysis, and whether it delivers substantial value compared with other pages in the results. Read that as a description of the only condition under which a new page has a reason to exist.

The test is a single sentence, written before the outline. For this post it was: every checklist in this SERP only adds keywords and none of them has a rule for rejecting one. If you cannot write that sentence, the keyword fails the gate. It might still be worth targeting later, once you have a proprietary number, a client result, or a position you can defend. Today it is a row you skip.

What usually supplies that sentence in B2B is operating data nobody else has. Your close rate by traffic source. What actually happened to the 40 accounts you ran a play against. The number you had to stop reporting because it was measuring the wrong thing. A company with two years of delivery history is sitting on more original material than the agency blogs it is competing with, and almost none of it is written down.

Gate four: check the page can survive being read by a machine

A keyword that clears the first three gates can still produce a page that fails, because increasingly the reader is a retrieval system deciding which passage to quote.

Two findings should change how you evaluate a target. An analysis of 18,012 verified ChatGPT citations reported by Search Engine Land found that 44.2% of citations come from the first 30% of a page, with content buried deep roughly 2.5 times less likely to be cited. The classic structure that withholds the payoff until the conclusion is actively hostile to citation. And Semrush's ghost citations study with Kevin Indig found that 61.7% of AI citations never name the brand: your page gets used as a source while the answer credits nobody.

For keyword selection this has a concrete consequence. Terms whose honest answer is a single paragraph will be answered by the machine and attributed to no one. Terms that require a judgement, a sequence, a set of trade-offs, or a number you own are ones where a retrieval system still has a reason to send a person to you, and where your name has a reason to appear in the sentence.

So at selection time, ask whether the answer to this query is a fact or a decision. Facts get absorbed. Decisions get visited. If you are targeting the term anyway, front-load the finding in the first 30% of the page and write the brand into the sentences carrying the claims, which is the structural half of the same problem we cover in how to show up in Google's AI Mode.

The kill list is the deliverable

Run all four gates and a 400-row spreadsheet comes out somewhere between 30 and 60 rows. That is the useful artefact, and the discarded rows are almost as useful, because a written reason for rejection stops the same keyword being re-proposed every quarter by someone who has only seen the volume column.

Structure the survivors into three buckets rather than one ranked list. Write now, for terms that clear all four gates and have a buyer behind them. Write when we have the proof, for terms that fail only gate three, held until a client result or a proprietary number exists to carry them. Do not write, for everything else, with one line saying which gate it failed.

Then commit the first bucket to actual dates. A prioritised list with no calendar attached behind it is the same spreadsheet you started with, sorted differently. Assign each survivor a publish week and an owner, cap the queue at your real publishing capacity rather than your ambition, and let the rest sit in the second bucket where they belong. Turning a selected keyword into a page someone can actually write is a separate discipline, and it is worth using a content brief template so the reason you picked the term survives the handoff to whoever drafts it.

The last step is the one most teams skip. Put a review date on the kill list. The gates move: AI Overview coverage expands into new query types, a competitor publishes the piece you were holding, your own delivery history finally produces the number that unlocks a term you rejected in March. A kill list reviewed quarterly is a working document. One written once is just a longer version of the problem.

If you want the condensed version of the gates plus the structural checks that follow them, our SEO and AEO checklist covers both, and the Demand Engine is where this selection logic runs as a system rather than a quarterly spreadsheet exercise.