AI keyword lists are helpful without search volume, but only as a draft of questions to check, not as a content plan. A language model is genuinely good at one part of keyword research, listing the many ways people might phrase a problem, and genuinely bad at another, knowing which of those phrasings real people use and how often. Used alone, an AI list mixes real questions and invented phrasings with equal confidence. Used as a starting point and validated against observed data, it saves hours. That validation step is what SQSEO is built around: it starts from the real questions people ask Google, ChatGPT and Perplexity, so the list you plan from is made of demand you can see.
What an AI keyword list actually is
When you ask ChatGPT or Claude for “fifty long-tail keywords about project management for architects”, the model does not look anything up. It generates phrases that are statistically plausible given everything it has read. Some will match what people really type. Some will be fluent combinations that no person has ever searched. The output looks identical in both cases.
That is not a flaw to be fixed with a better prompt. It is what a language model is. Research on query rewriting for retrieval, such as the work on rewriting queries for retrieval-augmented models, uses exactly this ability on purpose: models are good at producing alternative phrasings of a need. The same ability that makes them useful for expanding a query makes them unreliable as a record of demand.
Where AI lists help, and where they mislead
| Task | Can an AI list do it? | Why |
|---|---|---|
| Brainstorm ways a problem is phrased | Yes, well | Paraphrase is what models do best |
| Surface adjacent subtopics you missed | Yes, mostly | Models connect related concepts broadly |
| Group keywords by the question behind them | Yes, well | Clustering by meaning is a strength |
| Say which phrasings real people use | No | The model generates, it does not observe |
| Estimate how often a question is asked | No | No access to query logs; numbers are invented |
| Reflect very new or niche vocabulary | Poorly | Little training data, so it borrows generic terms |
| Match how your actual buyers speak | Partly | Only if you give it real customer language first |
The pattern is consistent: models are strong at language and structure, weak at facts about frequency. Plan around that split and AI lists become one of the cheapest research inputs you have.
Why “no volume” is not the same as “no demand”
The second half of the question is about volume, and here the instinct to discard anything without a number is usually wrong.
Keyword tools show zero, or a dash, for anything below their reporting threshold, and their data comes from samples. Google itself has said that 15 percent of the queries it sees every day are ones it has never seen before. A large share of real demand is therefore made of phrasings too rare or too new to carry a volume figure anywhere.
In AI assistants the effect is stronger. Prompts are longer and more personal than search queries, because people describe their situation. The NBER study How People Use ChatGPT found that practical guidance, seeking information and writing make up nearly 80% of conversations, which are exactly the kinds of requests people phrase in their own words. Almost none of those prompts would register in a volume tool, and many of them come from people close to a decision.
So the right question is not “does this keyword have volume?” but “is this a real question?” A zero-volume question that came from a sales call is real demand. A zero-volume phrase that exists only in an AI output is a guess.
How to validate an AI keyword list
Validation does not need expensive tools. It needs at least one observed source per entry.
- Generate broadly. Ask the model for phrasings, sub-questions, objections and comparisons around your topic. Give it real customer language first, such as a few sentences from sales calls or support tickets, so its output starts closer to how buyers speak.
- Check against search behaviour. Look for each phrasing, or its core, in search suggestions and People Also Ask boxes. Check Google Search Console for queries your site already appears for.
- Check against communities. Search Reddit, niche forums and review sites for the question. Real threads confirm real demand and show the exact words people use.
- Check against your own data. Sales notes, chat transcripts, support tickets and onboarding surveys are the most underused keyword source in most companies.
- Group and prune. Merge entries that ask the same question. Drop phrasings that appear in no observed source. Keep the ones confirmed at least once.
- Prioritise by value. Rank what remains by how close the question sits to a buying decision, then by how winnable it looks, then by any visible volume as a tiebreaker.
In our experience the validated list is usually around a third the length of the raw AI output, and it is the third worth writing for. Our guide to prompt-based keyword research covers steps one and five in more depth.
Why invented keywords cause real damage
An unvalidated list is not just inefficient. It creates three specific problems.
Content nobody needs. Each page written for an invented phrasing costs the same as one written for a real question, and earns nothing.
Cannibalisation. AI lists contain many near-identical phrasings. Writing a page for each produces several pages competing for the same intent, splitting signals that one strong page would have concentrated.
Thin, generic pages. Content written to a keyword that has no real asker tends to be vague, because there is no real situation to address. Search engines and AI assistants both favour pages that answer a specific need well. The GEO study found that adding concrete statistics, quotations and citations raised visibility in generative answers by up to 40% in its tests, which only happens when there is a specific question to be concrete about.
Which observed sources deserve the most trust?
Not every validation source is equal, and it helps to know how much weight each one carries before you start pruning.
Your own customer conversations rank first. A question that appears in sales calls, support tickets or onboarding surveys is confirmed demand from people who are, or nearly were, buyers. It also comes in their exact words, which is the language your page should use.
Google Search Console comes next for sites with some traffic. Queries your pages already appear for are observed demand with a click attached, and the long tail there is often full of zero-volume questions that tools never show.
Search suggestions and People Also Ask boxes are a strong public signal that a phrasing is common, although they lean toward the head of a topic rather than the specific tail.
Community threads on Reddit and niche forums are the best source for comparative and situational questions, the kind people put to assistants. Their weakness is that a single loud thread can make a rare question look common, so look for repetition across threads.
Competitor content is the weakest signal. A competitor ranking for a phrase shows that someone targeted it, not that anyone asks it.
Do AI keyword lists work in other languages?
They work, with a stronger caveat. Models have far more English training text than text in most other languages, so their phrasings outside English drift toward translated English: grammatically fine, but not how native speakers search. Local terms, abbreviations and brand nicknames are often missing.
The fix is the same validation, applied more strictly. Feed the model native customer language first, check every phrasing against local search suggestions and local communities, and ask a native speaker to read the final list. A list that sounds translated will produce pages that read translated, and both search engines and readers notice.
How AI search changes long-tail research
AI search makes the long tail more important, not less, and it changes what a good long-tail list looks like. When an assistant answers a question, it often breaks it into several sub-queries behind the scenes and retrieves sources for each. A page that answers one specific sub-question clearly can be retrieved even if the user’s original prompt was much broader.
That means the useful unit of research is the question, not the keyword string. Instead of fifty variations of “project management software architects”, a good list holds the twelve distinct questions architects actually ask: how to track fee stages, how to handle consultant coordination, which tools integrate with drawing software, what a small practice should pay. Each becomes a section or a page that answers one thing well. We argue the same point in is long-tail keyword research dead.
When do you actually need volume?
Some decisions genuinely depend on size, and for those an AI list is the wrong tool.
Paid search budgets. Bidding needs measured demand. Use the advertiser tools built for it.
Market choice. Deciding between two countries, languages or segments needs comparable numbers from the same source.
Traffic forecasts. Any forecast built on invented volumes is fiction with a chart. If someone must forecast, use measured search data and state its limits.
For content prioritisation, by contrast, volume is the least important input, and in AI search it barely exists, as are AI search volumes measurable explains. Value and winnability matter more, and a validated question list gives you both. If you want measured numbers without a paid tool, our guide to a free keyword planner with exact search volume covers the options.
How to prompt a model for a better first draft
The quality of the raw list depends heavily on what you put in. Four habits make AI keyword lists noticeably more useful.
Feed it real language. Paste a few anonymised customer emails, call notes or forum posts before asking for phrasings. The model mirrors the register it sees, so the output starts closer to how buyers talk.
Ask for questions, not keywords. “List the questions a first-time buyer asks before choosing” produces more useful output than “list keywords”, because questions carry intent.
Ask for stages. Request questions grouped by stage: understanding the problem, comparing options, deciding, and using the product. This spreads the list across the journey instead of clustering at the top.
Ask it to flag uncertainty. Instruct the model to mark phrasings it believes are common versus ones it constructed. It cannot know for certain, but the flag helps you decide where to check first.
None of this replaces validation. It just means the list you validate is better to begin with.
A worked example
Suppose you sell scheduling software for physiotherapy clinics. An AI tool returns sixty keywords. After validation against search suggestions, a physiotherapy forum and twenty support tickets, you keep nineteen. They group into seven questions: reducing no-shows, handling multi-practitioner calendars, insurance billing integration, online booking for new patients, migrating from paper diaries, pricing for a two-person clinic, and data protection for patient records.
Four of the seven show no volume in any tool. Two of those four appear repeatedly in support tickets and in forum threads where clinic owners compare software. Those two become your first pages, because they are close to a purchase and nobody else answers them well. The highest-volume phrase on the original AI list, a generic “best scheduling app”, is dropped, because it is broad, dominated by large publishers and far from your buyer.
Key takeaways
AI keyword lists are useful as drafts of questions and harmful as plans. Models are strong at phrasing, grouping and finding adjacent topics, and unable to tell you which phrasings people use or how often. Missing volume is not missing demand: Google says 15 percent of daily queries are new, and AI prompts are even more specific. Validate every entry against at least one observed source, merge duplicates, drop what appears nowhere, and prioritise by buyer value and winnability. Keep volume for decisions that depend on size, such as budgets and market choice.